Run an AI model on your computer

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. Qwen3.8-27B-NVFP4

   RadixArk · 2 tested configurations · 1 GPU

   • RTX 5090 · 32 GBNVFP4 · 8K → 2K tokens

     Output throughput · includes prefill · 1 request

     70.8tok/s

     Observed GPU occupancy

     27.3GiB

     First token

     633 ms

     Use this recipe · 1 request ↗Evidence report and run availability ↗Same test on other GPUs
     Test details

     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

     Protocol

     wbench-fixed-http-v1

     Dataset

     synthetic-tokenizer-v1:repeats=2

     Workload

     8192in-2048out

     Concurrency	Mean throughput (tok/s)	P95	TTFT	Observed occupancy (GiB)	GPU power (W)	Reports combined	View report
     C1	70.82	29.01 s	633 ms	27.29	405.5	1	bad294c0 ↗

   • 1 other test configurations on this GPU

     • NVFP4 · 32K → 2K tokens

       Output throughput · includes prefill · 1 request

       62.5tok/s

       Observed GPU occupancy

       27.3GiB

       First token

       3.6 s

       Use this recipe · 1 request ↗Evidence report and run availability ↗Same test on other GPUs
       Test details

       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

       Protocol

       wbench-fixed-http-v1

       Dataset

       synthetic-tokenizer-v1:repeats=2

       Workload

       32768in-2048out

       Concurrency	Mean throughput (tok/s)	P95	TTFT	Observed occupancy (GiB)	GPU power (W)	Reports combined	View report
       C1	62.48	32.79 s	3.6 s	27.29	433.3	1	746bbc39 ↗

2. 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

     Concurrency	Mean throughput (tok/s)	P95	TTFT	Observed occupancy (GiB)	GPU power (W)	Reports combined	View report
     C1	55.3	37.17 s	1.6 s	19.54	71.9	1	9301a9ba ↗
     C4	109.43	78.88 s	3.1 s	19.55	71.9	1	9301a9ba ↗

   • 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

       Concurrency	Mean throughput (tok/s)	P95	TTFT	Observed occupancy (GiB)	GPU power (W)	Reports combined	View report
       C1	57.39	35.7 s	60 ms	19.55	71.8	1	7e2cbfda ↗
       C4	175.39	46.75 s	114 ms	19.55	71.9	1	7e2cbfda ↗

   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

     Concurrency	Mean throughput (tok/s)	P95	TTFT	Observed occupancy (GiB)	GPU power (W)	Reports combined	View report
     C1	15.39	133.36 s	28.5 s	12.74	67.9	1	fcfe2d64 ↗
     C4	20.31	468.13 s	53.1 s	12.74	67.5	1	fcfe2d64 ↗