H20: specs & price history
H20 (96 GB VRAM, NVIDIA) is in our GPU registry, but no provider we track lists a current price within the 2-day freshness window. This page stays live and updates automatically the day a tracked provider relists it.
H20 has not yet appeared in a tracked provider's live pricing. This page stays live and shows a price the day one lists it.
H20 price history and range
Price changes
- 30 days: No matching dated observations for this comparison.
- 90 days: No matching dated observations for this comparison.
Changes use the geometric mean of price ratios for the same provider, variant, tier, GPU count, instance and region (RunPod's place is where that day's stock sat, so it is not matched). New listings are not counted as price changes. Full ledger on the history hub.
H20 has no recorded snapshot yet. Charts appear as daily checks accumulate; history cannot be backfilled.
H20 hardware specs
| Architecture | Hopper |
|---|---|
| VRAM | 96 GB |
| FP16 TFLOPS (dense) | 148 |
| Memory bandwidth | 4,000 GB/s |
| TDP | 400 W |
| Interconnect | NVLink 900GB/s |
| Launched | 2024 |
Manufacturer datasheet figures for the H20. Source: viperatech.com.
LLM memory estimates for this GPU
These models have an estimated inference footprint within 96 GB at one or more precisions. The scenario uses 4,096 tokens, batch 1, a 16-bit KV cache and 20% runtime headroom. A memory fit does not establish speed or software compatibility.
| Model | 16-bit GB | 8-bit GB | 4-bit GB |
|---|---|---|---|
| Mixtral 8x22B v0.1 | 338.6 | 169.9 | 85.5 ✓ |
| GLM 4.5 Air | 266.1 | 133.5 | 67.2 ✓ |
| Qwen2.5 72B | 176.1 | 88.9 ✓ | 45.2 ✓ |
| DeepSeek R1 Distill Llama 70B | 170.9 | 86.3 ✓ | 43.9 ✓ |
| Mixtral 8x7B v0.1 | 112.7 | 56.7 ✓ | 28.7 ✓ |
| Qwen2.5 32B | 79.9 ✓ | 40.6 ✓ | 20.9 ✓ |
| Qwen2.5 Coder 32B | 79.9 ✓ | 40.6 ✓ | 20.9 ✓ |
| DeepSeek R1 Distill Qwen 32B | 79.9 ✓ | 40.6 ✓ | 20.9 ✓ |
| QwQ 32B | 79.9 ✓ | 40.6 ✓ | 20.9 ✓ |
| Qwen3 32B | 79.9 ✓ | 40.6 ✓ | 20.9 ✓ |
| Qwen3 30B A3B | 73.8 ✓ | 37.1 ✓ | 18.8 ✓ |
| Qwen3 30B A3B 2507 | 73.8 ✓ | 37.1 ✓ | 18.8 ✓ |
| Qwen3 Coder 30B A3B | 73.8 ✓ | 37.1 ✓ | 18.8 ✓ |
| Mistral Small 3.2 24B 2506 | 58.4 ✓ | 29.6 ✓ | 15.2 ✓ |
| Mistral Small 24B 2501 | 57.4 ✓ | 29.1 ✓ | 14.9 ✓ |
| Devstral Small 2507 | 57.4 ✓ | 29.1 ✓ | 14.9 ✓ |
| Qwen2.5 14B | 36.4 ✓ | 18.7 ✓ | 9.8 ✓ |
| Qwen2.5 Coder 14B | 36.4 ✓ | 18.7 ✓ | 9.8 ✓ |
| DeepSeek R1 Distill Qwen 14B | 36.4 ✓ | 18.7 ✓ | 9.8 ✓ |
| Qwen3 14B | 36.2 ✓ | 18.5 ✓ | 9.7 ✓ |
| phi 4 | 36.2 ✓ | 18.6 ✓ | 9.8 ✓ |
| Phi 4 reasoning | 36.2 ✓ | 18.6 ✓ | 9.8 ✓ |
| Phi 4 reasoning plus | 36.2 ✓ | 18.6 ✓ | 9.8 ✓ |
| Mistral Nemo 2407 | 30.2 ✓ | 15.5 ✓ | 8.2 ✓ |
| Qwen3 8B | 20.4 ✓ | 10.6 ✓ | 5.6 ✓ |
| DeepSeek R1 Distill Llama 8B | 19.9 ✓ | 10.3 ✓ | 5.5 ✓ |
| Qwen2.5 7B | 18.6 ✓ | 9.4 ✓ | 4.9 ✓ |
| Qwen2.5 Coder 7B | 18.6 ✓ | 9.4 ✓ | 4.9 ✓ |
| DeepSeek R1 Distill Qwen 7B | 18.6 ✓ | 9.4 ✓ | 4.9 ✓ |
| Mistral 7B v0.3 | 18.0 ✓ | 9.3 ✓ | 5.0 ✓ |
| Qwen3 4B | 10.4 ✓ | 5.6 ✓ | 3.1 ✓ |
| Qwen3 4B 2507 | 10.4 ✓ | 5.6 ✓ | 3.1 ✓ |
| Qwen2.5 3B | 7.6 ✓ | 3.9 ✓ | 2.0 ✓ |
| Qwen2.5 Coder 3B | 7.6 ✓ | 3.9 ✓ | 2.0 ✓ |
| Qwen3 1.7B | 5.4 ✓ | 3.0 ✓ | 1.8 ✓ |
| DeepSeek R1 Distill Qwen 1.5B | 4.4 ✓ | 2.3 ✓ | 1.2 ✓ |
| Qwen2.5 1.5B | 3.8 ✓ | 2.0 ✓ | 1.1 ✓ |
| Qwen2.5 Coder 1.5B | 3.8 ✓ | 2.0 ✓ | 1.1 ✓ |
| Qwen3 0.6B | 2.4 ✓ | 1.5 ✓ | 1.0 ✓ |
| Qwen2.5 0.5B | 1.2 ✓ | 0.7 ✓ | 0.4 ✓ |
| Qwen2.5 Coder 0.5B | 1.2 ✓ | 0.7 ✓ | 0.4 ✓ |
Models that need more than one H20
- Qwen3 235B A22B at 16-bit needs about 565 GB with the same context and headroom, so at least 6 cards. No provider we track publishes a H20 node of that size today. At 8 or 4 bit it needs less; the model page recalculates.
- Qwen3 235B A22B 2507 at 16-bit needs about 565 GB with the same context and headroom, so at least 6 cards. No provider we track publishes a H20 node of that size today. At 8 or 4 bit it needs less; the model page recalculates.
- Mixtral 8x22B v0.1 at 16-bit needs about 339 GB with the same context and headroom, so at least 4 cards. No provider we track publishes a H20 node of that size today. At 8 or 4 bit it needs less; the model page recalculates.
Compare model memory requirements or change context and memory assumptions on a model's page.
When to choose the H20, and what to pick instead
Start from memory: your model, its context and your batch must fit in 96 GB (the model-fit table above estimates it). Then the H20 is the right card unless one of these fits your job better.
Rent or buy the H20
We have no confirmed H20 purchase price from the stores we check. A seller quote entered in the calculator gives the break-even.
Compare H20 purchase prices and quote availability, then enter your usage and electricity costs in the rent-versus-buy calculator. Purchase listings have their own variants, stock status and update dates. Compare two GPUs.
H20 questions
- How much VRAM does an H20 have?
- 96 GB.
- Is H20 available to rent right now?
- Not yet from a provider we track. H20 is in our GPU registry; this page shows a live price the day a tracked provider lists it.
GPUs near the H20
Related
All 80 GB+ GPUs · Price movers this week · Price history hub