Skip to content

GPT Neox 20B on H100 SXM: VRAM, speed and fit

GPT Neox 20B at 8-bit runs on the H100 SXM with 2K of context at about 95 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on H100 SXM · 8-bit (Q8_0) · Ollama / LM Studio · Chat

GPU memory
23.3 GB
of 80.0 GB
Context
2K
auto
Speed
95 tok/s
Very fast
First reply
0.3 s
2K prompt

Where the memory goes

Model
20.5 GB
Context
2.1 GB
Overhead
751 MB
OS
0.0 GB
Free
56.7 GB
Memory
23.3 GB of 80.0 GB
Context
2K tokens · auto
Speed
95 tok/s · very fast
First reply
0.3 s for a 2K prompt
Running cost
$11/mo at home · $33/mo on a cloud RTX A5000

Validation: Validated

What GPT Neox 20B on H100 SXM was checked against
CheckTargetResultStatus
Fits in GPU memory80.0 GB usable23.3 GB neededPass
Headroom5% free56.7 GB free (71%)Pass
Context for chat2K or more2KPass
Writing speed15 tok/s or more95 tok/sPass
First replyunder 3 s, 2K prompt0.3 sPass

Targets: a reply you can read along with, a first reply the use case tolerates, and 5% of memory left free. Calculated from published specs, not benchmarked.

The setup

Model
GPT Neox 20B
Parameters
20.7B
Precision
8-bit (Q8_0)
Context
2K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
H100 SXM
GPU can use
80.0 GB of 80.0 GB

Running cost

At home
$11/mo
500 W at the wall
Renting instead
$33/mo
RTX A5000 at $0.27/h

4 h a day, power at $0.18 per kWh, cloud at RunPod on-demand prices as of Nov 22, 2026.

How to run it

llama.cpp
llama-server -hf mradermacher/gpt-neox-20b-i1-GGUF
vLLM
vllm serve EleutherAI/gpt-neox-20b

GPT Neox 20B on similar machines

  • 2× RTX PRO 6000 BlackwellRuns
  • 4× RTX PRO 6000 BlackwellRuns
  • 8× RTX PRO 6000 BlackwellRuns
  • A100 80 GB SXMRuns
  • 4× H200 NVLRuns
  • 8× H200 NVLRuns

All machines: GPT Neox 20B VRAM requirements. Everything that runs here: what can the H100 SXM run?

Other models that run on H100 SXM

Questions

Can GPT Neox 20B run on H100 SXM?

GPT Neox 20B at 8-bit runs on the H100 SXM with 2K of context at about 95 tokens a second.

Report VR-1RIPBI2, calculated Nov 22, 2026 from published specs, not benchmarked.