E2E check and profiling
All 60 pairings have a disposition. The timing tables below remain the 2026-09-22 measurements. The configuration matrix and affected per-cell reasons incorporate the 2026-09-23 scoped recovery runs and WSL CPU/native Windows expansion. The expansion adds separate 20-request serial MiniCPM-o text-to-speech profiles on WSL CPU and native Windows GPU, a Spark output-head component profile on the HX370 AMD NPU, complete Spark text-request profiles through standalone llama.cpp and a bounded Omni whole-session stage on native Windows CPU and Radeon 890M, native Windows CPU Qwen3-TTS text-to-WAV profiles through both standalone PyTorch and a bounded Omni whole-session stage, and a hybrid CPU+Radeon 890M Qwen3-TTS profile likewise repeated through a bounded Omni whole-session stage. A pinned GGUF MiniCPM-o route adds combined image+audio to text+speech runs on WSL CPU, native Windows CPU and native Windows Radeon 890M+CPU. All three now pass a bounded complete request through an Omni graph stage; the WSL CPU path additionally passed one warmup plus 20 serial measured requests. These are scoped synthetic cases, not full release qualification.
The WSL CPU and CPU+Radeon InternVLA cells also have bounded Omni complete-policy profiles: one warmup and 20 serial patterned synthetic requests per placement, plus admission refusal, cancellation and a public hybrid AsyncOmni.generate request. The WSL RTX 5090 Laptop CUDA and native Windows RTX CUDA cells now have the same scoped Omni whole-policy pattern with 20 serial requests each, explicit host/VRAM admission, refusal, cancellation and public action requests. Separate native Windows CPU and CPU+Radeon Omni profiles add 20 serial patterned requests per placement, refusal, cancellation and public action requests. These advance only the synthetic action-policy path. Real observations/reference actions, physical action units/order/step time, task quality, loading peak and sustained power/thermal behavior remain unverified. The earlier Cosmos component and direct hybrid results remain separate experiments. A new 20-request native Radeon MiniCPM-o profile passed the scoped combined input/output path but exposed a large memory-sensitive latency regression versus its earlier one-request smoke; its row below supersedes the smoke for repeated-request timing.
A follow-up MiniCPM-o storage experiment copied the same ten GGUF weights to native NTFS. One warmup plus 20 complete requests passed on Windows CPU (p50/p95 23.62/23.88 s) and Radeon+CPU (16.35/17.20 s), with no prefetch warnings. Separate one-request UNC controls were much slower; the original UNC 20-request Radeon profile remains valid for that path. Run order and OS cache were not controlled, so no causal or cross-placement speedup claim follows. A real InternVLA Cosmos encoder NPU artifact attempt passed fixed-shape CPU export parity but its tested A16W8 candidates failed the CPU numerical gate or VitisAI compilation; the AMD NPU cell remains NOT E2E.
The WSL CPU Spark GGUF Omni profile adds a separate current-Omni complete-request path to the earlier vLLM INT8 CPU result. It passed two named prompts, 20/20 serial measured inventory requests at p50/p95 7.82/7.97 s, a public text request, context refusal and cancellation cleanup. The CPU-only llama.cpp build required explicit CPU-buffer placement evidence because it omits the Windows offload-count log line. This does not establish token streaming or broad quality.
On a hosted Galaxy S24 Spark attention-layer experiment, a real one-layer W8A16 QDQ artifact executed on NPU and matched prior S25 component outputs exactly on one dataset. The 100-sample layer p50/p95 was 1.109/1.205 ms; the service's 1.098 ms estimated-time field was the sample minimum. FP32-source versus S24 output hidden-state relative L2 was 1.351% without a downstream token-quality gate. This adds component C evidence only; device-local generation remains absent.
The same W8A16 layer also ran on hosted Snapdragon X Elite CRD and SA8775P ADP NPUs, with pinned-fixture outputs identical to S24/S25; the 100-sample component p50/p95 was 1.433/1.606 ms and 1.837/2.079 ms, respectively. On RB3 Gen 2 / QCS6490, that W8A16 graph failed QNN DLC loading even after target-specific recompilation. An FP32 ONNX CPU layer then passed with source-relative L2 below 9e-7, while a calibrated W8A8 NPU layer executed but had 20.9% hidden-state relative L2 and remains numerically unqualified. These hosted component outcomes do not establish full generation or comparable device speed.
A Qwen3-TTS Code2Wav follow-up recovered source and target artifacts, a retained fixture, and raw historical profiles. On SA8775P, fresh GPU-requested inference saturated 96.63% of waveform samples and differed from original ONNX CPU by 11.832 relative L2; the same binary's CPU-requested output on the exact SA device matched ONNX CPU within 4.13e-6 relative L2. The CPU-requested route also has 100 component samples at p50/p95 5.384/5.534 s, with all 738 execution-detail rows on CPU; the GPU-requested path remains numerically failed. A separate SA FP16 QNN NPU vocoder artifact compiled but failed at inference without waveform output. Fresh RB3 inference was unsaturated but differed from CPU by 15.73% relative L2 (16.06 dB SNR), without a listening-quality gate; its 100-sample mixed CPU/GPU component p50/p95 was 3.099/3.133 s. X Elite ONNX CPU vocoder inference passed on the same fixture with 8.09e-7 relative L2 versus local ONNX CPU; its 94-sample component p50/p95 was 6.384/6.644 s with all execution-detail rows on CPU. DirectML inference and profile both failed during provider initialization (Windows error 80070057); a separate FP16 QNN NPU vocoder inference returned an unsaturated waveform at 1.138% relative L2 / 38.88 dB SNR versus local ONNX CPU, but its NPU placement profile exceeded the service execution-time limit during graph preparation and returned no samples. None is a complete TTS stream.
The identical SA/RB3 TFLite binary matched original ONNX CPU within 3.45e-6 relative L2 in a separate local HX370 x86-64 LiteRT CPU control. Together with the exact SA CPU-requested parity result, this narrows the SA saturation to the GPU-requested route; the RB3 embedded CPU route and specific faulty delegate operation remain unverified.
A separate Qwen3-TTS talker decode-step attempt recovered a 28-layer ONNX export and synthetic fixed-cache fixture. Historical FP16 QNN context-binary inference executed on S24, S25, SA8775P and X Elite, but all four changed the CPU reference top-1 speech token with 8.57–9.43% logits relative L2; they are numerically unqualified on this fixture. An RB3 compile rejected floating-point input. Separate CPU-requested talker steps on X Elite ONNX and SA8775P TFLite preserved CPU token 80 with logits relative L2 2.07e-5 and 3.33e-5; X Elite had 100 component samples at p50/p95 191.414/309.428 ms with all 1,744 execution-detail rows on CPU; SA had 100 at 52.635/76.925 ms with all 1,800 rows on CPU. No complete talker loop or TTS stream follows.
A Galaxy S25 InternVLA Cosmos attempt packaged the real Place_Markpen image encoder into one ONNX file after two CPU parity fixtures. FLOAT16 QNN DLC inference returned a latent within 0.0524% relative L2 of the source on one pattern; its 100-sample six-frame component profile had p50/p95 49.665/51.382 ms and all 121 execution-detail rows on NPU. Injecting that latent into the real policy on local CPU changed synthetic actions by 0.742% relative L2, without a task tolerance. The same encoder compiled to FP16 QNN DLC for X Elite CRD and SA8775P; same-fixture NPU-requested inferences returned finite latents at 0.0608% and 0.0613% relative L2 versus source, with X Elite profiled on NPU and SA8775P placement pending. This advances component and action-sensitivity evidence, not device-local full-policy qualification. The Place_Markpen checkpoint names dataset a2d_real/a2d_pen_holder, which returned HTTP 404 under the current authenticated Hugging Face account. The official vLLM-Omni example names a different Place_Markpen archive in InternData-A1, but access to that gated dataset returned HTTP 403; the access record leaves real-observation reference-action validation open.
A MiniCPM-o speech-head study recovered the historical 20-layer source and matching S24/S25 fixtures. Both FP16 context-binary outputs preserved the ONNX CPU top-1 speech token on this one decode step with logits relative L2 of 0.346%/0.360%. S24/S25 100-sample component profiles attributed all 1,029 execution-detail rows to NPU, with p50/p95 8.537/9.006 ms and 7.197/7.615 ms. New FP16 QNN DLC compiles passed for SA8775P, X Elite and RB3. SA8775P same-fixture inference preserved CPU top-1 token 1867 with 0.3568% logits relative L2; its 100-sample component profile measured p50/p95 11.901/13.038 ms with all 944 execution-detail rows on NPU. X Elite same-fixture inference also preserved token 1867 with 0.3170% logits relative L2; its 100-sample component profile measured p50/p95 9.905/11.037 ms with all 944 execution-detail rows on NPU; RB3 FP16 QNN inference failed at graph load; its explicit TFLite CPU fallback returned 41 finite outputs, preserved CPU top-1 token 1867 and had logits relative L2 1.26e-6 versus ONNX CPU, with 100 CPU-requested component samples at p50/p95 59.518/71.347 ms and all 1,008 execution-detail rows on CPU. These are speech-head component results, not multimodal MiniCPM-o requests.
