Skip to content
m.recipes
Find a configuration
Sign in
32 GB · 53 reports

vGPU-32GB · 32 GB

Qualified results first; original reports and full tables stay below.
Reset

Speech synthesis · Qualified results

Speeds compare only within one exact test. Recipe availability is checked on each report against the published recipe executor.

  1. Kokoro-82M

    hexgrad · 3 tested configurations

    • fp32 · Text → Audio · kokoro-english-v1 · Test duration 6 s
      Throughput · Batch 1
      76.7audio-sec/sec
      Best measured · Batch 2
      79.8audio-sec/sec
      Observed GPU occupancy
      1GiB
      P95 latency
      45 ms
      Evidence report and run availability ↗Same test on other GPUs
      Test details · 2 batch sizes

      Occupancy includes engine buffers and reserved KV cache. It is not the minimum VRAM required by the model.

      Reports combined counts reports, not inference requests. Request successes and failures are listed in each report.

      Engine
      kokoro · PyTorch 2.13.0+cu130
      Protocol
      audio-native-v1
      Dataset
      kokoro-english-v1
      Workload
      tts
      Batch sizeMean throughput (audio-sec/sec)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
      B176.7245 ms—183.51d90a5c29 ↗
      B279.84100 ms—1.03100.81d90a5c29 ↗
    • 2 other test configurations on this GPU
      • fp32 · Text → Audio · kokoro-english-v1 · Test duration 6 s
        Throughput · Batch 1
        40.4audio-sec/sec
        Best measured · Batch 4
        41.8audio-sec/sec
        Observed GPU occupancy
        1GiB
        P95 latency
        104 ms
        Evidence report and run availability ↗Same test on other GPUs
        Test details · 2 batch sizes

        Occupancy includes engine buffers and reserved KV cache. It is not the minimum VRAM required by the model.

        Reports combined counts reports, not inference requests. Request successes and failures are listed in each report.

        Engine
        kokoro · PyTorch 2.13.0+cu130
        Protocol
        audio-native-v1
        Dataset
        kokoro-english-v1
        Workload
        tts
        Batch sizeMean throughput (audio-sec/sec)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
        B140.36104 ms—0.9960.51c0488108 ↗
        B441.8426 ms—1.0365.91c0488108 ↗
      • fp32 · Text → Audio · kokoro-english-v1 · Test duration 6 s
        Throughput · Batch 1
        43.4audio-sec/sec
        Best measured · Batch 4
        44.9audio-sec/sec
        Observed GPU occupancy
        1GiB
        P95 latency
        101 ms
        Evidence report and run availability ↗Same test on other GPUs
        Test details · 2 batch sizes

        Occupancy includes engine buffers and reserved KV cache. It is not the minimum VRAM required by the model.

        Reports combined counts reports, not inference requests. Request successes and failures are listed in each report.

        Engine
        kokoro
        Protocol
        audio-native-v1
        Dataset
        kokoro-english-v1
        Workload
        tts
        Batch sizeMean throughput (audio-sec/sec)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
        B143.43101 ms—0.9964.11d8e4b3ce ↗
        B444.88396 ms—1.0369.11d8e4b3ce ↗

Occupancy includes engine buffers and reserved KV cache. It is not the minimum VRAM required by the model.

