Best AI Models for NVIDIA GH200 Grace Hopper Superchip (144.0GB)
Grace Hopper superchip: 144 GB HBM3e GPU pool plus a large NVLink-coherent LPDDR5X extension. Server/cloud part, listed for comparison.
With 144 GB of memory, this is a high-end configuration for local AI. You can comfortably run most open-source LLMs including large 70B parameter models at good quantization levels, making it one of the best setups for serious local AI work.
At this memory tier, nearly every popular open-source model is within reach. You can run Llama 3 70B at Q4_K_M or even Q5_K_M quantization with room to spare, handle coding assistants like DeepSeek Coder 33B at high quality, and easily run any 7B–30B model at full or near-full precision. Context windows remain generous even with larger models, so multi-turn conversations and long-document processing work smoothly.
Runs Well
- 70B models (Llama 3 70B, Qwen 72B) at Q4–Q5
- 30B models at Q6–Q8 quality
- 7B–14B models at full FP16 precision
- Vision models (LLaVA, CogVLM) without compromise
Challenging
- Mixture-of-experts models like Mixtral 8x22B at higher quants
- 120B+ models still require lower quantizations
What LLMs Can NVIDIA GH200 Grace Hopper Superchip Run?
302 models · 271 excellent · 12 good
Showing compatibility for NVIDIA GH200 Grace Hopper Superchip
| Model | Quant | VRAM | Speed | Context | Status | Grade |
|---|---|---|---|---|---|---|
Q4_K_M·619.8 t/s tok/s·131K ctx·RUNS GREAT | Q4_K_M | 5.3 GB | 619.8 t/s | 131K | RUNS GREAT | S100 |
Q4_K_M·154.2 t/s tok/s·262K ctx·RUNS GREAT | Q4_K_M | 20.7 GB | 154.2 t/s | 262K | RUNS GREAT | S97 |
BF16·189.1 t/s tok/s·262K ctx·RUNS GREAT | BF16 | 72.6 GB | 189.1 t/s | 262K | RUNS GREAT | S99 |
Q4_K_M·306.6 t/s tok/s·RUNS GREAT | Q4_K_M | 23.3 GB | 306.6 t/s | — | RUNS GREAT | S100 |
Q4_K_M·977.0 t/s tok/s·262K ctx·RUNS GREAT | Q4_K_M | 3.3 GB | 977.0 t/s | 262K | RUNS GREAT | S100 |
Q4_K_M·872.6 t/s tok/s·262K ctx·RUNS GREAT | Q4_K_M | 3.6 GB | 872.6 t/s | 262K | RUNS GREAT | S100 |
Q4_K_M·1085.9 t/s tok/s·33K ctx·RUNS GREAT | Q4_K_M | 2.4 GB | 1085.9 t/s | 33K | RUNS GREAT | S100 |
Q4_K_M·511.2 t/s tok/s·41K ctx·RUNS GREAT | Q4_K_M | 6.2 GB | 511.2 t/s | 41K | RUNS GREAT | S100 |
Q4_K_M·928.6 t/s tok/s·131K ctx·RUNS GREAT | Q4_K_M | 3.4 GB | 928.6 t/s | 131K | RUNS GREAT | S100 |
Q4_K_M·512.9 t/s tok/s·262K ctx·RUNS GREAT | Q4_K_M | 6.2 GB | 512.9 t/s | 262K | RUNS GREAT | S100 |
Q4_K_M·154.2 t/s tok/s·262K ctx·RUNS GREAT | Q4_K_M | 20.6 GB | 154.2 t/s | 262K | RUNS GREAT | S97 |
Q4_K_M·152.1 t/s tok/s·200K ctx·RUNS WELL | Q4_K_M | 131.6 GB | 152.1 t/s | 200K | RUNS WELL | A81 |
Q4_K_M·188.7 t/s tok/s·262K ctx·RUNS GREAT | Q4_K_M | 16.9 GB | 188.7 t/s | 262K | RUNS GREAT | S99 |
Q2_K·166.0 t/s tok/s·1049K ctx·RUNS WELL | Q2_K | 129.6 GB | 166.0 t/s | 1049K | RUNS WELL | A84 |
Q4_K_M·1561.3 t/s tok/s·131K ctx·RUNS GREAT | Q4_K_M | 2.0 GB | 1561.3 t/s | 131K | RUNS GREAT | S100 |
Q4_K_M·549.1 t/s tok/s·37K ctx·RUNS GREAT | Q4_K_M | 5.8 GB | 549.1 t/s | 37K | RUNS GREAT | S100 |
NVIDIA GH200 Grace Hopper Superchip Specifications
- Brand
- NVIDIA
- Architecture
- Hopper (Grace Hopper)
- Compute Capability
- 9.0 (CUDA SM version)
- VRAM
- 144.0 GB HBM3e
- Memory Bandwidth
- 4900.0 GB/s
- CUDA Cores
- 16,896
- Tensor Cores
- 528
- FP16 Performance
- 989.50 TFLOPS
- TDP
- 1000W
- Release Date
- 2024-01-01
Get Started
Prompt Processing
Estimated for Llama 3 1 Nemotron 51B Instruct, the compute-bound phase that reads your prompt before the first reply token appears.
