---
title: RTX PRO 4000 Hub
emoji: 📊
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: "6.19.0"
app_file: app.py
pinned: false
---

# RTX PRO 4000 Blackwell Community Hub

Benchmarks and configs, not a support channel.

This Hugging Face organization is for owner-measured inference results and
workstation configuration notes around the NVIDIA RTX PRO 4000 Blackwell.

Anchored on the RTX PRO 4000 Blackwell (desktop, 24 GB, 672 GB/s,
single-slot). Open to 24 GB-class Blackwell workstation GPUs.

This is a reputation and skill-building project. There is no affiliate,
sponsor, or monetization framing.

## Anchor Card

The desktop RTX PRO 4000 Blackwell is the reference SKU for this hub.

| Field | Desktop RTX PRO 4000 Blackwell |
|---|---|
| Architecture | Blackwell, GB203, 4nm |
| CUDA cores | 8,960 |
| Tensor cores | 280, fifth generation |
| RT cores | 70, fourth generation |
| Memory | 24 GB GDDR7 ECC |
| Bus | 192-bit |
| Bandwidth | 672 GB/s |
| Power | 140 W total board power |
| Form factor | Full-height, single-slot |
| PCIe | Gen5 x16 |

Source note: desktop power uses NVIDIA's 140 W reference value. Some board or
retailer pages list 145 W; treat that as board/listing variance, not the
canonical hub value. This card uses measured benchmark data, not marketing TOPS
figures.

## What Belongs Here

- Inference benchmarks: tok/s across models, quant levels, context lengths, and
  backends such as llama.cpp, vLLM, and TensorRT-LLM.
- Quant compatibility: what GGUF, NVFP4, FP8, BF16, and other formats fit in a
  24 GB-class Blackwell workstation card.
- Workstation configs: systemd units, Docker launch patterns, thermal notes, and
  multi-GPU setup notes with secrets and local network details stripped.
- Owner-operated results from real hardware.

## Variant Boundary

Rows must declare `variant`: `desktop`, `sff`, or `laptop`.

The same "RTX PRO 4000 Blackwell" name covers multiple physical products. Their
memory systems are not interchangeable, so results are never merged across
variants.

VRAM, memory type, and bus width are not desktop/SFF identifiers. Desktop and
SFF rows are both 24 GB GDDR7 ECC on a 192-bit bus, so benchmark rows also carry
bandwidth, TDP, and PCIe lanes.

| Variant | VRAM | Memory | Bus | Bandwidth | Power | PCIe | Notes |
|---|---:|---|---:|---:|---:|---:|---|
| desktop | 24 GB | GDDR7 ECC | 192-bit | 672 GB/s | 140 W | x16 | Anchor SKU, full-height single-slot |
| sff | 24 GB | GDDR7 ECC | 192-bit | 432 GB/s | 70 W | x8 | Low-profile dual-slot half-height |
| laptop | 16 GB | GDDR7 ECC | 256-bit | 896 GB/s | OEM | Gen5 | Mobile GPU, OEM TGP varies |

SFF source: NVIDIA's RTX PRO 4000 Blackwell SFF Edition datasheet lists 24 GB
GDDR7 ECC, 192-bit, 432 GB/s, PCIe 5.0 x8, 70 W, and dual-slot half-height form
factor.

## Repos

- `benchmarks`: JSONL dataset, schema, validation, methodology, and submission
  template.
- `configs`: sanitized templates for running and measuring local inference on
  workstation cards.

Tag related model cards with `rtx-pro-4000` so people can find work targeting
this card family.