A raw sustained-window audit found that native Windows RTX Qwen3-TTS playback slows sharply after about 26 minutes: all 17 serial medium requests in minutes 27–30 underrun in the simulated first-audio playout, with median RTF 1.331. The separate WSL run stayed below real time without underruns. The cause is not established; a timestamped GPU telemetry repeat is underway.
Configuration × model
The WSL CPU Qwen3-TTS cell also has a separate bounded Omni whole-session text-to-complete-WAV profile with one warmup, 20 measured requests, a public audio request, pinned ASR proxy, memory refusal and cancellation evidence. It does not replace the earlier short-stream finding or establish playable streaming.
NOT E2E means no complete device/model workload was verified. “Scoped E2E” identifies only the named input/output path; “standalone” means the whole model ran outside Omni. The Spark GGUF Omni stage uses a complete-request control path; it has no incremental token stream or restart-after-cancel contract yet. “Synthetic policy pass” does not establish a real robot-task result. See the per-cell reasons below.
| Device | Spark-X2.5 | Qwen3-TTS 0.6B CustomVoice | Qwen3.8-27B | MiniCPM-o 4.5 | InternVLA-A1 |
|---|---|---|---|---|---|
| PC HX370: CPU / WSL | scoped E2E: Omni GGUF Spark text | scoped E2E: Omni CPU text → complete WAV | NOT E2E | scoped E2E: Omni GGUF combined image+audio → text+speech | synthetic policy pass: Omni CPU |
| PC HX370: CPU / native Windows | scoped E2E: Omni Spark text | scoped E2E: Omni CPU text → complete WAV | NOT E2E | scoped E2E: Omni GGUF image+audio → text+speech (native-storage 20-run profile) | synthetic policy pass: Omni CPU |
| PC HX370 + RTX 5090 Laptop / WSL | profile completed | profile completed | scoped E2E: text + image | scoped E2E: text/image/audio → speech | synthetic policy pass: admitted Omni CUDA |
| PC HX370 + RTX 5090 Laptop / Windows | profile completed | profile completed | NOT E2E | scoped E2E: text/image/audio → speech | synthetic policy pass: admitted Omni CUDA |
| PC HX370 + Radeon 890M / Windows worker | scoped E2E: Omni Spark text | scoped E2E: Omni Radeon talker/codec + CPU predictor | NOT E2E | scoped E2E: Omni Radeon+CPU image+audio → text+speech (native-storage 20-run profile) | synthetic policy pass: Omni CPU+Radeon Cosmos (WSL/native Windows) |
| PC HX370 + AMD NPU / Windows worker | NOT E2E (output-head component passed) | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
| PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
| PC Snapdragon X Elite CRD / AI Hub | NOT E2E (one Spark attention layer on NPU) | NOT E2E (CPU talker/vocoder steps; QNN vocoder numeric) | NOT E2E | NOT E2E (NPU speech-head step profiled) | NOT E2E (Cosmos NPU encoder profiled) |
| Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | NOT E2E (one Spark attention layer on NPU) | NOT E2E (FP16 talker token mismatch) | NOT E2E | NOT E2E (NPU speech head step profiled) | NOT E2E (Cosmos NPU encoder profiled) |
| Mobile Galaxy S24 / Snapdragon 8 Gen 3 | NOT E2E (one Spark attention layer on NPU) | NOT E2E (FP16 talker token mismatch) | NOT E2E | NOT E2E (NPU speech head step profiled) | NOT E2E |
| Embedded SA8775P ADP / AI Hub | NOT E2E (one Spark attention layer on NPU) | NOT E2E (CPU talker/vocoder steps; NPU vocoder failed) | NOT E2E | NOT E2E (NPU speech-head step profiled) | NOT E2E (Cosmos encoder numeric; placement pending) |
| Embedded RB3 Gen 2 / QCS6490 / AI Hub | NOT E2E (Spark CPU layer passed; W8A8 NPU quality open) | NOT E2E (mixed vocoder numeric unqualified) | NOT E2E | NOT E2E (CPU speech-head step profiled) | NOT E2E |
Per-cell reasons and next steps
| Device | Model | Current profiling disposition | Reason / remaining work | Next step |
|---|---|---|---|---|
| PC HX370: CPU / WSL | Spark-X2.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The pinned Spark-X2.5-1.7B Q4_K_M GGUF ran on a CPU-only Linux llama.cpp server through the current Omni 0.29 StageRuntime/StagePool whole-session backend. Both named prompts passed (Paris, 1511); one warmup plus 20/20 serial inventory requests had nearest-rank complete-request wall p50/p95 7.82/7.97 s, excluding 1.61 s startup. The stage verified model/server hashes, explicit CPU-only tensor buffers (CPU_Mapped, CPU_REPACK) and no usable GPU, reserved 4 GiB host RAM, refused an over-context request without truncation, and after in-flight cancellation returned no stale output or reservation. A separate public AsyncOmni.generate text request passed. The Linux CPU-only build omits the Windows 0/29 offload line; the verifier was extended to require positive CPU buffer evidence rather than inferring CPU from an absent GPU count. This is separate from the historical vLLM INT8 CPU profile and does not prove model-wide quality, incremental token streaming, restart-after-cancel, loading peak or sustained performance. |
Broaden text/token quality and context suite, qualify matching vLLM 0.29 CPU execution/token parity, then test streaming, restart, loading peak and sustained load; WSL GGUF Omni raw evidence. |
| PC HX370: CPU / WSL | Qwen3-TTS 0.6B CustomVoice | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The earlier vLLM 0.28 CPU path passed four short streams but had four simulated playback underruns in one measured request; its wheel remains mismatched with Omni 0.29. Separately, the pinned BF16 0.6B CustomVoice checkpoint ran two named text-to-complete-WAV requests through an Omni 0.29 whole-session stage with an isolated CPU worker on WSL. One warmup plus 20 serial measured requests returned identical 3.44 s audio; nearest-rank complete-request wall p50/p95 was 8.27/8.47 s (RTF 2.40/2.46), excluding 19.82 s startup. Pinned Whisper tiny.en transcribed the first measured WAV exactly (WER 0, one-output intelligibility proxy). Public AsyncOmni audio output passed; the stage verified weights, interpreter and CPU placement, reserved 10 GiB host RAM, refused 4 GiB before launch, and drained in-flight cancellation with zero remaining reservation. WSL PCM differs from the same-checkpoint native Windows outputs despite the declared settings; its cause is unqualified. Sampled process-tree RSS reached 3.39 GB, not a loading peak. This worker delivers only a complete WAV, not a playable stream. |
Qualify cross-OS numerical/quality difference, playable streaming without underruns, post-cancel restart, loading peak, concurrency and sustained power/thermal behavior; WSL Omni raw evidence, earlier stream evidence. |
| PC HX370: CPU / WSL | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No compatible CPU artifact/deployment for this model was found in the checked model roots. Available Qwen 27B FP8/NVFP4 CUDA artifacts do not establish a CPU route. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| PC HX370: CPU / WSL | MiniCPM-o 4.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The BF16/vLLM 0.28 three-stage path previously passed separate text, synthetic red-square image and synthetic 440 Hz audio inputs to text+speech; its 20-request serial text-only profile was p50/p95 21.86/22.97 s with swap and shutdown warnings. A separate pinned Q4_K_M/F16 GGUF worker now runs a combined spoken question+image → text+speech request through a bounded Omni StageRuntime/StagePool graph stage on WSL CPU, with installed vLLM 0.29 for the controller and CPU-only C++ model execution. All 20 measured requests after one warmup answered “The square in the image is red.”, emitted complete nonzero speech and a terminal audio event; nearest-rank complete-request wall p50/p95 was 22.67/24.83 s, speech duration 1.60–2.92 s, and maximum sampled worker RSS 13.70 GB. The first measured WAV independently transcribed exactly with pinned Whisper tiny.en (WER 0, one-output intelligibility proxy). Ten weight hashes and CPU placement were verified; 24 GiB shared RAM was reserved, 8 GiB was refused before launch, and in-flight cancellation left no stale output or reservation. Host RAM/swap samples covered only part of the profile. One synthetic fixture and variable speech PCM do not establish general quality, playable streaming, loading peak or sustained performance. | Validate real image/audio suites, speech alignment and quality, video input, concurrent admission, post-cancel restart, loading peak and sustained thermal behavior; GGUF Omni raw evidence, BF16/vLLM CPU evidence, earlier standalone GGUF run. |