Short chat
117 ms
512 tok prompt
Long chat
938 ms
4,096 tok prompt
Document / codebase
7.5 s
32,768 tok prompt
Prefill is compute-bound and a different number from the decode tok/s shown elsewhere on this page — how prompt processing works →
Efficiency & Value
Based on Llama 3 1 Nemotron 51B Instruct at ~28.4 tok/s decode.
Tokens per watt
Higher is better.
Performance figures are estimates calibrated as of 2026-09-21 — see calibration basis →
GPUs to Consider Over NVIDIA GH200 Grace Hopper Superchip
Similar GPUs and upgrades with more VRAM or higher bandwidth for AI
AMD Instinct MI350X
AMD · CDNA 4
AMD Instinct MI355X
AMD · CDNA 4
NVIDIA B200
NVIDIA · Blackwell
NVIDIA B300
NVIDIA · Blackwell Ultra
AMD Instinct MI325X
AMD · CDNA 3
AMD Instinct MI300X
AMD · CDNA 3
Frequently Asked Questions
- Can NVIDIA GH200 Grace Hopper Superchip run Qwen3.8 27B?
Yes, the NVIDIA GH200 Grace Hopper Superchip with 144 GB can run Qwen3.8 27B, Muse Glimmer 30B, Gemma 4 26B A4B IT, and 2652 other models. 2485 models run at excellent quality, and 124 at good quality. Check the compatibility table above for the full list with VRAM usage and estimated speed.
- Is NVIDIA GH200 Grace Hopper Superchip good for AI?
The NVIDIA GH200 Grace Hopper Superchip has 144 GB of HBM3e, making it excellent for running local AI models. It supports 2609 models at good quality or better. With 4900.0 GB/s memory bandwidth, it delivers fast token generation speeds. This is an enthusiast-grade GPU that handles most popular open-source LLMs.
- How many parameters can NVIDIA GH200 Grace Hopper Superchip handle?
With 144 GB, the NVIDIA GH200 Grace Hopper Superchip supports models from 3B to 70B+ parameters depending on quantization level. At Q4_K_M (the recommended sweet spot), you can fit roughly 240B parameters. This means 7B models at high quality (Q6/Q8) or 30B+ models at Q4.
- What quantization should I use on NVIDIA GH200 Grace Hopper Superchip?
For the best balance of quality and speed on the NVIDIA GH200 Grace Hopper Superchip, start with Q4_K_M — it preserves ~85% of the original model quality while keeping VRAM usage reasonable. With 24+ GB, you have the headroom to run 7B models at Q5_K_M or even Q6_K for noticeably better output quality. For larger 30B models, Q4_K_M remains the sweet spot.
- How fast is NVIDIA GH200 Grace Hopper Superchip for AI inference?
With 4900.0 GB/s memory bandwidth, the NVIDIA GH200 Grace Hopper Superchip achieves approximately 708 tokens/sec on a 7B model at Q4_K_M — that's very fast, well above conversational speed. A 14B model runs at ~354 tok/s. Token generation speed scales inversely with model size — smaller models are significantly faster.
tok/s = (4900 GB/s ÷ model GB) × efficiency
Smaller models = faster inference. Memory bandwidth is the main bottleneck for token generation speed.
Estimated speed on NVIDIA GH200 Grace Hopper Superchip
~183 tok/s~174 tok/s~411 tok/s~156 tok/sReal-world results typically within ±20%. Speed depends on quantization kernel, batch size, and software stack.
- What's the best model for NVIDIA GH200 Grace Hopper Superchip?
The top-rated models for the NVIDIA GH200 Grace Hopper Superchip are Qwen3.8 27B, Muse Glimmer 30B, Gemma 4 26B A4B IT. The best choice depends on your use case: coding assistants benefit from code-tuned models, while general chat works well with instruction-tuned models like Llama or Qwen.
- What power supply and cooling does NVIDIA GH200 Grace Hopper Superchip need?
The NVIDIA GH200 Grace Hopper Superchip has a TDP of 1000 W. A good rule of thumb is to provide at least double the GPU's TDP to cover the rest of the system — that means a 1600 W PSU or larger. At this power level, a high-airflow case matters: aim for at least two front intake fans and one rear exhaust, with tidy cabling so hot air isn't trapped around the card. LLM inference sustains full GPU load continuously — longer and more consistently than most gaming workloads — so also make sure your CPU cooler can keep up under combined load.
- Anything to watch out for with NVIDIA GH200 Grace Hopper Superchip?
Grace Hopper superchip: 144 GB HBM3e GPU pool plus a large NVLink-coherent LPDDR5X extension. Server/cloud part, listed for comparison.