Skip to content
m.recipes
Find a configuration
Sign in
24 GB · 9 reports

RTX 3090 · 24 GB

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

All tasks · Qualified results

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

No qualified results yet · All tasks · RTX 3090 · 24 GB

9 original reports are kept as evidence but do not qualify for comparison; each report states why. This does not show that the model or GPU cannot run it.

Check model launch supportView original reportsMeasured alternatives for this taskTest your own GPU

Measured alternatives on other GPUs

Measured on different GPUs with the same or smaller listed capacity. These are references for choosing hardware, not performance or compatibility evidence for your selected GPU.

  1. Qwen3-4B-AWQ

    Qwen · 3 tested configurations · 2 GPUs

    • L4 · 24 GBAWQ · 8K → 2K tokens · No cache reuse
      Output throughput · includes prefill · 1 request
      55.3tok/s
      Best measured · 4 concurrent requests
      109tok/s
      Observed GPU occupancy
      19.5GiB
      First token
      1.6 s
      Use this recipe · 4 concurrent requests ↗Evidence report and run availability ↗Same test on other GPUs
      Test details · 2 concurrency levels

      Total output tokens divided by the full request window, including prefill and queueing.

      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
      vllm · vLLM 0.29.0
      Protocol
      wbench-fixed-http-v3
      Dataset
      migration-analysis-v1:sha256=68590c248e43f32c04dfd8062464fd1e4856917795fe307225e5c6112a54fe94:cache=isolated:repeats=3
      Workload
      8192in-2048out
      ConcurrencyMean throughput (tok/s)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
      C155.337.17 s1.6 s19.5471.919301a9ba ↗
      C4109.4378.88 s3.1 s19.5571.919301a9ba ↗
    • 1 other test configurations on this GPU
      • AWQ · 8K → 2K tokens · Prefix cache reuse
        Output throughput · includes prefill · 1 request
        57.4tok/s
        Best measured · 4 concurrent requests
        175tok/s
        Observed GPU occupancy
        19.5GiB
        First token
        60 ms
        Use this recipe · 4 concurrent requests ↗Evidence report and run availability ↗Same test on other GPUs
        Test details · 2 concurrency levels

        Total output tokens divided by the full request window, including prefill and queueing.

        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
        vllm · vLLM 0.29.0
        Protocol
        wbench-fixed-http-v3
        Dataset
        migration-analysis-v1:sha256=68590c248e43f32c04dfd8062464fd1e4856917795fe307225e5c6112a54fe94:cache=prefix-reuse:repeats=3
        Workload
        8192in-2048out
        ConcurrencyMean throughput (tok/s)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
        C157.3935.7 s60 ms19.5571.817e2cbfda ↗
        C4175.3946.75 s114 ms19.5571.917e2cbfda ↗
    Show 1 more GPUs
    • T4 · 16 GBAWQ · 8K → 2K tokens · No cache reuse
      Output throughput · includes prefill · 1 request
      15.4tok/s
      Best measured · 4 concurrent requests
      20.3tok/s
      Observed GPU occupancy
      12.7GiB
      First token
      28.5 s
      Use this recipe · 4 concurrent requests ↗Evidence report and run availability ↗Same test on other GPUs
      Test details · 2 concurrency levels

      Total output tokens divided by the full request window, including prefill and queueing.

      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
      vllm · vLLM 0.29.0
      Protocol
      wbench-fixed-http-v3
      Dataset
      migration-analysis-v1:sha256=68590c248e43f32c04dfd8062464fd1e4856917795fe307225e5c6112a54fe94:cache=isolated:repeats=3
      Workload
      8192in-2048out
      ConcurrencyMean throughput (tok/s)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
      C115.39133.36 s28.5 s12.7467.91fcfe2d64 ↗
      C420.31468.13 s53.1 s12.7467.51fcfe2d64 ↗