| PC HX370: CPU / WSL | InternVLA-A1 | SYNTHETIC_POLICY_PASS_OMNI; NOT_TASK_E2E | The real Place_Markpen checkpoint ran through an admitted Omni StageRuntime whole-policy stage with a CPU-only isolated worker. One warmup plus 20 serial patterned synthetic camera/state/noise requests returned identical finite [1,50,32] action chunks with terminal action events; nearest-rank complete-request wall p50/p95 was 3.175/3.219 s, excluding 40.03 s startup. The stage verified checkpoint/interpreter hashes and CPU placement, reserved 16 GiB of a declared 30 GiB WSL host-RAM pool, refused 8 GiB before launch, and released the reservation on shutdown. The earlier direct zero-input policy/CUDA comparison was a different fixture and remains separate. No real-observation task-quality tolerance exists. |
Validate real observations/reference actions, physical units/order/step time, action deadlines, loading peak, concurrency and sustained power/thermal behavior; Omni raw evidence, earlier direct policy evidence. |
| PC HX370: CPU / native Windows | Spark-X2.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The pinned Spark-X2.5-1.7B Q4_K_M GGUF ran in native Windows llama.cpp CPU (-dev none -ngl 0) as a whole-session Omni StageRuntime/StagePool backend. Paris and the correct 1511 inventory answer passed; one warmup plus 20 serial complete requests were 20/20 correct, nearest-rank wall p50/p95 6.93/7.05 s. The public AsyncOmni short request passed. The stage verifies artifact hashes and CPU placement (0/29 layers offloaded), reserves 4 GiB host RAM, emits a terminal text event and releases reservation after in-flight cancellation. The separate prior standalone profile was 6.65/6.86 s. No vLLM CPU attention repair, incremental stream, restart-after-cancel, loading peak or sustained power/thermal gate is claimed. |
Validate broader text/token quality, streaming and cancellation recovery, memory peak and sustained load; Omni raw evidence, standalone evidence. |
| PC HX370: CPU / native Windows | Qwen3-TTS 0.6B CustomVoice | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The pinned 0.6B CustomVoice BF16/SDPA checkpoint ran wholly on CPU through an isolated Qwen wrapper worker owned by one bounded Omni StageRuntime stage. Two named 24 kHz WAVs matched the earlier standalone PCM SHA256 values exactly. One warmup + 20 serial complete Omni requests for 4.56 s of audio had 20/20 identical PCM and nearest-rank wall p50/p95 15.32/15.60 s (RTF 3.36/3.42); independent Whisper tiny.en ASR gave WER 0.20 on the first Omni WAV, an intelligibility proxy only. A public AsyncOmni audio request passed. The stage verified pinned weights/interpreter and CPU placement, reserved 10 GiB host RAM, rejected an insufficient 4 GiB demand before worker launch, and after in-flight cancellation emitted no stale output and returned the ledger to zero. Sampled process-tree private memory peaked at 5.39 GB, not a loading peak. Native vLLM CPU attention operators remain absent; playable streaming, post-cancel restart, broader speech quality, loading peak and sustained power/thermal behavior remain unverified. | Validate playable streaming and restart/recovery, broader voice quality, peak memory, concurrency and sustained behavior; Omni raw evidence, standalone baseline. |
| PC HX370: CPU / native Windows | Qwen3.8-27B | NOT_E2E_PROFILEABLE | The installed native Windows vLLM extension lacks cpu_attn_reshape_and_cache and cpu_attn_get_scheduler_metadata. The separate ORT CPU graph worker is not a whole-model text/TTS adapter; a qualified CPU build/binding is missing. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| PC HX370: CPU / native Windows | MiniCPM-o 4.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST; NATIVE_STORAGE_PROFILED | The pinned Q4_K_M/F16 GGUF combined synthetic JPEG+spoken-question → text+speech route passed one warmup and 20/20 serial complete requests with weights on native NTFS: nearest-rank p50/p95 23.62/23.88 s, excluding 9.23 s controller startup; every request returned the expected red-square sentence and nonzero 1.60–2.88 s 24 kHz speech. The first measured WAV exactly matched pinned Whisper tiny.en ASR (WER 0, one-output intelligibility proxy). CPU placement and ten weight hashes were verified; a 21 GiB shared-RAM demand was admitted, sampled worker RSS peaked at 13.10 GB, no worker logged a prefetch warning, and in-flight cancellation left no stale output or reservation. A same-artifact one-request control from WSL UNC weights took 79.20 s, versus 21.16 s from native weights, with 40.90 vs 12.75 s worker load and two vs zero prefetch warnings. That sequential comparison associates UNC storage with the older slowdown but does not isolate causality or prove stable cross-placement speedup. The older UNC CPU result remains valid for that path. This is a whole-request C++ worker under Omni control, not native vLLM execution or playable streaming; the synthetic fixture is not general quality qualification. | Validate real image/audio suites, playable streaming, concurrent admission, post-cancel restart, loading peak and sustained thermal/power behavior; native/UNC raw evidence, earlier UNC CPU control, earlier Omni evidence. |
| PC HX370: CPU / native Windows | InternVLA-A1 | SYNTHETIC_POLICY_PASS_OMNI; NOT_TASK_E2E | The real Place_Markpen checkpoint now runs through an admitted native Windows Omni StageRuntime whole-policy stage with the policy and output on CPU. One warmup plus 20 serial patterned synthetic camera/state/noise requests returned identical finite [1,50,32] action chunks; nearest-rank complete-request wall p50/p95 was 4.878/4.902 s, excluding 135.60 s startup. A separate public AsyncOmni.generate request returned a terminal action event with the same hash. The stage verified checkpoint/interpreter hashes, used a 16 GiB host-RAM capacity and demand within observed Windows availability, refused 8 GiB before worker launch, and after in-flight abort emitted no stale output and released the ledger. Compared with the WSL CPU Omni output for this fixture, action relative L2 difference was 2.1795% and cosine 0.999764, without a task tolerance. The earlier one-forward zero-input policy result is a different fixture. |
Validate real observations/reference actions, physical units/order/step time and deadlines, loading peak, post-cancel restart, concurrency and sustained power/thermal behavior; native Omni evidence, earlier direct policy evidence. |
| PC HX370 + RTX 5090 Laptop / WSL | Spark-X2.5 | PROFILE_COMPLETED | not fully qualified: reference-quality and remaining reliability/performance gates must be reviewed | Broaden only with a named workload/device/precision qualification. |
| PC HX370 + RTX 5090 Laptop / WSL | Qwen3-TTS 0.6B CustomVoice | PROFILE_COMPLETED | not fully qualified: reference-quality and remaining reliability/performance gates must be reviewed | Resolve and measure playback stalls; run quality, interruption and thermal tests. |
| PC HX370 + RTX 5090 Laptop / WSL | Qwen3.8-27B | SCOPED_E2E | A new one-stage Omni binding ran the real NVFP4 checkpoint: one 16-token text request and one synthetic red-image request (answer Red) at an explicit 512-token context with host offload. This is not the full multimodal, long-context or quality workload. |
Qualify image/video suites, longer contexts, latency distributions and sustained power; recovery evidence. |
| PC HX370 + RTX 5090 Laptop / WSL | MiniCPM-o 4.5 | SCOPED_E2E | The verified checkpoint completed three-stage text, one synthetic image and one synthetic audio input to text+speech under separate constrained plans with 5 GiB host offload. The image response correctly described a 448×448 red square on white and produced a nonzero 5.2 s WAV in 17.49 s; a 2 s/440 Hz tone was described as “A loud whistling sound” with a nonzero 2.16 s WAV in 7.78 s. The original 0.58 thinker GPU budget failed image/KV admission; 0.61 admitted the same context and image. Image shutdown force-killed stage 0, and both multimodal runs reported shared-memory cleanup warnings. General image/audio quality and speech meaning remain unverified. | Validate real image/audio suites, video input, speech intelligibility, interruption, repeated latency, clean shutdown and sustained power; text evidence, image/audio evidence. |
| PC HX370 + RTX 5090 Laptop / WSL | InternVLA-A1 | SYNTHETIC_POLICY_PASS_OMNI_CUDA; NOT_TASK_E2E | The real Place_Markpen checkpoint ran through an admitted Omni whole-policy graph stage with BF16 policy and Cosmos encoder on the RTX 5090 Laptop GPU. One warmup plus 20/20 serial patterned synthetic observation/state/noise requests returned identical finite float32 [1,50,32] actions with terminal events; nearest-rank complete-request wall p50/p95 was 0.382/0.399 s, excluding 43.82 s startup. The 16 GiB host-RAM and 16 GiB CUDA demands were within explicit 28/21 GiB ceilings checked against available memory. Actual CUDA placement was verified, PyTorch peak reserved VRAM reached 7.86 GB, and an 8 GiB VRAM plan was refused before worker launch. An in-flight abort and shutdown released both ledgers with no stale output; a separate public AsyncOmni.generate action request passed. Relative to a separate patterned WSL CPU output, action relative L2 difference was 2.006% with no task tolerance. The earlier one-forward zero-input direct CUDA run was a different experiment. This is synthetic action-policy integration, not robot-task E2E; units/order/step time remain unverified and control_ready=false. | Validate real observations/reference actions and physical action semantics, post-cancel restart, concurrent admission, loading peak, deadlines and sustained latency/power/thermal behavior; Omni CUDA evidence, earlier direct result. |
