H100 PCIe Price, 80GB VRAM & Local LLMs
NVIDIA's H100 PCIe (2022) has 80GB of VRAM and runs models up to 35B at 8-bit with full context. Native FP8 also makes it a good card for several users to share. The 94GB H100 NVL is a different card, so don't mix up the two.
eBay
H100 PCIe price history
eBay
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
H100 PCIe VRAM & full specs
- Memory80 GB · HBM2e
- Bandwidth2,000 GB/s
- FP16 / BF16756 TFLOPS
- FP8 / INT81,513 / 1,513 TOPS
- Bus width
- 5,120-bit
- Architecture
- Hopper (GH100)
- Cores
- 14,592
- Tensor cores
- 456
- Ecosystem
- CUDA
- Interface
- PCIe 5.0 x16
- NVLink
- 600 GB/s
- Power
- 350 W
- Power connector
- 16-pin
- Slots
- 2
- Length
- 268 mm
- Type
- Data center
- Launch date
- Mar 22, 2022
What LLMs can a single H100 PCIe run?
One H100 PCIe tops out at 70B-class and gpt-oss 120B at 4-bit. Up to 35B, use 8-bit: Qwen3.8 27B at about 49 tokens/s, and FP8 models run as-is.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 170.8 | 256K |
 Qwen3.6 27B27.8B | 11.8 GB | Runs | 107.9 | 256K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 125.6 | 256K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 178 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 104.4 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 184.4 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 154.3 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 79.8 | 256K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 81.1 | 256K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 165.7 | 198K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 72.4 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 167.9 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Runs | 133.9 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Runs | 49.4 | 256K |
| Qwen3.8 27B27.8B | 29 GB | Runs | 48.8 | 256K |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Runs | 144.6 | 198K |
| Gemma 4 31B30.7B | 32.6 GB | Runs | 43.2 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Runs | 147.9 | 256K |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Needs RAM offloadMoE only: expert weights sit in system RAM, so speed is estimated from 70 GB/s RAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
What LLMs can multiple H100 PCIe cards run?
Two H100 PCIe cards (160GB) run Qwen3.8 Flash Next at 4-bit at about 135 tokens/s; four run DeepSeek V4 Flash, and eight run trillion-parameter models. NVLink connects cards in pairs.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 2 cards | No price | 160 GB | 149 | 700 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 2 cards | No price | 160 GB | 130.7 | 700 W |
 Hy3295B (21B active) | 107.2 GB | 2 cards | No price | 160 GB | 76.3 | 700 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 2 cards | No price | 160 GB | 113.6 | 700 W |
| GLM 5.3744B (40B active) | 253.9 GB | 4 cards | No price | 320 GB | 68.8 | 1,400 W |
| Hy4 preview770B (49B active) | 307.2 GBest. | 8 cards | No price | 640 GB | 54.4 | 2,800 W |
 Muse Spark Open816B (42B active) | 311.2 GB | 8 cards | No price | 640 GB | 62.7 | 2,800 W |
 Kimi K2.7 Code1,000B (32B active) | 339.5 GB | 8 cards | No price | 640 GB | 79.9 | 2,800 W |
| Kimi K2.61,000B (32B active) | 340.5 GB | 8 cards | No price | 640 GB | 79.7 | 2,800 W |
 Mistral Large 41,000B (52B active) | 378.3 GB | 8 cards | No price | 640 GB | 54.7 | 2,800 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 2 cards | No price | 160 GB | 134.6 | 700 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 4 cards | No price | 320 GB | 106.5 | 1,400 W |
| Hy3295B (21B active) | 182.2 GB | 4 cards | No price | 320 GB | 60.3 | 1,400 W |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | 4 cards | No price | 320 GB | 82.3 | 1,400 W |
| GLM 5.3744B (40B active) | 467.3 GB | 8 cards | No price | 640 GB | 45.4 | 2,800 W |
| Hy4 preview770B (49B active) | 467.3 GB | 8 cards | No price | 640 GB | 39.9 | 2,800 W |
| Muse Spark Open816B (42B active) | 505.5 GB | 8 cards | No price | 640 GB | 44.5 | 2,800 W |
| DeepSeek V4.1 Flash763B (8 / 16B active) | 510.3 GB | 8 cardsEngram in RAM | No price | 640 GB | 73.3 | 2,800 W |
| Kimi K2.61,000B (32B active) | 583.7 GB | 8 cards | No price | 640 GB | 56.8 | 2,800 W |
| Kimi K2.7 Code1,000B (32B active) | 583.7 GB | 8 cards | No price | 640 GB | 56.8 | 2,800 W |
