NVIDIA GH200 Grace Hopper, AI & Machine Learning Benchmarks & Specs

141GB · AI Score 65.6/100 · anchored estimate vs 51 measured cards

65.6 AI Score Includes estimates

We have not run NVIDIA GH200 Grace Hopper on our bench. These figures are anchored estimates, interpolated per workload against the 51 GPUs we did measure (confidence: high (sibling silicon)). On Llama 3.1 8B (Q4_K_M) NVIDIA GH200 Grace Hopper should deliver about 273.9 tokens/sec. Stepping up to Qwen3 32B it should hold roughly 78.2 tok/s. The full Llama 3.3 70B still runs, at about 43.5 tok/s. For image generation, SDXL should run near 18.58 it/s, and FLUX.1-dev at 4.44 it/s. All 12 workloads fit in 141GB. There is no model in our suite this card has to turn down.

AI & Machine Learning benchmark results

Text Generation tok/s 5

Qwen3 4B325.5

Llama 3.1 8B273.9

Qwen2.5-Coder 14B151.5

Qwen3 32B78.2

Llama 3.3 70B43.5

Workload	Result	Telemetry	Data
Qwen3 4B	325.5 tok/s	estimated	Est.
Llama 3.1 8B	273.9 tok/s	estimated	Est.
Qwen2.5-Coder 14B	151.5 tok/s	estimated	Est.
Qwen3 32B	78.2 tok/s	estimated	Est.
Llama 3.3 70B	43.5 tok/s	estimated	Est.

Image Generation images/min 3

Stable Diffusion XL37.16

Z-Image Turbo23.175

FLUX.1 dev9.514

Workload	Result	Telemetry	Data
Stable Diffusion XL	37.16 images/min	estimated	Est.
FLUX.1 dev	9.51 images/min	estimated	Est.
Z-Image Turbo	23.18 images/min	estimated	Est.

Image Editing images/min 2

Workload	Result	Telemetry	Data
FLUX.1 Kontext dev	4.39 images/min	estimated	Est.
Qwen-Image-Edit	3.76 images/min	estimated	Est.

Video Generation frames/s 2

Workload	Result	Telemetry	Data
LTX-Video (distilled)	17.87 frames/s	estimated	Est.
Wan 2.2 5B (720p)	1.41 frames/s	estimated	Est.

How this estimate is derived. This card hasn’t been through our bench yet, so its numbers are anchored estimates, interpolated from the 51 first-party measured cards (Sibling to measured H200: same GH100-class GPU (141GB HBM3e, 4.9 vs 4.8 TB/s) → ×~1.0). The VRAM “won’t fit” gates are exact, since they’re pure capacity limits. Confidence: high (sibling silicon). Estimates are replaced with measured data as more silicon goes through the bench. Full methodology →

NVIDIA GH200 Grace Hopper specifications

Architecture	Hopper
CUDA cores	16,896
VRAM	141GB HBM3e
Memory bus	6144-bit
Memory bandwidth	4900 GB/s
Boost clock	1,980 MHz
TDP	700 W
Process	TSMC 4N
Interface	SXM (Grace superchip)
Release date	2024-02-01
Launch MSRP	$40,000

Verdict, nothing in our suite slows it down

NVIDIA GH200 Grace Hopper scores 65.6/100, #4 of 102. It ran all 12 workloads. Figures are anchored estimates, not measurements, we flag that on every row.

Relative performance: where the NVIDIA GH200 Grace Hopper lands

100% = this card, AI & Machine Learning headline metric (AI Score). #4 of 21 datacenter cards in this vertical.

GPU	%	AI Score
NVIDIA B300	143%	93.8
NVIDIA B200	119%	78
NVIDIA B100	102%	67
NVIDIA GH200 Grace Hopper	100%	65.6
NVIDIA H200	99%	65
NVIDIA H100 80GB HBM3	95%	62.6
NVIDIA H800 80GB	95%	62.6
NVIDIA H100 NVL	89%	58.1

One tier up: NVIDIA B100 (+2%)
One tier down: NVIDIA H200 (-1%)
Best value alternative ($31,000): NVIDIA H200 (-1%)

← All AI & Machine Learning GPU rankings

The silicon

Transistors	80,000 million
Die size	814 mm²
Process node	4 nm
Fabricated by	TSMC
Transistor density	98.3 million per mm²

Denser than 90% of the 746 cards we have silicon data for. Density is the clearest measure of what a process node bought: a card that gained it without growing the die got its speed from the fab rather than the architecture.

Silicon figures from Wikipedia (CC BY-SA 4.0). Benchmarks on this page are our own. Compare every chip.

What this card can build

Whole-job timings, composed from our measured per-model results on this card.

Workflow	Time	Energy	Basis
24-frame storyboard	80 s	n/a	estimate, 0 of 2 stages measured
60-second AI short film	2 min	n/a	estimate, 0 of 3 stages measured
6-panel comic page	2.1 min	n/a	estimate, 0 of 3 stages measured
Character sheet, 12 poses	2.8 min	n/a	estimate, 0 of 2 stages measured
Short social clips	6.3 min	n/a	estimate, 0 of 3 stages measured
Full codebase review	6.6 min	n/a	estimate, 0 of 1 stage measured
Product photo shoot	10.2 min	n/a	estimate, 0 of 2 stages measured
Long-form article batch	10.7 min	n/a	estimate, 0 of 1 stage measured
Photo restoration batch	22.8 min	n/a	estimate, 0 of 1 stage measured