| PC HX370 + RTX 5090 Laptop / Windows | Spark-X2.5 | PROFILE_COMPLETED | not fully qualified: reference-quality and remaining reliability/performance gates must be reviewed | Broaden only with a named workload/device/precision qualification. |
| PC HX370 + RTX 5090 Laptop / Windows | Qwen3-TTS 0.6B CustomVoice | PROFILE_COMPLETED; SUSTAINED_PLAYBACK_FAIL | The complete-model CUDA stream passed short measured requests, but the 30-minute serial medium-prompt run slowed sharply around minute 26–27. In the final 27–30 minute window, all 17 requests had simulated playback underruns; nearest-rank p50 RTF was 1.331 and complete wall 10.379 s, versus 0.185 / 1.478 s in minutes 0–25. Avoiding all recorded late underruns required up to 4.650 s additional prebuffer. A separate WSL run on the same laptop had no sustained underruns, but power/start conditions were not controlled. The historical run lacks clock/temperature/power timeline, so the cause is unknown. | Repeat with timestamped GPU telemetry, identify and fix the late slowdown, then run speech-quality, interruption/recovery and sustained thermal gates; raw-window analysis. |
| PC HX370 + RTX 5090 Laptop / Windows | Qwen3.8-27B | NOT_E2E_PROFILEABLE | The Omni pipeline binding now exists and passed scoped text/image cases in WSL, but no native Windows Omni run has verified the NVFP4 checkpoint, memory admission or modality behavior. | Run the same checkpoint and explicit context on native Windows, then broaden quality and profiling; WSL-only evidence. |
| PC HX370 + RTX 5090 Laptop / Windows | MiniCPM-o 4.5 | SCOPED_E2E | The verified checkpoint completed native Windows three-stage BF16 text, one synthetic image, and one synthetic audio input to text+speech in separate constrained plans with 5 GiB thinker host offload and explicit 2,048/512-token limits. Text and all 32 thinker token IDs matched WSL CUDA; a 5.2 s WAV was nonzero. The red-square image produced the same description and 27 thinker token IDs as WSL plus a nonzero 5.2 s WAV in 17.15 s. The 2 s/440 Hz tone produced the same whistling description and nine IDs as WSL plus a nonzero 2.16 s WAV in 8.26 s. The initial text run needed 900/600 s startup limits after a 300 s timeout; a separate 1-warmup/20-measured serial text run had request wall p50 18.38 s/p95 18.73 s, with sampled host available RAM as low as 0.78 GB and pagefile use up to 11.17 GB. These synthetic inputs do not establish general image/audio or speech quality. | Check real image/audio suites, video input, speech intelligibility/alignment, concurrency, interruption, loading-peak admission and sustained thermal/power behavior before broader qualification; native evidence. |
| PC HX370 + RTX 5090 Laptop / Windows | InternVLA-A1 | SYNTHETIC_POLICY_PASS_OMNI_CUDA; NOT_TASK_E2E | The real Place_Markpen checkpoint ran through an admitted native Windows Omni whole-policy graph stage with BF16 policy and Cosmos encoder on the RTX 5090 Laptop GPU. After one warmup, 20/20 serial patterned synthetic requests returned identical finite float32 [1,50,32] actions with terminal events; nearest-rank complete-request wall p50/p95 was 0.746/0.806 s, excluding 147.77 s startup. The 16 GiB host-RAM and 16 GiB CUDA demands were within explicit 18/21 GiB ceilings checked against available memory. Actual CUDA placement was verified, PyTorch peak reserved VRAM reached 7.86 GB, and an 8 GiB VRAM plan was refused before worker launch. In-flight cancellation and shutdown released both ledgers with no stale output; a separate public AsyncOmni.generate request returned the same action hash. Relative to the separate native CPU patterned output, action relative L2 difference was 2.340% with no task tolerance; native-vs-WSL CUDA relative L2 was 1.555%. The earlier one-forward zero-input direct CUDA run was a different experiment. This is synthetic action-policy integration, not robot-task E2E; units/order/step time remain unverified and control_ready=false. | Validate real observations/reference actions and physical action semantics, post-cancel restart, concurrent admission, loading peak, deadlines and sustained latency/power/thermal behavior; Omni Windows CUDA evidence, earlier direct result. |
| PC HX370 + Radeon 890M / Windows worker | Spark-X2.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The same pinned Spark-X2.5-1.7B Q4_K_M GGUF completed text prefill/decode through a bounded Omni whole-session stage with llama.cpp Vulkan1. The log identifies Radeon 890M and 29/29 layers offloaded there, rather than RTX 5090; an intentionally wrong device name was refused. Paris and 1511 passed; one warmup plus 20 serial complete requests were 20/20 correct, nearest-rank wall p50/p95 2.10/2.14 s. A public AsyncOmni request passed; an oversized context was rejected with a later correct request; in-flight cancellation produced no stale output and released the ledger. The prior standalone run was 2.18/2.20 s. No streaming, restart-after-cancel, GPU-memory peak or sustained power/thermal check is claimed. |
Qualify public configuration, broader text/token quality, streaming and cancellation recovery, shared-RAM/GPU peak and sustained use; Omni raw evidence, standalone evidence. |
| PC HX370 + Radeon 890M / Windows worker | Qwen3-TTS 0.6B CustomVoice | SCOPED_E2E_OMNI_COMPLETE_REQUEST_HYBRID | The pinned Q8_0 CustomVoice talker, F16 codec and punctuation GGUFs completed two named text-to-WAV requests through an Omni whole-session CrispASR stage on native Windows. The runtime log identifies Radeon 890M Vulkan0 for talker/codec and the 432 MB code predictor pinned to CPU FP32; this is not all-Radeon or NPU execution. Both PCM hashes exactly match the earlier standalone outputs. Whisper tiny.en transcribed the first Omni WAV and both standalone WAVs exactly (WER 0 as an intelligibility proxy). One warmup plus 20 serial complete Omni requests for 2.56 s of audio had identical PCM and nearest-rank wall p50/p95 3.298/3.336 s (RTF 1.29/1.30). A public AsyncOmni audio request passed; cancellation drained the worker without stale output or remaining reservation. The earlier standalone profile was 3.310/3.345 s. The default Vulkan configuration previously fell back to CPU, all-Vulkan ran away to a 360-frame cap, and DirectML failed waveform quality. No playable stream, post-cancel restart, loading peak, broad speech quality or sustained power/thermal run is claimed. | Validate playable streaming, restart/recovery, broader voice quality, shared-RAM/GPU peak, concurrency and sustained real-time behavior before wider qualification; Omni raw evidence, standalone evidence, failed DirectML route. |
| PC HX370 + Radeon 890M / Windows worker | Qwen3.8-27B | NOT_E2E_PROFILEABLE | FP32 448x448 vision tower: DirectML 85/113 nodes; normalized max error 9.095e-06 versus torch CPU. Historical component path, not full 27B or current graph-stage image-to-text integration. | Recover the hashed tower bundle and qualify its current Omni graph handoff with the language stage. |
| PC HX370 + Radeon 890M / Windows worker | MiniCPM-o 4.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST_HYBRID; NATIVE_STORAGE_PROFILED | The pinned Q4_K_M/F16 GGUF combined synthetic spoken-question+JPEG → text+speech route passed one warmup and 20/20 serial complete requests through a bounded Omni whole-session stage with weights on native NTFS. Nearest-rank wall p50/p95 was 16.35/17.20 s, range 14.08–17.30 s, excluding 8.75 s startup; all requests returned the red-square sentence and nonzero 2.08–2.88 s 24 kHz speech with terminal audio events. The first WAV exactly matched pinned Whisper tiny.en ASR (WER 0, one-output intelligibility proxy). All 21 successful worker logs identified Radeon 890M Vulkan0 language (37/37 layers) and vision plus CPU Token2Wav, with zero prefetch warnings; sampled RSS peaked at 15.30 GB. The 21 GiB shared-RAM reservation admitted after file-cache reclaim; an initial identical attempt was explicitly refused before model load when available RAM had fallen below reservation. Cancellation left no stale output or reservation. A sequential same-artifact one-request UNC control took 90.63 s versus 16.89 s native and logged two prefetch warnings; the previous UNC 20-request profile remains p50/p95 87.27/91.45 s for that path. Storage is strongly associated with the slowdown, but OS cache/run order are uncontrolled and the separate CPU/Radeon profiles are not a qualified iGPU speedup comparison. This is CPU+iGPU joint execution, not all-Radeon, NPU, general quality, playable streaming or sustained thermal/power qualification. | Validate real multimodal quality, playable streaming, post-cancel restart, concurrency, loading peak and sustained power/thermal behavior; capture matched OS/shared-iGPU memory traces. Native/UNC raw evidence, earlier UNC 20-request evidence. |
