Side-by-side GPU comparison: specs, memory, compute performance, and live cloud pricing.
Verdict
GH200 has 48 GB more VRAM, making it better suited for large models and long context windows. For compute-bound workloads like training, GH200 delivers 6.6× higher FP16 throughput. At $0.44/hr vs $2.29/hr, A40 is the more cost-efficient choice for inference. GH200 supports a broader range of models (637 vs 624 from this catalog), giving more flexibility.
Specifications
| A40 | GH200 | |
|---|---|---|
| VRAM | 48 GB | 96 GB |
| VRAM Type | GDDR6 | HBM3 |
| Memory Bandwidth | 0.7 TB/s | 4.0 TB/s |
| FP16 Performance | 150 TFLOPS | 990 TFLOPS |
| Manufacturer | NVIDIA | NVIDIA |
| FP8 Support | No | Yes |
| FP4 Support | No | No |
Price / Performance
Based on cheapest single-GPU on-demand pricing. Lower $/TFLOP = better compute value; lower $/GB = better memory value.
| A40 | GH200 | |
|---|---|---|
| $/hr (cheapest) | $0.44✓ best | $2.29 |
| $/TFLOP (compute value) | $0.0029 | $0.0023✓ best |
| $/GB VRAM (memory value) | $0.0092✓ best | $0.0239 |
Cloud Pricing
Cheapest on-demand price per provider (single GPU).
Model Compatibility
Models from the catalog that fit on each GPU, grouped by required precision.
A40 (624 models)
GH200 (637 models)
You might also compare…
Pricing data refreshed hourly · Last updated April 11, 2026 · Browse all comparisons