  2. embeddinggemma-2

    google · 4 tested configurations · 2 GPUs

    • L4 · 24 GBnf4 · Text → Vectors · embedding-mteb-tasks-v1 · Test duration 90 s
      Throughput · Batch 1
      175vectors/sec
      Best measured · Batch 16
      1,639vectors/sec
      Observed GPU occupancy
      0.9GiB
      P95 latency
      6 ms
      Use this recipe · Batch 16 ↗Evidence report and run availability ↗Same test on other GPUs
      Test details · 3 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
      sentence-transformers · PyTorch 2.12.1+cu126
      Protocol
      native-forward-v1
      Dataset
      embedding-mteb-tasks-v1
      Workload
      embedding
      Batch sizeMean throughput (vectors/sec)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
      B1174.886 ms—0.8553.167ad6a821 ↗
      B4571.037 ms—0.855767ad6a821 ↗
      B161,639.0810 ms—0.8559.867ad6a821 ↗
    • 1 other test configurations on this GPU
      • fp16 · Text → Vectors · embedding-mteb-tasks-v1 · Test duration 90 s
        Throughput · Batch 1
        177vectors/sec
        Best measured · Batch 16
        2,039vectors/sec
        Observed GPU occupancy
        1GiB
        P95 latency
        6 ms
        Use this recipe · Batch 16 ↗Evidence report and run availability ↗Same test on other GPUs
        Test details · 3 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
        sentence-transformers · PyTorch 2.12.1+cu126
        Protocol
        native-forward-v1
        Dataset
        embedding-mteb-tasks-v1
        Workload
        embedding
        Batch sizeMean throughput (vectors/sec)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
        B1177.466 ms—0.9949.86e0700e04 ↗
        B4650.166 ms—0.9956.96e0700e04 ↗
        B162,039.178 ms—0.9958.46e0700e04 ↗
    Show 1 more GPUs
    • T4 · 16 GBnf4 · Text → Vectors · embedding-mteb-tasks-v1 · Test duration 90 s
      Throughput · Batch 1
      90vectors/sec
      Best measured · Batch 16
      815vectors/sec
      Observed GPU occupancy
      0.7GiB
      P95 latency
      11 ms
      Use this recipe · Batch 16 ↗Evidence report and run availability ↗Same test on other GPUs
      Test details · 3 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
      sentence-transformers · PyTorch 2.12.1+cu126
      Protocol
      native-forward-v1
      Dataset
      embedding-mteb-tasks-v1
      Workload
      embedding
      Batch sizeMean throughput (vectors/sec)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
      B190.0211 ms—0.7554.6243c67a74 ↗
      B4303.3513 ms—0.7560.1243c67a74 ↗
      B16814.6420 ms—0.7564.1243c67a74 ↗
    • 1 other test configurations on this GPU
      • fp16 · Text → Vectors · embedding-mteb-tasks-v1 · Test duration 90 s
        Throughput · Batch 1
        109vectors/sec
        Best measured · Batch 16
        1,004vectors/sec
        Observed GPU occupancy
        0.9GiB
        P95 latency
        9 ms
        Use this recipe · Batch 16 ↗Evidence report and run availability ↗Same test on other GPUs
        Test details · 3 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
        sentence-transformers · PyTorch 2.12.1+cu126
        Protocol
        native-forward-v1
        Dataset
        embedding-mteb-tasks-v1
        Workload
        embedding
        Batch sizeMean throughput (vectors/sec)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
        B1108.659 ms—0.8949.322846de88 ↗
        B4365.6511 ms—0.8957.922846de88 ↗
        B161,004.1116 ms—0.8963.922846de88 ↗

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

Measurements without a comparable protocol · 53 rows

These measured rates remain visible, but are excluded from comparison curves because the protocol is unknown.

