---
license: mit
tags:
  - blackwell
  - sm120
  - rtx-pro-6000
  - vllm
  - sglang
  - llama-cpp
  - fp8
  - nvfp4
  - inference
  - benchmark
pretty_name: Blackwell SM120 Serving Matrix
size_categories:
  - n<1K
---

# 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-NVFP4` on vLLM 0.25.1 with `--moe-backend=flashinfer_b12x` reserves 11.47 GiB more than the `marlin` baseline before the KV cache is sized, which is why it OOMs 16-32 GB Blackwell cards during `profile_run` but starts on 96 GB. [Full repro.](https://conatus.jahn.ai/ai-engineering/sm120-b12x-workspace/)
- `Qwen/Qwen3-8B` FP8 on vLLM gives roughly 1.5x aggregate throughput over BF16 with no regression on a fixed 20-prompt check. [Report.](https://conatus.jahn.ai/ai-engineering/sample-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](https://conatus.jahn.ai/ai-engineering/#sku-e).

Operator: Conatus AI. Measurements are on our own hardware; corrections and additions welcome.