NVIDIA H100 PCIe, AI & Machine Learning Benchmarks & Specs

80GB · AI Score 46.4/100 · anchored estimate vs 51 measured cards

49.8 AI Score Includes estimates

We have not run NVIDIA H100 PCIe 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 H100 PCIe should deliver about 156.3 tokens/sec. Stepping up to Qwen3 32B it should hold roughly 44.2 tok/s. The full Llama 3.3 70B still runs, at about 24.5 tok/s. For image generation, SDXL should run near 14.93 it/s, and FLUX.1-dev at 3.65 it/s. All 12 workloads fit in 80GB. There is no model in our suite this card has to turn down. NVIDIA H100 PCIe isn't a retail purchase for most people. It's rented by the hour. You can run this exact card on RunPod.

AI & Machine Learning benchmark results

Text Generation tok/s 10

gpt-oss-20b251.11

Qwen3-4B245.86

Qwen3 30B A3B235.56

Llama-3.1-8B194.29

Qwen3 8B185.87

Gemma 4 12B115.77

Qwen3 14B112.57

Qwen2.5-Coder-14B103.65

Qwen3-32B54.34

Llama-3.3-70B28.44

Workload	Result	Telemetry	Data
Qwen3-4B	245.86 tok/s	148 W41°CQ4_K_M	✓ Measured
Llama-3.1-8B	194.29 tok/s	171 W44°CQ4_K_M	✓ Measured
Qwen3 8B	185.87 tok/s	175 W44°CQ4_K_M	✓ Measured
Gemma 4 12B	115.77 tok/s	179 W45°CQ4_K_M	✓ Measured
Qwen2.5-Coder-14B	103.65 tok/s	194 W46°CQ4_K_M	✓ Measured
Qwen3 14B	112.57 tok/s	197 W47°CQ4_K_M	✓ Measured
gpt-oss-20b	251.11 tok/s	121 W42°CQ4_K_M	✓ Measured
Qwen3 30B A3B	235.56 tok/s	113 W42°CQ4_K_M	✓ Measured
Qwen3-32B	54.34 tok/s	222 W49°CQ4_K_M	✓ Measured
Llama-3.3-70B	28.44 tok/s	229 W53°CQ4_K_M	✓ Measured

Image Generation images/min 3

Stable Diffusion XL29.86

Z-Image Turbo19.125

FLUX.1 dev7.821

Workload	Result	Telemetry	Data
Stable Diffusion XL	29.86 images/min	estimated	Est.
FLUX.1 dev	7.82 images/min	estimated	Est.
Z-Image Turbo	19.13 images/min	estimated	Est.

Image Editing images/min 2

Workload	Result	Telemetry	Data
FLUX.1 Kontext dev	3.69 images/min	estimated	Est.
Qwen-Image-Edit	3.18 images/min	estimated	Est.

Video Generation frames/s 2

Workload	Result	Telemetry	Data
LTX-Video (distilled)	15.09 frames/s	estimated	Est.
Wan 2.2 5B (720p)	1.15 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 H100 SXM: LLM ×0.597 (HBM2e 2.0 vs 3.35 TB/s), diffusion ×0.864 (14592 vs 16896 cores)). 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 H100 PCIe specifications

Architecture	Hopper
CUDA cores	14,592
VRAM	80GB HBM2e
Memory bus	5120-bit
Memory bandwidth	2000 GB/s
Boost clock	1,755 MHz
TDP	350 W
Process	TSMC 4N
Interface	PCIe 5.0 x16
Release date	2022-03-22
Launch MSRP	$25,000

Verdict, NVIDIA H100 PCIe on real AI workloads

NVIDIA H100 PCIe scores 46.4/100, #11 of 102. It ran all 12 workloads. Figures are anchored estimates, not measurements, we flag that on every row.

Relative performance: where the NVIDIA H100 PCIe lands

100% = this card, AI & Machine Learning headline metric (AI Score). #10 of 21 datacenter cards in this vertical.

GPU	%	AI Score
NVIDIA H100 80GB HBM3	126%	62.6
NVIDIA H800 80GB	126%	62.6
NVIDIA H100 NVL	117%	58.1
NVIDIA RTX PRO 6000 Blackwell Server Edition	100%	49.9
NVIDIA H100 PCIe	100%	49.8
NVIDIA A100 80GB SXM4	66%	33.1
NVIDIA A800 80GB	66%	33.1
NVIDIA A100 80GB PCIe	64%	31.7
NVIDIA L40S	56%	27.7

One tier up: NVIDIA RTX PRO 6000 Blackwell Server Edition (0%) · compare head-to-head
One tier down: NVIDIA A100 80GB SXM4 (-34%) · compare head-to-head
Best value alternative ($8,565): NVIDIA RTX PRO 6000 Blackwell Server Edition (0%) · compare head-to-head

← 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	1.8 min	1.59 Wh	estimate, 1 of 2 stages measured
60-second AI short film	2.5 min	1.02 Wh	estimate, 1 of 3 stages measured
6-panel comic page	2.7 min	0.79 Wh	estimate, 1 of 3 stages measured
Character sheet, 12 poses	3.4 min	n/a	estimate, 0 of 2 stages measured
Short social clips	7.9 min	0.57 Wh	estimate, 1 of 3 stages measured
Full codebase review	9.7 min	31.21 Wh	measured
Product photo shoot	12.2 min	n/a	estimate, 0 of 2 stages measured
Long-form article batch	16.6 min	62.54 Wh	measured
Photo restoration batch	27.1 min	n/a	estimate, 0 of 1 stage measured

Rent or buy?

This card is $25,000 to buy. The cheapest listed rate on RunPod is $1.990/hour, but that is the floor: we budget $2.388/hour, a 20% premium, because idle time, storage and unavailable cheap instances all land on the same bill. At that rate buying wins after 10,469 GPU-hours. Below it you are paying for idle silicon.

How you would use it	GPU-hours a year	Rental cost a year	Time to break even
2 hours a day, hobby	730	$1,743	14.3 years
8 hours a day, working on it	2,920	$6,973	3.6 years
24/7, always-on agent	8,760	$20,919	1.2 years

At hobby usage this card is very unlikely to pay for itself before it is superseded. Rent it. Rental figures include a 20% premium over the cheapest listed rate. Ignores electricity, resale and the fact that a rented card can be a newer one tomorrow.

Rental price

$1.990/hr+0.0% since 2026-08-14low $1.936 · high $1.990

Cheapest of the RunPod and Vast on-demand rates we see, sampled daily. Spot and interruptible pricing runs lower.