| PC HX370 + Radeon 890M / Windows worker | InternVLA-A1 | SYNTHETIC_POLICY_PASS_HYBRID_OMNI; NOT_TASK_E2E | The real Place_Markpen checkpoint ran in admitted Omni whole-policy stages on both WSL CPU+Radeon and native Windows CPU+Radeon. In each, the BF16 Qwen3-VL/flow action policy stayed on CPU and a pinned six-frame FP32 Cosmos encoder used Omni's external DirectML worker; the selected Radeon 890M and graph output device were verified, without internal per-op attribution. Each placement returned 20/20 identical finite [1,50,32] actions after one warmup. WSL hybrid complete-request p50/p95 was 3.120/3.197 s versus its separate CPU baseline 3.175/3.219 s, with action relative L2 difference 2.305%. Native Windows hybrid was 4.575/4.725 s versus its separate CPU baseline 4.878/4.902 s, with action relative L2 difference 2.4391%. Native hybrid startup was 140.31 s versus CPU 135.60 s. These are separate one-fixture serial runs, not paired speedup or task-quality proof. FP32 encoder is an explicit precision change after DirectML BF16 aborted. Both stages verified hashes, reserved 16 GiB shared host RAM, rejected insufficient demand before launch, emitted terminal actions through public AsyncOmni.generate, and released reservations after in-flight cancellation/shutdown. An initial WSL hybrid shutdown quarantined 16 GiB; the lifecycle fix and reruns cleared it. The older direct hybrid zero-input profile is a different experiment. |
Validate real observations/reference actions, physical units/order/step time, post-cancel restart, loading peak, concurrency and sustained tail/power before deployment; WSL Omni evidence, native Windows Omni evidence, earlier direct component evidence. |
| PC HX370 + AMD NPU / Windows worker | Spark-X2.5 | COMPONENT_PASS; NOT_E2E_PROFILEABLE | The real 1.7B output head (RMS norm and full 131,072-logit projection) ran through Omni external.graph.v1 with VitisAI: one fused NPU partition and seven CPU nodes per request. The A16W8 graph matched top-1 on four captured activations (full-logit SNR 55.43–56.81 dB). One warmup plus 20 serial repeated-input component requests gave p50/p95 10.21/12.44 ms, versus 17.97/20.29 ms for the original FP32 output head on the same Omni CPU stage. NPU startup was 18.21 s versus 4.48 s for FP32 CPU, and NPU worker peak RSS was 3.52 GiB versus 2.06 GiB. No full prefill/decode/KV/sampling, real handoff, power, or sustained run was verified. |
Measure an actual vLLM-to-output-head handoff and whole-generation cost, state/cancellation, memory admission and quality before considering a split plan; component evidence. |
| PC HX370 + AMD NPU / Windows worker | Qwen3-TTS 0.6B CustomVoice | NOT_E2E_PROFILEABLE | No placement- and quality-verified whole-model graph bundle/stateful adapter for this model on this AMD route. Small backend graphs do not establish model support. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| PC HX370 + AMD NPU / Windows worker | Qwen3.8-27B | NOT_E2E_PROFILEABLE | The tested A16W8 whole-vision-tower export failed NPU placement. This conclusion is specific to that export, not all future Qwen/NPU combinations. | Validate a supported export or smaller useful component, including same-checkpoint numerical quality. |
| PC HX370 + AMD NPU / Windows worker | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | No placement- and quality-verified whole-model graph bundle/stateful adapter for this model on this AMD route. Small backend graphs do not establish model support. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| PC HX370 + AMD NPU / Windows worker | InternVLA-A1 | COMPONENT_EXPORT_CPU_PASS; NPU_ARTIFACTS_REJECTED; NOT_E2E_PROFILEABLE | The real Place_Markpen Cosmos CI8x8 encoder exported to fixed FP32 [6,3,256,256] -> [6,16,32,32] ONNX; ORT CPU matched the source on two fixtures with relative L2 8.74e-7/2.29e-6. A16W8 Conv+MatMul QDQ changed encoder output by 11.77%/5.03% on CPU and the VitisAI compiler emitted a missing-scale fatal assertion. MatMul-only A16W8 retained near-source CPU parity, but its original and ORT-preprocessed graphs each hit a native SmallVector compiler assertion. No VitisAI node executed, so this is not NPU component placement or complete InternVLA policy support. Failures are specific to these artifacts and EP 1.8.63.0. | Produce a VitisAI-supported whole-encoder or coarse substage artifact with representative calibration and numerical/policy tolerances; then verify actual NPU nodes, same-device handoff, full-policy behavior and memory/power before changing this disposition. raw export and failure evidence. |
| PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | Spark-X2.5 | NOT_E2E_PROFILEABLE | No qualifying same-model multi-accelerator deployment/artifact plan or overlap/recovery evidence. Individual routes are not a joint pipeline. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | Qwen3-TTS 0.6B CustomVoice | NOT_E2E_PROFILEABLE | No qualifying same-model multi-accelerator deployment/artifact plan or overlap/recovery evidence. Individual routes are not a joint pipeline. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No qualifying same-model multi-accelerator deployment/artifact plan or overlap/recovery evidence. Individual routes are not a joint pipeline. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | No qualifying same-model multi-accelerator deployment/artifact plan or overlap/recovery evidence. Individual routes are not a joint pipeline. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | InternVLA-A1 | NOT_E2E_PROFILEABLE | No qualifying same-model multi-accelerator deployment/artifact plan or overlap/recovery evidence. Individual routes are not a joint pipeline. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| PC Snapdragon X Elite CRD / AI Hub | Spark-X2.5 | COMPONENT_NPU_INFERENCE_PROFILED; NOT_E2E_PROFILEABLE | A real Spark-X2.5-1.7B one-layer W8A16 QDQ attention artifact, compiled as QNN DLC, ran on the exact hosted Snapdragon X Elite CRD (Windows 11 ARM64). Inference returned finite hidden/new-K/new-V tensors bitwise identical to the prior S24/S25 NPU outputs on the same fixture; the profile attributed all 73 execution-detail rows to NPU. Across 100 component samples, nearest-rank p50/p95 was 1.433/1.606 ms; the 1.346 ms estimated-time field is the minimum. Hidden/K/V relative L2 versus original FP32 ONNX CPU was 1.351%/0.377%/0.650%, without a next-token quality tolerance. There is no complete 28-layer prefill/decode/KV-ring/sampling loop, resident Omni controller or complete-request timing. | Qualify a complete X Elite resident artifact/controller with state, token quality, memory admission/cancel and sustained power; X Elite NPU raw evidence. |
| PC Snapdragon X Elite CRD / AI Hub | Qwen3-TTS 0.6B CustomVoice | COMPONENT_ONNX_CPU_VOCODER_PARITY_PROFILED; DIRECTML_RUNTIME_FAILED; TALKER_FP16_NUMERIC_FAIL; QNN_NPU_VOCODER_NUMERIC_MEASURED_PROFILE_TIMEOUT; TALKER_ONNX_CPU_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | The fresh ONNX vocoder inference on the exact Windows 11 ARM64 X Elite CRD requested CPU and returned an unsaturated waveform on the pinned fixture, matching local ONNX Runtime CPU with 8.09e-7 relative L2 and 121.84 dB SNR. Its 94-sample component profile had nearest-rank p50/p95 6.384/6.644 s and all 634 execution-detail rows on CPU; 5.917 s estimated time is the sample minimum. This is one component, without a listening-quality gate. Corrected DirectML inference and profile both failed at provider initialization with Windows error 80070057. A separate FP16 QNN DLC vocoder inference returned an unsaturated waveform at 1.138% relative L2 / 38.88 dB SNR versus local ONNX CPU on the same fixture; the NPU node-placement profile timed out during graph preparation with no timing or per-node samples, and no listening-quality tolerance is set. The 28-layer FP16 QNN talker step executed, yet changed CPU top-1 token 80 to 1720 with 8.57% logits relative L2 on a synthetic cache fixture; a separate ONNX CPU-requested talker step preserved CPU token 80 with logits relative L2 2.07e-5 and maximum K/V relative L2 2.18e-5. Its 100-sample component profile had p50/p95 191.414/309.428 ms and all 1,744 execution-detail rows on CPU. The older TFLite vocoder target was rejected on Windows ARM; no complete stream exists. | Reduce or partition the QNN vocoder profile graph and audit CPU talker placement; investigate DirectML failure, then qualify complete local streaming, memory and sustained behavior; vocoder and talker evidence. |
