NVIDIA H100 SXM5 96 GB
Live eBay prices, cheapest first. The best one right now is $40,000 (new).
Not a PCIe card. SXM module: needs an SXM-to-PCIe adapter board, its own cooling and power to run in a normal PC. The PCIe version: NVIDIA H100 PCIe 96 GB, from $29,802 →
Specs
- VRAM
- 96 GB HBM3
- Bandwidth
- 3,350 GB/s
- FP32
- 67 TFLOPS
- FP16
- 989 TFLOPS
- Power
- 700 W
- Connector
- SXM5 (PCIe 5.0)
- Released
- 2022
- CUDA compute
- 9.0
- Llama 3.1 8B at Q4 (est.)
- 345–518 tokens/s
Price right now
Live on eBayFrom the live listings below; shipping included.
Listings
Live on eBay| Listing | Condition | Price | Checked |
|---|---|---|---|
| New | $40,000or best offerfree shipping |
Auctions ending soon
No auctions ending soon right now.
Prices include shipping and were read from eBay at the time shown. They can change: check the listing before you buy.
Owner reviews
No owner reviews yet. Own one? Tell others how it runs.
From signed-in owners, checked before they're published. PerGB doesn't pay for or edit reviews.
How it compares
Live on eBay- Software: CUDA 11.8 and newer.
As much VRAM, about as fast, for less
- AMD Instinct MI300X: $27,550, 1021% of the GeForce RTX 3090, 192 GB
- NVIDIA H200 NVL: $32,300, 777% of the GeForce RTX 3090, 141 GB
More VRAM, for not much more
- AMD Instinct MI250: $4,995, 422% of the GeForce RTX 3090, 128 GB
- AMD Instinct MI300X: $27,550, 1021% of the GeForce RTX 3090, 192 GB
- NVIDIA H200 NVL: $32,300, 777% of the GeForce RTX 3090, 141 GB
From the live listings of every GPU, shipping included; performance against the GeForce RTX 3090 (FP16 or FP32 compute and memory bandwidth). Updated with every build.
Local AI: what it runs
| Model (Q4, 8K context) | H100 SXM5 96 GB needed | Costs now | Speed (est.) |
|---|---|---|---|
| Llama 3.1 8B | 1 card | $40,000 | 345–518 tokens/s |
| Qwen3 14B | 1 card | $40,000 | 187–280 tokens/s |
| gpt-oss 20B | 1 card | $40,000 | 460–691 tokens/s |
| Gemma 3 27B | 1 card | $40,000 | 101–151 tokens/s |
| Qwen3 32B | 1 card | $40,000 | 84–126 tokens/s |
| Llama 3.3 70B | 1 card | $40,000 | 39–59 tokens/s |
| gpt-oss 120B | 1 card | $40,000 | 325–488 tokens/s |
| Qwen3 235B-A22B | 2 cards | not enough listings | 75–113 tokens/s |
| DeepSeek R1 / V3 671B | 5 cards | not enough listings | 45–67 tokens/s |
Cost: that many of the cheapest live listings. Speed: an estimate from the memory bandwidth for one person chatting. Try other models, sizes and contexts.
Decision models
| Model (Q4, 8K context) | H100 SXM5 96 GB needed | Costs now | Per question (est.) |
|---|---|---|---|
| Julia-10.144B | 1 card | $40,000 | 1.1–2.3 ms |
| Laya0.421B | 1 card | $40,000 | 1.8–3.9 ms |
| Kev-4B4B | 1 card | $40,000 | 4.4–9.4 ms |
| lev4B | 1 card | $40,000 | 13–28 ms |
| OpenJev27B | 1 card | $40,000 | 16–34 ms |
| Clef27B | 1 card | $40,000 | 16–34 ms |
| Perplexity Decider v1.227B | 1 card | $40,000 | 16–34 ms |
| mammoth-127B | 1 card | $40,000 | 16–34 ms |
Decision models score the options you give them instead of writing text, so they're timed per question. The time is scaled from each model's reported speed on an RTX PRO 6000 by this card's compute and memory bandwidth, within about half either way. About the models.
Compare elsewhere
Search other shops for NVIDIA H100 SXM5 96 GB. These open the shop's own search; we don't show their prices here.
