Specifications Compared
| Spec | GH200 | H100 |
|---|---|---|
| TDP | 900W | 700W |
| VRAM | 96 GB | 80-94 GB |
| CUDA Cores | 16,896 | 16,896 |
| FP8 (dense) | 1,979 TFLOPS | 1,979 TFLOPS |
| Memory Type | HBM3 | HBM3 |
| Architecture | Hopper | Hopper |
| FP16 (dense) | 989 TFLOPS | 989 TFLOPS |
| Form Factors | SXM | SXM5, PCIe, NVL |
| INT8 (dense) | 1,979 TOPS | 1,979 TOPS |
| Interconnect | NVLink-C2C, PCIe 5.0 | NVLink, PCIe 5.0, InfiniBand |
| Tensor Cores | 528 | 528 |
| FP32 Performance | 67 TFLOPS | 67 TFLOPS |
| FP64 Performance | 34 TFLOPS | 34 TFLOPS |
| Memory Bandwidth | 4,000 GB/s | 3,350 GB/s |
| FP8 (with sparsity) | 3,958 TFLOPS | 3,958 TFLOPS |
| FP16 (with sparsity) | 1,979 TFLOPS | 1,979 TFLOPS |
| INT8 (with sparsity) | 3,958 TOPS | 3,958 TOPS |
Performance Analysis
The FP16 dense figure of 989 TFLOPS on both GPUs produces the same throughput for dense matrix operations during training and inference workloads. The FP16 with sparsity figure of 1979 TFLOPS likewise matches exactly between the two. Higher memory bandwidth of 4000 GB/s on the GH200 compared with 3350 GB/s on the H100 allows larger batch sizes before memory becomes the limiting factor. The 96 GB HBM3 capacity on the GH200 versus the 80 to 94 GB range on the H100 further extends the feasible model size before swapping or reduced batching occurs. FP8 dense performance remains 1979 TFLOPS on both devices so quantization benefits stay equivalent when sparsity is not applied.
Current On-Demand Offers
Cheapest secure on-demand offer per provider that is reported in stock right now, per GPU per hour; multi-GPU instances show the whole-instance price alongside. Deploy opens the DeployGPU console for providers on the platform, otherwise the provider's own site. Stock is rechecked every minute.
GH200
GH200 is not offered on-demand by any provider we track right now. See the GH200 rental page for last-seen listed prices and a price alert.
H100
| Provider | Region | GPUs | Per GPU / hr | Instance / hr | Deploy |
|---|---|---|---|---|---|
| Hyperstack | CANADA-1 | 2 | $2.50 | $5.00 | Deploy |
| QuantaCloud | us-midwest-2 | 2 | $2.59 | $5.18 | Deploy |
| Massed Compute | us-central-3 | 1 | $2.73 | — | Deploy |
| Ori | dallas-2 | 4 | $2.90 | $11.60 | Deploy |
| Scaleway | Paris, France (PAR2) | 1 | $3.23 | — | Deploy |
9 providers in stock, 19 offers (cheapest per provider shown). All H100 offers, price history and alerts
Notify me when GH200 drops below a price
One email when the cheapest in-stock on-demand price per GPU-hour falls below your threshold.
QuantaCloud
Comparing H-series providers? We broker across all of them.
Hopper stock changes by the hour and prices differ by provider. If you need 16+ GPUs reserved or a cluster in the next 90 days, we quote H-series or B300 inventory at partner rates: one quote, 24h turnaround.
When to Choose the GH200
The GH200 suits workloads that require the full 96 GB HBM3 capacity together with 4000 GB/s bandwidth. Training runs that exceed the 94 GB limit on the H100 benefit directly from the additional memory. Inference pipelines that move large intermediate activations also gain from the bandwidth increase at the cost of 900W TDP.
When to Choose the H100
The H100 fits deployments constrained to 700W TDP or that require PCIe or NVL form factors. Multi-node clusters that rely on InfiniBand interconnect select the H100 for its native support. Applications that fit within 80 to 94 GB HBM3 avoid the extra power draw of the GH200.
Use Cases
The GH200 supplies 96 GB HBM3 and 4000 GB/s bandwidth that exceed the H100 limits for large batch training.
The GH200 4000 GB/s bandwidth supports larger concurrent request batches than the H100 3350 GB/s.
Both GPUs deliver the same 989 TFLOPS FP16 dense and 1979 TFLOPS FP16 with sparsity so fine-tuning throughput matches when memory fits.
The H100 700W TDP and multiple form factors suit image generation clusters that do not require the GH200 memory capacity.
The GH200 67 TFLOPS FP32 combined with 4000 GB/s bandwidth accelerates memory-bound scientific kernels beyond the H100.
Frequently Asked Questions
What is the memory bandwidth difference between the two GPUs?▾
The GH200 reaches 4000 GB/s while the H100 reaches 3350 GB/s. This gap affects achievable batch sizes in bandwidth-limited phases.
Do the GPUs share the same FP16 performance numbers?▾
Both list 989 TFLOPS FP16 dense and 1979 TFLOPS FP16 with sparsity. Training and inference throughput therefore match when memory capacity is not the constraint.
How do the power limits compare?▾
The GH200 TDP is 900W while the H100 TDP is 700W. Systems must accommodate the higher power draw when selecting the GH200.
Which interconnect options exist on each GPU?▾
The GH200 uses NVLink-C2C and PCIe 5.0. The H100 uses NVLink, PCIe 5.0 and InfiniBand depending on the chosen form factor.
Which is cheaper to rent, the GH200 or the H100?▾
Cloud rental prices for both the GH200 and H100 vary by provider, configuration, and availability. This page shows live pricing from 25+ providers updated every 60 seconds. Scroll to the Live Cloud Pricing section to compare current rates.
How much VRAM does the GH200 have compared to the H100?▾
The GH200 has 96 GB of HBM3 memory. The H100 has 80 to 94 GB of HBM3 memory.
Can I find GH200 and H100 GPUs available to rent right now?▾
Yes. This page shows real-time availability across 25+ cloud GPU providers. The Live Cloud Pricing section displays only in-stock offers with current pricing.
What is the main difference between the GH200 and the H100?▾
The GH200 uses the Hopper architecture (2023) while the H100 uses Hopper (2022). Both deliver the same dense FP16 throughput (989 TFLOPS without sparsity), while the GH200 has 1.2x the memory bandwidth of the H100.
Rent these GPUs
Each GPU page lists every current on-demand offer by provider, per GPU-hour, updated every minute.
Related comparisons
How this page is made
- Specifications come from the NVIDIA, AMD and Intel datasheets for the GH200 and the H100. Dense and sparse throughput are listed separately.
- Prices and availability are live: every offer shown is an in-stock, on-demand listing from a provider we track, rechecked every minute.
- The written comparison (overview, performance notes, when to choose each card, verdict, use cases and FAQ) was drafted with an AI model from the specification table above and passed an automated check that rejects any figure not in that table. It was last generated on . It contains no prices; those are always read live.
- Read how we collect the data, or report an error on this page.
Next steps
- Rent GH200 Grace HopperEvery current offer by provider, daily price history and a price alert.
- Rent H100Every current offer by provider, daily price history and a price alert.
- GPU price indexHow on-demand prices for the major GPUs have moved, updated daily.
- All GPUsEvery GPU we track, with current lows and provider counts.