| PC Snapdragon X Elite CRD / AI Hub | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| PC Snapdragon X Elite CRD / AI Hub | MiniCPM-o 4.5 | COMPONENT_SPEECH_HEAD_NPU_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | The real 20-layer MiniCPM-o speech-head FP16 QNN DLC ran on the exact X Elite CRD. Same-fixture inference returned 41 finite tensors, preserved ONNX CPU top-1 speech token 1867, and had 0.3170% logits relative L2 and 1.221% maximum K/V relative L2. Its 100-sample component profile had nearest-rank p50/p95 9.905/11.037 ms, with all 944 execution-detail rows on NPU. Reported inference peak memory was 79,208,448 bytes and first-load peak 1,081,573,376 bytes. Thinker/encoders/vocoder and a device-local stateful Omni backend remain unqualified. | Validate the complete multimodal pipeline, state, quality, co-resident memory and sustained behavior; X Elite raw evidence. |
| PC Snapdragon X Elite CRD / AI Hub | InternVLA-A1 | COMPONENT_COSMOS_NPU_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | The real Place_Markpen Cosmos FP16 QNN image encoder returned a finite FP32 latent for the six-frame synthetic pattern on exact X Elite CRD. Against the pinned source, relative L2 was 0.0608%, SNR 64.32 dB and cosine 0.999999815. The 100-sample six-frame encoder profile had p50/p95 82.939/83.837 ms, with all 121 execution-detail rows on NPU. Reported component inference peak memory was 24,059,904 bytes and first-load peak 692,891,648 bytes. No real-observation action tolerance, complete policy or local stateful Omni backend exists. | Validate real observations, action units/timing, state, admission, cancellation and sustained behavior; X Elite Cosmos raw evidence. |
| Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | Spark-X2.5 | COMPONENT_NPU_INFERENCE_PROFILED; NOT_E2E_PROFILEABLE | The real 1.7B checkpoint one-layer W8A16 full-attention decoder artifact executed for a fixed 1,024-token cache bucket. Same-fixture outputs matched S24; the hosted S25 profile had 100 component samples at nearest-rank p50/p95 0.957/1.028 ms with all 73 execution-detail rows on NPU. FP32-source hidden-state relative L2 was 1.351% without a downstream token-quality gate. No full 28-layer prefill/decode, KV/ring, sampling or resident controller ran on S25. | Implement and validate local 128-token generation, 512/1024 transitions, cancellation, resident memory and sustained power; S25/S24 component evidence. |
| Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | Qwen3-TTS 0.6B CustomVoice | TALKER_FP16_NUMERIC_FAIL; NOT_E2E_PROFILEABLE | Fresh FP16 predictor inference passed; its one output exactly matches the same compiled artifact/dataset reference. The real 28-layer FP16 QNN talker step executed on the synthetic cache fixture, but changed CPU top-1 speech token 80 to 412 with 8.99% logits relative L2 and maximum new-K/V relative L2 11.71%. This tested talker artifact is numerically unqualified; no complete co-resident NPU/GPU stream exists. | Improve and revalidate the talker artifact, then qualify the device-local predictor/vocoder pipeline, output quality, recovery and 30-minute thermal behavior. |
| Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | MiniCPM-o 4.5 | COMPONENT_FP16_SPEECH_HEAD_NPU_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | Fresh FP16 speech-head inference passed; all 41 outputs exactly matched the same artifact/dataset reference. Against the recovered original ONNX CPU source on the same fixture, logits relative L2 was 0.3605%, maximum new-K/V relative L2 1.311%, and top-1 speech token 1867 matched. Historical FP16 parity passed; tested W8A16 top-1 parity failed. The 100-sample component profile had p50/p95 7.197/7.615 ms, with all 1,029 execution-detail rows on NPU and reported inference peak memory 185,991,168 bytes. Generated-audio quality and qualified thinker/encoders/vocoder remain unverified. | Qualify a complete mobile checkpoint/export and the full multimodal/stateful pipeline; speech-head evidence. |
| Mobile Galaxy S25 / Snapdragon 8 Elite for Galaxy | InternVLA-A1 | COMPONENT_COSMOS_NPU_NUMERIC_PROFILED; SYNTHETIC_ACTION_SENSITIVITY_MEASURED; NOT_E2E_PROFILEABLE | The real Place_Markpen Cosmos image encoder FP32 ONNX export matched its source on two six-frame fixtures. Galaxy S25 FP16 QNN DLC inference returned a finite latent within 0.0524% relative L2 / 65.61 dB SNR of the pinned pattern source. Injecting this device latent into the same real policy on local CPU changed synthetic [1,50,32] actions by 0.742% relative L2; the CPU source repeated bitwise identically. The 100-sample six-frame encoder profile had nearest-rank p50/p95 49.665/51.382 ms, all 121 execution-detail rows on NPU, and a reported inference peak of 158,232,576 bytes. No real-observation task tolerance exists, and the action policy was not resident on S25. ADB has no attached device. |
Extend the S25 encoder component, then validate real action quality, device-local state, admission, cancellation and sustained behavior. |
| Mobile Galaxy S24 / Snapdragon 8 Gen 3 | Spark-X2.5 | COMPONENT_NPU_INFERENCE_PROFILED; NOT_E2E_PROFILEABLE | The real 1.7B checkpoint's fixed-shape full-attention decoder one-layer W8A16 QDQ artifact compiled as QNN DLC and ran on an exact hosted Galaxy S24 (Android 14). A same-dataset inference returned finite hidden/new-K/new-V outputs identical to the prior S25 NPU outputs; the profile attributed all 73 execution-detail rows to NPU. Across 100 component samples, nearest-rank p50/p95 was 1.109/1.205 ms; the 1.098 ms estimated_inference_time field is the minimum, not a median. Relative L2 versus original FP32 ONNX CPU was 1.351%/0.377%/0.650% for hidden/K/V; no token-quality tolerance is known. This is one cache bucket, not 28-layer prefill/decode, KV-ring, sampling, a resident local controller or complete-request timing. ADB has no attached device. |
Qualify a complete device-resident artifact/controller with state and token-quality gates, 128-token generation across 512/1024 context transitions, admission/cancel and sustained memory/power; S24 NPU raw evidence. |
| Mobile Galaxy S24 / Snapdragon 8 Gen 3 | Qwen3-TTS 0.6B CustomVoice | TALKER_FP16_NUMERIC_FAIL; NOT_E2E_PROFILEABLE | The real 28-layer FP16 QNN talker step executed on the synthetic cache fixture, but changed CPU top-1 speech token 80 to 1720 with 9.33% logits relative L2 and maximum new-K/V relative L2 11.23%. This differs from an earlier FP16 decoder parity test (39.13 dB SNR) on a different component/fixture; the tested INT8 decoder parity failed with negative SNR. No full local TTS stream exists. | Resolve the talker numerical difference, then validate the complete local pipeline and a quality-passing artifact set. |
| Mobile Galaxy S24 / Snapdragon 8 Gen 3 | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| Mobile Galaxy S24 / Snapdragon 8 Gen 3 | MiniCPM-o 4.5 | COMPONENT_FP16_SPEECH_HEAD_NPU_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | The retained FP16 speech-head job used a dataset with identical tensors to the recovered S25 fixture. Against original ONNX CPU, logits relative L2 was 0.3459%, maximum new-K/V relative L2 1.235%, and top-1 speech token 1867 matched. Earlier parity reported 49.22 dB SNR/top-1 match; tested W8A16 top-1 failed. The 100-sample component profile had p50/p95 8.537/9.006 ms, with all 1,029 execution-detail rows on NPU and reported inference peak memory 165,433,344 bytes. Generated-audio quality and the rest of the model remain unqualified. | Qualify thinker/encoders/vocoder and stateful local integration; speech-head evidence. |
| Mobile Galaxy S24 / Snapdragon 8 Gen 3 | InternVLA-A1 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| Embedded SA8775P ADP / AI Hub | Spark-X2.5 | COMPONENT_NPU_INFERENCE_PROFILED; NOT_E2E_PROFILEABLE | A real Spark-X2.5-1.7B one-layer W8A16 QDQ attention artifact, compiled as QNN DLC, ran on the exact hosted SA8775P ADP (Android 14). Inference returned finite hidden/new-K/new-V tensors bitwise identical to the prior S24/S25 NPU outputs on the same fixture; the profile attributed all 73 execution-detail rows to NPU. Across 100 component samples, nearest-rank p50/p95 was 1.837/2.079 ms; the 1.749 ms estimated-time field is the minimum. Hidden/K/V relative L2 versus original FP32 ONNX CPU was 1.351%/0.377%/0.650%, without a next-token quality tolerance. There is no complete 28-layer prefill/decode/KV-ring/sampling loop, resident Omni controller or complete-request timing. | Qualify a complete embedded resident artifact/controller with state, token quality, memory admission/cancel and sustained power; SA8775P NPU raw evidence. |