Measurements without a comparable protocol · 53 rows
ModelFormatEngineWorkloadConcurrencyThroughputGPU power (W)
DeepSeek-R1-Qwen3-8B ↗BF16vLLM2K143.85 tok/s297.22
DeepSeek-R1-Qwen3-8B ↗BF16vLLM2K8292.08 tok/s307.05
DeepSeek-R1-Qwen3-8B ↗BF16vLLM2K16507.73 tok/s311.83
DeepSeek-R1-Qwen3-8B ↗BF16vLLM2K32673.23 tok/s317.44
DeepSeek-R1-Qwen3-8B ↗BF16vLLM2K64941.99 tok/s323.55
Qwen2.5-3B-AWQ ↗AWQvLLM2K148.84 tok/s131.06
Qwen2.5-3B-AWQ ↗AWQvLLM2K8317.37 tok/s139.33
Qwen2.5-3B-AWQ ↗AWQvLLM2K16542.31 tok/s157.19
Qwen2.5-3B-AWQ ↗AWQvLLM2K321,025.24 tok/s232.82
Qwen2.5-3B-AWQ ↗AWQvLLM2K641,437.7 tok/s257.38
Qwen2.5-3B-AWQ ↗AWQvLLM2K1282,430.05 tok/s313.21
Qwen2.5-7B-AWQ ↗AWQvLLM2K160.02 tok/s238.76
Qwen2.5-7B-AWQ ↗AWQvLLM2K8446.87 tok/s256.88
Qwen2.5-7B-AWQ ↗AWQvLLM2K16819.45 tok/s297.6
Qwen2.5-7B-AWQ ↗AWQvLLM2K321,290.84 tok/s313.56
Qwen2.5-7B-AWQ ↗AWQvLLM2K642,232.82 tok/s333.05
Qwen2.5-7B-AWQ ↗AWQvLLM2K1282,406.34 tok/s332.53
Qwen2.5-14B-AWQ ↗AWQvLLM2K135.72 tok/s258.34
Qwen2.5-14B-AWQ ↗AWQvLLM2K8220.76 tok/s285.93
Qwen2.5-14B-AWQ ↗AWQvLLM2K16339.55 tok/s294.11
Qwen2.5-14B-AWQ ↗AWQvLLM2K32661.02 tok/s331.82
Qwen2.5-14B-AWQ ↗AWQvLLM2K64982.56 tok/s338.63
Qwen2.5-14B-AWQ ↗AWQvLLM2K128958.2 tok/s339.67
Qwen3-0.6B ↗BF16vLLM2K158.92 tok/s124.05
Qwen3-0.6B ↗BF16vLLM2K8429.4 tok/s139.97
Qwen3-0.6B ↗BF16vLLM2K16720.7 tok/s143.06
Qwen3-0.6B ↗BF16vLLM2K321,353.52 tok/s186.69
Qwen3-0.6B ↗BF16vLLM2K642,470.97 tok/s251.17
Qwen3-0.6B ↗BF16vLLM2K1283,614.78 tok/s301.66
Qwen3-1.7B ↗BF16vLLM2K157.35 tok/s160.17
Qwen3-1.7B ↗BF16vLLM2K8431.03 tok/s221.41
Qwen3-1.7B ↗BF16vLLM2K16723.51 tok/s219.04
Qwen3-1.7B ↗BF16vLLM2K321,381.34 tok/s249.81
Qwen3-1.7B ↗BF16vLLM2K642,478.34 tok/s289.71
Qwen3-1.7B ↗BF16vLLM2K1283,674.22 tok/s314.55
Qwen3-4B ↗BF16vLLM2K144.05 tok/s269.55
Qwen3-4B ↗BF16vLLM2K8340.14 tok/s282.51
Qwen3-4B ↗BF16vLLM2K16568.12 tok/s276.74
Qwen3-4B ↗BF16vLLM2K321,067.35 tok/s299.14
Qwen3-4B ↗BF16vLLM2K641,977.47 tok/s313.49
Qwen3-4B ↗BF16vLLM2K1281,654.61 tok/s325.49
Qwen3-8B-AWQ ↗AWQvLLM2K142.6 tok/s169.39
Qwen3-8B-AWQ ↗AWQvLLM2K8317.55 tok/s240.12
Qwen3-8B-AWQ ↗AWQvLLM2K16517.51 tok/s248.81
Qwen3-8B-AWQ ↗AWQvLLM2K32693.83 tok/s298.86
Qwen3-8B-AWQ ↗AWQvLLM2K641,307.89 tok/s329.99
Qwen3-8B-AWQ ↗AWQvLLM2K1281,494.79 tok/s333.79
Qwen3-14B-AWQ ↗AWQvLLM2K138.55 tok/s276.73
Qwen3-14B-AWQ ↗AWQvLLM2K8241.69 tok/s292.69
Qwen3-14B-AWQ ↗AWQvLLM2K16393.49 tok/s299.59
Qwen3-14B-AWQ ↗AWQvLLM2K32712.21 tok/s332.19
Qwen3-14B-AWQ ↗AWQvLLM2K641,031.21 tok/s339.16
Qwen3-14B-AWQ ↗AWQvLLM2K1281,237.67 tok/s337.88

Original reports

Every report for this task, including quality, diagnostic and legacy runs that do not enter comparisons.
Original reports · 9 rows
ModelTest taskFormatEngineWorkloadProtocolQuality scoreTest date
Qwen3-14B-AWQ ↗Text generationAWQvLLM2KMethod unknown—1/12/2026, 12:22:46 PM
Qwen3-8B-AWQ ↗Text generationAWQvLLM2KMethod unknown—1/12/2026, 12:15:24 PM
Qwen3-4B ↗Text generationBF16vLLM2KMethod unknown—1/12/2026, 12:08:00 PM
Qwen3-1.7B ↗Text generationBF16vLLM2KMethod unknown—1/12/2026, 12:00:47 PM
Qwen3-0.6B ↗Text generationBF16vLLM2KMethod unknown—1/12/2026, 11:54:49 AM
Qwen2.5-14B-AWQ ↗Text generationAWQvLLM2KMethod unknown—1/11/2026, 6:39:16 PM
DeepSeek-R1-Qwen3-8B ↗Text generationBF16vLLM2KMethod unknown—1/11/2026, 6:31:36 PM
Qwen2.5-7B-AWQ ↗Text generationAWQvLLM2KMethod unknown—1/11/2026, 6:23:40 PM
Qwen2.5-3B-AWQ ↗Text generationAWQvLLM2KMethod unknown—1/11/2026, 6:18:10 PM

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