COMPARE / H20 VS H800
H20 vs H800
Specs from the chip makers' datasheets; prices from the live Compute Cafe panel, refreshed hourly. Every rate is an on-demand published price with a source receipt.
CHEAPEST H20
$1.010/hr
PPIO · median $1.29 · 3 providers
CHEAPEST H800
$1.560/hr
AutoDL · median $1.79 · 2 providers
CHEAPER RIGHT NOW
H20
live panel, 2026-09-27
| H20 | H800 | |
|---|---|---|
| MAKER | NVIDIA | NVIDIA |
| ARCHITECTURE | Hopper | Hopper |
| LAUNCHED | 2024 | 2023 |
| MEMORY | 96 GB HBM3 | 80 GB HBM3 |
| MEMORY BANDWIDTH | 4.0 TB/s | 3.35 TB/s |
| TENSOR COMPUTE | 148 TFLOPS FP16/BF16 tensor (296 TFLOPS FP8, 44 TFLOPS FP32) | 1,513 TFLOPS FP16/BF16 tensor (3,026 TFLOPS FP8, TF32 756 TFLOPS) |
| POWER DRAW | 400 W | 700 W (SXM5; 350-400 W PCIe/NVL SKUs) |
| INTERCONNECT | SXM5 (HGX H20, NVLink) | NVLink 400 GB/s (SXM5), PCIe 5.0 |
HOW TO CHOOSE
H20 vs H800
The H800 is the earlier export-compliant Hopper: far more compute (about 10x the H20's FP16) but 80 GB and lower interconnect. For training under the rules, the H800 wins; for pure inference-per-dollar, the H20's newer memory system competes.
H800 vs H20
The H20 is the later, tighter compliance cut: 96 GB of memory but only 148 TFLOPS FP16. The H800 has roughly 10x the compute; the H20 has the newer memory system. Training says H800, memory-per-dollar inference says H20.
Workload fit
H20: large-model inference, memory-bound hpc and analytics, llm training. H800: llm training, large-batch inference, fp64 hpc.
H20 VS H800 - FAQ
Which is cheaper to rent, H20 or H800?
By today's published on-demand rates: H20 starts at $1.010/GPU/hr (median $1.29, 3 providers) and H800 starts at $1.560/GPU/hr (median $1.79, 2 providers). The cheapest live listing is the H20 at PPIO. Figures recompute from the panel on every refresh.
H20 or H800 - which should I rent?
H20: The H20 is the China-market Hopper: full 96 GB of HBM3 and 4.0 TB/s of memory bandwidth, but tensor compute capped at 148 TFLOPS of FP16 - a fraction of the H100's. H800: The H800 was the first export-compliant Hopper for the China market: near-H100 tensor throughput but NVLink cut from 900 to 400 GB/s and FP64 essentially removed. The side-by-side spec table and live price rows above give the full picture.
Where can I rent H20 and H800?
3 providers publish H20 rates and 2 publish H800 rates on Compute Cafe. Every row links to the provider's own published price - committed-use and reserved rates are excluded.
Full pricing: H20 · H800 · all comparisons