| Embedded SA8775P ADP / AI Hub | Qwen3-TTS 0.6B CustomVoice | COMPONENT_CPU_VOCODER_PARITY_AND_PLACEMENT_PASS; TALKER_TFLITE_CPU_NUMERIC_PROFILED; QNN_NPU_VOCODER_INFERENCE_FAILED; COMPONENT_GPU_NUMERIC_FAIL; TALKER_FP16_NUMERIC_FAIL; NOT_E2E_PROFILEABLE | On the exact SA8775P, the same TFLite binary's CPU-requested Code2Wav inference matched original ONNX CPU within 4.13e-6 relative L2 / 107.67 dB SNR, without saturation. Its 100-sample CPU-requested component profile had p50/p95 5.384/5.534 s with all 738 execution-detail rows on CPU; the service estimated-time field 5.250 s is the sample minimum. The GPU-requested output saturated 96.63% of samples with 11.832 relative L2; its fresh 100-sample mixed CPU/GPU profile p50/p95 was 1.776/1.880 s, timing failed audio. Even with FP32-preserving GPU options, 84.89% of samples saturated with 11.389 relative L2. The 28-layer FP16 QNN talker step executed but changed CPU top-1 token 80 to 1720 with 9.43% logits relative L2. A separate TFLite CPU-requested talker step preserved CPU top-1 token 80 with logits relative L2 3.33e-5 and maximum K/V relative L2 3.41e-5. Its 100-sample component profile had p50/p95 52.635/76.925 ms with all 1,800 execution-detail rows on CPU. A separate FP16 QNN DLC vocoder compile passed, but same-fixture NPU-requested inference failed after compilation without a waveform or placement samples. No complete pipeline exists. | Isolate the FP16 QNN vocoder runtime failure, GPU delegate failure and FP16 talker numerical drift, then establish listening quality, co-resident streaming, memory and sustained behavior; vocoder and talker evidence. |
| Embedded SA8775P ADP / AI Hub | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| Embedded SA8775P ADP / AI Hub | MiniCPM-o 4.5 | COMPONENT_SPEECH_HEAD_NPU_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | The real 20-layer MiniCPM-o speech-head FP16 QNN DLC ran on the exact SA8775P ADP. Same-fixture inference returned 41 finite tensors, preserved ONNX CPU top-1 token 1867, and had 0.3568% logits relative L2 and 1.225% maximum K/V relative L2. Its 100-sample component profile had nearest-rank p50/p95 11.901/13.038 ms, with all 944 execution-detail rows on NPU. Reported inference peak memory was 134,696,960 bytes and first-load peak 1,322,835,968 bytes; no complete thinker/encoders/vocoder bundle or device-local stateful controller exists. | Validate complete multimodal generation, state, speech quality, co-resident memory, cancellation and sustained power; SA8775P raw evidence. |
| Embedded SA8775P ADP / AI Hub | InternVLA-A1 | COMPONENT_COSMOS_NUMERIC_MEASURED; PLACEMENT_PENDING; NOT_E2E_PROFILEABLE | The real Place_Markpen Cosmos FP16 QNN image encoder returned a finite FP32 latent for the six-frame synthetic pattern on exact SA8775P ADP. Against the pinned source, relative L2 was 0.0613%, SNR 64.25 dB and cosine 0.999999812. The NPU-requested route awaits actual node-placement/timing profile. No real-observation action tolerance, complete policy or local stateful Omni backend exists. | Audit the SA8775P Cosmos placement profile, then validate real observations, action units/timing, state, admission, cancellation and sustained behavior. |
| Embedded RB3 Gen 2 / QCS6490 / AI Hub | Spark-X2.5 | COMPONENT_CPU_PARITY_PASS; COMPONENT_NPU_NUMERIC_UNQUALIFIED; NOT_E2E_PROFILEABLE | On the exact hosted Dragonwing RB3 Gen 2 Vision Kit (QCS6490, QC Linux 1.6), the reused S24 W8A16 QNN DLC and a separately RB3-compiled W8A16 DLC each failed inference/profile at QnnModel_composeGraphsFromDlc: MODEL_GRAPH_ERROR. An FP32 ONNX one-layer CPU route passed with all 56 profile rows on CPU, source-relative hidden/K/V L2 of 8.73e-7/2.98e-7/3.37e-7 and 100-sample component p50/p95 19.304/21.113 ms. A calibrated W8A8 QNN DLC ran with all 61 rows on NPU and p50/p95 5.610/5.761 ms, but hidden/K/V relative L2 was 20.914%/1.702%/4.413% versus FP32 CPU, without a next-token quality gate. These are different precisions, not a qualified speedup; no complete 28-layer prefill/decode/KV-ring/sampling, local Omni controller or complete-request timing exists. | Improve representative W8A8 calibration or find another compatible quality-passing artifact; validate token parity, then full device-resident generation, admission/cancel and sustained power; RB3 raw evidence. |
| Embedded RB3 Gen 2 / QCS6490 / AI Hub | Qwen3-TTS 0.6B CustomVoice | COMPONENT_MIXED_VOCODER_PROFILED_NUMERIC_UNQUALIFIED; NOT_E2E_PROFILEABLE | Fresh same-fixture TFLite GPU-requested vocoder inference returned a finite unsaturated waveform, but relative L2 versus original ONNX CPU was 15.73% (16.06 dB SNR), without a listening-quality tolerance; it is not a qualified audio path. The fresh 100-sample component profile had nearest-rank p50/p95 3.099/3.133 s and 403 CPU plus 335 GPU execution-detail rows. Its 3.070 s estimated-time field is the sample minimum, not median. The historical NPU talker build was rejected for floating-point inputs; no full model pipeline exists. | Establish waveform quality and a compatible integer talker artifact, then validate complete local execution, memory and sustained behavior; vocoder evidence. |
| Embedded RB3 Gen 2 / QCS6490 / AI Hub | Qwen3.8-27B | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
| Embedded RB3 Gen 2 / QCS6490 / AI Hub | MiniCPM-o 4.5 | COMPONENT_SPEECH_HEAD_FP16_QNN_LOAD_FAILED; CPU_FALLBACK_NUMERIC_PROFILED; NOT_E2E_PROFILEABLE | The real 20-layer MiniCPM-o speech-head FP16 QNN DLC compiled for exact RB3 Gen 2, but same-fixture NPU-requested inference failed at graph load with QnnModel_composeGraphsFromDlc: MODEL_GRAPH_ERROR. RB3 device attributes do not advertise FP16 HTP support. An explicit TFLite CPU-requested fallback returned 41 finite tensors, preserved ONNX CPU top-1 token 1867, and had logits relative L2 1.26e-6 and maximum K/V relative L2 1.58e-6. Its 100-sample component profile had p50/p95 59.518/71.347 ms with all 1,008 execution-detail rows on CPU; reported inference peak memory was 820,981,760 bytes. No complete model bundle or local stateful controller exists. | Qualify complete multimodal generation and sustained behavior; RB3 speech-head CPU evidence. |
| Embedded RB3 Gen 2 / QCS6490 / AI Hub | InternVLA-A1 | NOT_E2E_PROFILEABLE | No complete qualified target artifact and device-local stateful Omni backend. Hub jobs provide component execution, not an on-device generation/streaming controller; ADB has no attached device. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
Measured request timing
Nearest-rank p50/p95; warmups and sustained requests excluded. Values are observations under recorded power/cache conditions, not latency targets. Spark CPU uses 1.7B INT8; CUDA uses 4B BF16, so these are not same-model acceleration comparisons.
PC HX370: CPU / WSL — Spark-X2.5
Status: completed. Full timing protocol complete: True. Startup: 20.767 s. Sustained run: 1805.768 s.
| Length | Concurrent requests | n | Metric | p50 | p95 |
|---|---|---|---|---|---|
| long | 1 | 20 | ttft_s | 3.377 | 3.668 |
| long | 1 | 20 | wall_s | 11.838 | 15.504 |
| long | 1 | 20 | decode_tok_per_s | 14.923 | 15.631 |
| long | 2 | 20 | ttft_s | 3.630 | 6.942 |
| long | 2 | 20 | wall_s | 15.120 | 17.792 |
| long | 2 | 20 | decode_tok_per_s | 11.509 | 15.800 |
| long | 4 | 20 | ttft_s | 8.034 | 14.609 |
| long | 4 | 20 | wall_s | 22.493 | 26.072 |
| long | 4 | 20 | decode_tok_per_s | 8.815 | 15.010 |
| medium | 1 | 20 | ttft_s | 0.762 | 0.801 |
| medium | 1 | 20 | wall_s | 8.830 | 9.446 |
| medium | 1 | 20 | decode_tok_per_s | 15.663 | 16.069 |
| medium | 2 | 20 | ttft_s | 0.824 | 1.584 |
| medium | 2 | 20 | wall_s | 9.409 | 10.202 |
| medium | 2 | 20 | decode_tok_per_s | 15.142 | 16.564 |
| medium | 4 | 20 | ttft_s | 1.618 | 3.173 |
| medium | 4 | 20 | wall_s | 11.504 | 11.861 |
| medium | 4 | 20 | decode_tok_per_s | 13.201 | 15.387 |
| short | 1 | 20 | ttft_s | 0.148 | 0.175 |
| short | 1 | 20 | wall_s | 8.234 | 9.761 |
| short | 1 | 20 | decode_tok_per_s | 15.704 | 16.269 |
| short | 2 | 20 | ttft_s | 0.232 | 0.301 |
| short | 2 | 20 | wall_s | 7.996 | 10.014 |
| short | 2 | 20 | decode_tok_per_s | 16.237 | 16.804 |
| short | 4 | 20 | ttft_s | 0.447 | 0.571 |
| short | 4 | 20 | wall_s | 8.661 | 10.261 |
| short | 4 | 20 | decode_tok_per_s | 15.451 | 16.508 |
PC HX370 + RTX 5090 Laptop / WSL — Spark-X2.5
Status: completed. Full timing protocol complete: True. Startup: 23.693 s. Sustained run: 1802.602 s.