| Mistral Large 41,000B (52B active) | 604.8 GB | 8 cards | No price | 640 GB | 38.5 | 2,800 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 4 cards | No price | 320 GB | 104.3 | 1,400 W |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | 4 cards | No price | 320 GB | 109.4 | 1,400 W |
| Hy3295B (21B active) | 317.7 GB | 8 cards | No price | 640 GB | 43.7 | 2,800 W |
| GLM 5.3 Flash320B (18B active) | 341 GB | 8 cards | No price | 640 GB | 57.7 | 2,800 W |
| Kimi K2.61,000B (32B active) | 594.5 GB | 8 cards | No price | 640 GB | 56 | 2,800 W |
| Kimi K2.7 Code1,000B (32B active) | 594.5 GB | 8 cards | No price | 640 GB | 56 | 2,800 W |
Rows marked “Engram in RAM” need about 203 GB of extra system RAM. Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Running FP8 models
Hopper supports FP8 natively, so FP8 models get a real speedup, and FlashAttention 3 is available
- Long prompts, RAG and many concurrent users
It has 2.4 times the A100's compute
- Fitting 70B- to 120B-class models on one card
Llama 3.3 70B at 8-bit and gpt-oss 120B at 4-bit both fit
- Two cards linked with NVLink bridges
Qwen3.8 Flash Next at 4-bit runs at about 135 tokens/s
Not a good fit
- Just one person chatting
Generation speed on the same model is almost the same as the A100 80GB's, and the extra compute goes unused
- You don't have a server
Passive cooling depends on case airflow, and if the power cable signals too low a power tier, the card gets capped at 310W or won't run at all
Before you buy
Give the 16-pin connector enough power
The H100 PCIe uses a single 16-pin power connector. The card reads the sense pins on the cable to judge how much power the supply can deliver, then decides how fast to run: only a cable signaling the above-300W tier lets it run at the full 350W; the 151 to 300W tier caps it at 310W; and the 150W tier isn't supported at all. If you use a CPU 8-pin to 16-pin adapter cable, first check which tier it reports to the card.
Don't confuse it with the H100 NVL or SXM
Listings often lump several H100 variants together. The H100 NVL is a different PCIe card, with 94GB and 400W. SXM5 modules have to plug into a dedicated HGX baseboard and can't be used in a personal machine. When the card arrives, check the model in nvidia-smi: a factory H100 PCIe shows “H100 PCIe” (lspci device ID 2331; the NVL is 2321).
Airflow and room for NVLink
The card has no fan. Its heatsink accepts airflow from either direction, but the volume of air has to meet server standards. When installing NVLink bridges, leave at least 2.5 mm of clearance above the card and 2.67 mm behind it; you only get full bandwidth with all three bridges in place.
Software: Hopper is the main platform
New inference software almost always supports Hopper first: vLLM computes FP8 models directly in FP8, and FlashAttention 3 supports only the H100 and H800. MIG can split the card into up to seven 10GB slices, or into two 40GB slices. But if you're just running llama.cpp for one person, speed depends mainly on bandwidth, and the FP8 and compute advantages don't come into play.
Compared with similar GPUs
H100 PCIe vs A100 80GB PCIe and RTX PRO 6000: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | H100 PCIe | A100 80GB PCIe | RTX PRO 6000 |
|---|---|---|---|
| Memory | 80 GB | 80 GBSame | 96 GB20% more |
| Bandwidth | 2,000 GB/s | 1,935 GB/s3% lower | 1,792 GB/s10% lower |
| BF16 / FP16 compute | 756 TFLOPS | 312 TFLOPS59% lower | 500 TFLOPS34% lower |
| INT8 compute | 1,513 TOPS | 624 TOPS59% lower | 1,000 TOPS34% lower |
| Qwen3.8 27B speed (2-bit) | 125.6 t/s | 122.2 t/sSame | 114.6 t/s9% slower |
| Largest dense model (4-bit) | 136B | 136BSame | 165B21% more |
| Power | 350 W | 300 W14% lower | 600 W71% higher |
| GPU interconnect | NVLink 600 GB/s | NVLink 600 GB/s | Not supported |
| Interface | PCIe 5.0 x16 | PCIe 4.0 x16 | PCIe 5.0 x16 |
| Released | 2022-03 | 2021-06 | 2025-03 |
| eBay median |
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
H100 PCIe rental price per hour
On-demand rental of the H100 PCIe starts at $1.99 per hour (RunPod), about $484/month at 8 hours a day.
| Option | Price | 4 h/day | 8 h/day | 24 h/day | View |
|---|---|---|---|---|---|
 RunPodRentUpdated 29 min. ago | $1.99/hr | $242/mo | $484/mo | $1,453/mo | Rent |
 Vast.aiRentUpdated 29 min. ago | $3.54/hr | $431/mo | $862/mo | $2,585/mo | Rent |
Rental links are referral links: we may receive a commission or credit if you sign up through them.
Amazon.jp
Xianyu