Sortable tables by exact test condition · 6 rows

C: concurrency · B: batch · —: not measured

Speech synthesis · Text → Audio · kokoro-english-v1 · Test duration 6 s · Batch 1 · audio-sec/sec · 1 rows
ModelFormatEngineWorkloadConcurrency / batchMean throughputBest measuredTTFT (ms)P95 (s)Observed occupancy (GiB)GPU power (W)Lowest measuredThroughput / WReports combined
hexgrad/Kokoro-82MScaling curves ↗fp32kokorottsaudio-native-v1B176.7276.72—0.045183.576.720.921 ↗
Speech synthesis · Text → Audio · kokoro-english-v1 · Test duration 6 s · Batch 1 · audio-sec/sec · 1 rows
ModelFormatEngineWorkloadConcurrency / batchMean throughputBest measuredTTFT (ms)P95 (s)Observed occupancy (GiB)GPU power (W)Lowest measuredThroughput / WReports combined
hexgrad/Kokoro-82MScaling curves ↗fp32kokorottsaudio-native-v1B140.3640.36—0.1040.9960.540.360.671 ↗
Speech synthesis · Text → Audio · kokoro-english-v1 · Test duration 6 s · Batch 1 · audio-sec/sec · 1 rows
ModelFormatEngineWorkloadConcurrency / batchMean throughputBest measuredTTFT (ms)P95 (s)Observed occupancy (GiB)GPU power (W)Lowest measuredThroughput / WReports combined
hexgrad/Kokoro-82MScaling curves ↗fp32kokorottsaudio-native-v1B143.4343.43—0.1010.9964.143.430.681 ↗
Speech synthesis · Text → Audio · kokoro-english-v1 · Test duration 6 s · Batch 2 · audio-sec/sec · 1 rows
ModelFormatEngineWorkloadConcurrency / batchMean throughputBest measuredTTFT (ms)P95 (s)Observed occupancy (GiB)GPU power (W)Lowest measuredThroughput / WReports combined
hexgrad/Kokoro-82MScaling curves ↗fp32kokorottsaudio-native-v1B279.8479.84—0.11.03100.879.840.791 ↗
Speech synthesis · Text → Audio · kokoro-english-v1 · Test duration 6 s · Batch 4 · audio-sec/sec · 1 rows
ModelFormatEngineWorkloadConcurrency / batchMean throughputBest measuredTTFT (ms)P95 (s)Observed occupancy (GiB)GPU power (W)Lowest measuredThroughput / WReports combined
hexgrad/Kokoro-82MScaling curves ↗fp32kokorottsaudio-native-v1B441.841.8—0.4261.0365.941.80.631 ↗
Speech synthesis · Text → Audio · kokoro-english-v1 · Test duration 6 s · Batch 4 · audio-sec/sec · 1 rows
ModelFormatEngineWorkloadConcurrency / batchMean throughputBest measuredTTFT (ms)P95 (s)Observed occupancy (GiB)GPU power (W)Lowest measuredThroughput / WReports combined
hexgrad/Kokoro-82MScaling curves ↗fp32kokorottsaudio-native-v1B444.8844.88—0.3961.0369.144.880.651 ↗
Scaling curves and individual runs

One model and test protocol per curve; different workloads are never joined into one line.

hexgrad/Kokoro-82M · fp32 · kokoro · tts · audio-native-v1

Mean throughput audio-sec/sec
Batch size 1: 40.4 audio-sec/sec; Batch size 4: 41.8 audio-sec/sec011.52334.54614Batch size · log₂

Measured points only. Select a point to open its source report.

Mean run P95 latency s
Batch size 1: 0.104 s; Batch size 4: 0.426 s00.1170.2340.3520.46914Batch size · log₂

Measured points only. Select a point to open its source report.

Mean GPU memory used GB
Batch size 1: 0.994 GB; Batch size 4: 1.03 GB00.2830.5670.851.1314Batch size · log₂

Measured points only. Select a point to open its source report.

Each point is the mean of recorded runs at that level. A source report may differ from the mean; all runs are listed below.

Runs behind the curve · hexgrad/Kokoro-82M · audio-sec/sec · 2 rows
Batch sizeThroughputDecode (tok/s)Uncached prefill (tok/s)GPU power (W)TTFT (ms)P95 (s)Observed occupancy (GiB)Test dateView report
140.36 audio-sec/sec—Uncached prefill unavailableView report60.45—0.1040.9910/7/2026, 12:54:04 PMc0488108 ↗
441.8 audio-sec/sec—Uncached prefill unavailableView report65.94—0.4261.0310/7/2026, 12:54:04 PMc0488108 ↗

Original reports

Every report for this task, including quality, diagnostic and legacy runs that do not enter comparisons.
Original reports · 4 rows
ModelTest taskFormatEngineWorkloadProtocolQuality scoreTest date
Kokoro-82M ↗Speech synthesisfp32kokorottsaudio-native-v1—10/8/2026, 3:30:29 AM
Kokoro-82M ↗Speech synthesisfp32kokorottsaudio-native-v1—10/7/2026, 12:54:04 PM
Kokoro-82M ↗Speech synthesisfp32kokorottsaudio-native-v1—10/7/2026, 10:23:16 AM
Kokoro-82M ↗Speech synthesisfp32vLLMttsaudio-native-v1—10/7/2026, 9:40:16 AM

Running benchmarks

0Updates automatically

No benchmarks are currently reported as running.

Methodology

Average of error-free runs for the selected recorded configuration. Missing methods and fallback runs are not ranked. Performance index = average / fastest matching hardware × 100. Value index = performance per purchase cost / best comparable value × 100. Quality is separate by dataset and protocol. Prices are dated seller listings, not completed transactions or market averages.

Updated 11/5/2026, 11:07:29 PM