| Length | Concurrent requests | n | Metric | p50 | p95 |
|---|---|---|---|---|---|
| long | 1 | 20 | ttft_s | 0.363 | 0.386 |
| long | 1 | 20 | wall_s | 2.987 | 3.163 |
| long | 1 | 20 | decode_tok_per_s | 48.210 | 50.041 |
| long | 2 | 20 | ttft_s | 0.457 | 0.828 |
| long | 2 | 20 | wall_s | 3.566 | 3.982 |
| long | 2 | 20 | decode_tok_per_s | 41.427 | 45.763 |
| long | 4 | 20 | ttft_s | 0.848 | 1.430 |
| long | 4 | 20 | wall_s | 4.175 | 4.354 |
| long | 4 | 20 | decode_tok_per_s | 38.124 | 45.331 |
| medium | 1 | 20 | ttft_s | 0.090 | 0.103 |
| medium | 1 | 20 | wall_s | 2.810 | 2.987 |
| medium | 1 | 20 | decode_tok_per_s | 46.545 | 48.369 |
| medium | 2 | 20 | ttft_s | 0.097 | 0.192 |
| medium | 2 | 20 | wall_s | 2.847 | 3.007 |
| medium | 2 | 20 | decode_tok_per_s | 46.499 | 48.199 |
| medium | 4 | 20 | ttft_s | 0.185 | 0.369 |
| medium | 4 | 20 | wall_s | 3.031 | 3.065 |
| medium | 4 | 20 | decode_tok_per_s | 44.901 | 47.283 |
| short | 1 | 20 | ttft_s | 0.035 | 0.045 |
| short | 1 | 20 | wall_s | 2.552 | 2.703 |
| short | 1 | 20 | decode_tok_per_s | 50.203 | 51.126 |
| short | 2 | 20 | ttft_s | 0.047 | 0.064 |
| short | 2 | 20 | wall_s | 2.576 | 2.604 |
| short | 2 | 20 | decode_tok_per_s | 50.216 | 50.462 |
| short | 4 | 20 | ttft_s | 0.065 | 0.069 |
| short | 4 | 20 | wall_s | 2.655 | 2.675 |
| short | 4 | 20 | decode_tok_per_s | 48.944 | 49.154 |
PC HX370 + RTX 5090 Laptop / WSL — Qwen3-TTS 0.6B CustomVoice
Status: completed. Full timing protocol complete: True. Startup: 345.168 s. Sustained run: 1801.183 s.
| Length | Concurrent requests | n | Metric | p50 | p95 |
|---|---|---|---|---|---|
| long | 1 | 20 | ttfa_ms | 68.518 | 100.566 |
| long | 1 | 20 | rtf_total | 0.176 | 0.183 |
| long | 1 | 20 | stall_at_ttfa_ms | 0.000 | 0.000 |
| long | 2 | 20 | ttfa_ms | 79.379 | 138.556 |
| long | 2 | 20 | rtf_total | 0.215 | 0.217 |
| long | 2 | 20 | stall_at_ttfa_ms | 0.000 | 0.000 |
| long | 4 | 20 | ttfa_ms | 117.755 | 144.011 |
| long | 4 | 20 | rtf_total | 0.259 | 0.274 |
| long | 4 | 20 | stall_at_ttfa_ms | 0.000 | 0.000 |
| medium | 1 | 20 | ttfa_ms | 73.956 | 111.604 |
| medium | 1 | 20 | rtf_total | 0.175 | 0.182 |
| medium | 1 | 20 | stall_at_ttfa_ms | 0.000 | 0.000 |
| medium | 2 | 20 | ttfa_ms | 82.497 | 120.043 |
| medium | 2 | 20 | rtf_total | 0.219 | 0.224 |
| medium | 2 | 20 | stall_at_ttfa_ms | 0.000 | 0.000 |
| medium | 4 | 20 | ttfa_ms | 127.905 | 160.158 |
| medium | 4 | 20 | rtf_total | 0.274 | 0.285 |
| medium | 4 | 20 | stall_at_ttfa_ms | 0.000 | 0.000 |
| short | 1 | 20 | ttfa_ms | 67.583 | 91.833 |
| short | 1 | 20 | rtf_total | 0.203 | 0.215 |
| short | 1 | 20 | stall_at_ttfa_ms | 0.000 | 0.000 |
| short | 2 | 20 | ttfa_ms | 95.470 | 120.562 |
| short | 2 | 20 | rtf_total | 0.246 | 0.255 |
| short | 2 | 20 | stall_at_ttfa_ms | 0.000 | 0.000 |
| short | 4 | 20 | ttfa_ms | 118.914 | 145.742 |
| short | 4 | 20 | rtf_total | 0.302 | 0.323 |
| short | 4 | 20 | stall_at_ttfa_ms | 0.000 | 0.000 |
PC HX370 + RTX 5090 Laptop / Windows — Spark-X2.5
Status: completed. Full timing protocol complete: True. Startup: 95.155 s. Sustained run: 1803.249 s.
| Length | Concurrent requests | n | Metric | p50 | p95 |
|---|---|---|---|---|---|
| long | 1 | 20 | ttft_s | 0.363 | 0.415 |
| long | 1 | 20 | wall_s | 3.045 | 3.986 |
| long | 1 | 20 | decode_tok_per_s | 47.550 | 48.902 |
| long | 2 | 20 | ttft_s | 0.412 | 0.738 |
| long | 2 | 20 | wall_s | 3.383 | 3.497 |
| long | 2 | 20 | decode_tok_per_s | 44.115 | 48.104 |
| long | 4 | 20 | ttft_s | 0.751 | 1.375 |
| long | 4 | 20 | wall_s | 4.009 | 4.211 |
| long | 4 | 20 | decode_tok_per_s | 39.814 | 46.459 |
| medium | 1 | 20 | ttft_s | 0.100 | 0.129 |
| medium | 1 | 20 | wall_s | 2.827 | 3.537 |
| medium | 1 | 20 | decode_tok_per_s | 44.656 | 48.311 |
| medium | 2 | 20 | ttft_s | 0.122 | 0.250 |
| medium | 2 | 20 | wall_s | 3.595 | 4.379 |
| medium | 2 | 20 | decode_tok_per_s | 36.433 | 46.294 |
| medium | 4 | 20 | ttft_s | 0.189 | 0.367 |
| medium | 4 | 20 | wall_s | 3.034 | 3.138 |
| medium | 4 | 20 | decode_tok_per_s | 44.989 | 47.472 |
| short | 1 | 20 | ttft_s | 0.052 | 0.097 |
| short | 1 | 20 | wall_s | 2.677 | 3.256 |
| short | 1 | 20 | decode_tok_per_s | 48.084 | 50.892 |
| short | 2 | 20 | ttft_s | 0.058 | 0.145 |
| short | 2 | 20 | wall_s | 2.652 | 2.816 |
| short | 2 | 20 | decode_tok_per_s | 48.831 | 49.710 |
| short | 4 | 20 | ttft_s | 0.068 | 0.106 |
| short | 4 | 20 | wall_s | 2.642 | 2.710 |
| short | 4 | 20 | decode_tok_per_s | 49.173 | 50.890 |
PC HX370 + RTX 5090 Laptop / Windows — Qwen3-TTS 0.6B CustomVoice
Status: completed. Full timing protocol complete: True. Startup: 79.792 s. Sustained run: 1804.298 s.
| Length | Concurrent requests | n | Metric | p50 | p95 |
|---|---|---|---|---|---|
| long | 1 | 20 | ttfa_ms | 98.490 | 311.112 |
| long | 1 | 20 | rtf_total | 0.175 | 0.192 |
| long | 1 | 20 | stall_at_ttfa_ms | 0.000 | 118.271 |
| long | 2 | 20 | ttfa_ms | 131.331 | 344.353 |
| long | 2 | 20 | rtf_total | 0.207 | 0.219 |
| long | 2 | 20 | stall_at_ttfa_ms | 0.000 | 105.641 |
| long | 4 | 20 | ttfa_ms | 228.525 | 299.547 |
| long | 4 | 20 | rtf_total | 0.256 | 0.263 |
| long | 4 | 20 | stall_at_ttfa_ms | 0.000 | 118.746 |
| medium | 1 | 20 | ttfa_ms | 111.521 | 302.522 |
| medium | 1 | 20 | rtf_total | 0.178 | 0.210 |
| medium | 1 | 20 | stall_at_ttfa_ms | 0.000 | 40.773 |
| medium | 2 | 20 | ttfa_ms | 128.019 | 255.811 |
| medium | 2 | 20 | rtf_total | 0.224 | 0.246 |
| medium | 2 | 20 | stall_at_ttfa_ms | 0.000 | 77.119 |
| medium | 4 | 20 | ttfa_ms | 158.285 | 265.148 |
| medium | 4 | 20 | rtf_total | 0.281 | 0.296 |
| medium | 4 | 20 | stall_at_ttfa_ms | 73.005 | 75.003 |
| short | 1 | 20 | ttfa_ms | 103.066 | 206.838 |
| short | 1 | 20 | rtf_total | 0.203 | 0.260 |
| short | 1 | 20 | stall_at_ttfa_ms | 0.000 | 44.172 |
| short | 2 | 20 | ttfa_ms | 123.335 | 137.089 |
| short | 2 | 20 | rtf_total | 0.258 | 0.292 |
| short | 2 | 20 | stall_at_ttfa_ms | 0.000 | 134.419 |
| short | 4 | 20 | ttfa_ms | 161.902 | 299.999 |
| short | 4 | 20 | rtf_total | 0.291 | 0.327 |
| short | 4 | 20 | stall_at_ttfa_ms | 0.000 | 120.313 |