model_id stringclasses 3
values | runtime stringclasses 3
values | runtime_version stringclasses 6
values | gpu stringclasses 1
value | gpu_arch stringclasses 1
value | vram_gb int64 96 96 | quantization stringclasses 4
values | tp int64 1 1 | backend stringclasses 3
values | context_len int64 180k 180k ⌀ | concurrency int64 1 32 ⌀ | startup_status stringclasses 1
value | correctness_status stringclasses 5
values | agg_tokens_per_second int64 428 2.6k ⌀ | ttft_ms_p50 int64 39 316 ⌀ | e2e_s_p99 float64 3.4 16.3 ⌀ | peak_vram_gb float64 89 91.8 ⌀ | artifact_url stringclasses 3
values | tested_at timestamp[s]date 2026-08-25 00:00:00 2026-08-26 00:00:00 | kv_cache_dtype stringclasses 1
value | available_kv_cache_gib float64 52.9 64.4 ⌀ | max_concurrency_x float64 29.7 36.2 ⌀ | note stringclasses 2
values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Qwen/Qwen3-8B | vllm | 0.27.1 | RTX PRO 6000 Blackwell | sm_120 | 96 | bf16 | 1 | null | null | 32 | ok | greedy-matched | 1,725 | 39 | 3.4 | null | https://conatus.jahn.ai/ai-engineering/sample-report/ | 2026-08-25T00:00:00 | null | null | null | null |
Qwen/Qwen3-8B | sglang | 0.5.9 | RTX PRO 6000 Blackwell | sm_120 | 96 | bf16 | 1 | null | null | 32 | ok | greedy-matched | 1,327 | 42 | 5 | null | https://conatus.jahn.ai/ai-engineering/sample-report/ | 2026-08-25T00:00:00 | null | null | null | null |
Qwen/Qwen3-8B | llama.cpp | cuda | RTX PRO 6000 Blackwell | sm_120 | 96 | bf16 | 1 | null | null | 32 | ok | greedy-matched | 428 | 316 | 16.3 | null | https://conatus.jahn.ai/ai-engineering/sample-report/ | 2026-08-25T00:00:00 | null | null | null | null |
Qwen/Qwen3-8B | vllm | 0.27.1 | RTX PRO 6000 Blackwell | sm_120 | 96 | fp8 | 1 | null | null | 32 | ok | no-regression-20prompt | 2,597 | null | null | null | https://conatus.jahn.ai/ai-engineering/sample-report/ | 2026-08-25T00:00:00 | null | null | null | null |
nvidia/Qwen3.6-35B-A3B-NVFP4 | vllm | 0.25.1 | RTX PRO 6000 Blackwell | sm_120 | 96 | nvfp4 | 1 | flashinfer_b12x | 180,224 | null | ok | not-audited | null | null | null | 91.8 | https://conatus.jahn.ai/ai-engineering/sm120-b12x-workspace/ | 2026-08-26T00:00:00 | fp8 | 52.89 | 29.73 | OOMs on 16-32GB Blackwell during profile_run; starts on 96GB |
nvidia/Qwen3.6-35B-A3B-NVFP4 | vllm | 0.25.1 | RTX PRO 6000 Blackwell | sm_120 | 96 | nvfp4 | 1 | marlin | 180,224 | null | ok | not-audited | null | null | null | 88.97 | https://conatus.jahn.ai/ai-engineering/sm120-b12x-workspace/ | 2026-08-26T00:00:00 | fp8 | 64.36 | 36.17 | marlin baseline; 11.47 GiB more KV headroom than flashinfer_b12x |
unsloth/Qwen3.8-27B-NVFP4 | sglang | c7e2c08d1 (pre-#35228; all releases <=0.5.18) | RTX PRO 6000 Blackwell | sm_120 | 96 | nvfp4+fp8-lm_head (compressed-tensors mixed) | 1 | flashinfer | null | 1 | ok | FAIL: degenerate repetition, empty content (FP8 lm_head weight_scale never applied, sglang#34895) | null | null | null | null | https://github.com/sgl-project/sglang/issues/34895#issuecomment-5420022002 | 2026-08-26T00:00:00 | null | null | null | null |
unsloth/Qwen3.8-27B-NVFP4 | sglang | main 07a9de25b (>=5375babb, PR #35228) | RTX PRO 6000 Blackwell | sm_120 | 96 | nvfp4+fp8-lm_head (compressed-tensors mixed) | 1 | flashinfer | null | 1 | ok | greedy spot-check correct (lm_head weight_scale loads, 0 warnings) | null | null | null | null | https://github.com/sgl-project/sglang/issues/34895#issuecomment-5420022002 | 2026-08-26T00:00:00 | null | null | null | null |
Blackwell SM120 Serving Matrix
Measured serving results for LLMs on workstation and server Blackwell (sm_120, RTX PRO 6000, 96 GB), across vLLM, SGLang and llama.cpp, with FP8 and NVFP4 passes. Each row is one exact cell: a model, a runtime and version, a quantization, a topology and a load point, with the startup result, throughput or latency where measured, and a link to the raw artifact.
The point of the matrix is the cells that are known-broken as much as the ones that are known-good. Serving a given model on this GPU class often fails on one specific runtime and backend combination while another boots fine, and the failure is rarely in the model. This dataset records both, on real hardware.
Schema
| field | meaning |
|---|---|
model_id |
Hugging Face model id |
runtime / runtime_version |
serving stack and version |
gpu / gpu_arch / vram_gb |
hardware |
quantization / kv_cache_dtype |
weight and KV formats |
tp / backend / context_len / concurrency |
serving configuration |
startup_status |
ok or the failure class |
correctness_status |
how correctness was checked, if at all |
agg_tokens_per_second / ttft_ms_p50 / e2e_s_p99 |
throughput and latency where measured |
available_kv_cache_gib / max_concurrency_x / peak_vram_gb |
memory profile |
artifact_url / tested_at |
evidence link and date |
Notable cells
nvidia/Qwen3.6-35B-A3B-NVFP4on vLLM 0.25.1 with--moe-backend=flashinfer_b12xreserves 11.47 GiB more than themarlinbaseline before the KV cache is sized, which is why it OOMs 16-32 GB Blackwell cards duringprofile_runbut starts on 96 GB. Full repro.Qwen/Qwen3-8BFP8 on vLLM gives roughly 1.5x aggregate throughput over BF16 with no regression on a fixed 20-prompt check. Report.
Missing a cell?
If the exact combination you need is not in the matrix, it can be measured on the reference hardware and added: one public model and revision, one runtime, one quantization, one topology and one load point, returned with the exact command and raw logs. Fixed-scope single-cell verification is described at conatus.jahn.ai/ai-engineering#sku-e.
Operator: Conatus AI. Measurements are on our own hardware; corrections and additions welcome.
- Downloads last month
- 28