// Hardware catalog
Every GPU worth running.
From H100 to GB200 NVL72 — the full NVIDIA lineup, available on-demand and reserved.
// Detailed specs
H100
LLM training, inference
- VRAM
- 80 GB HBM3
- Bandwidth
- 3.35 TB/s
- FP8 throughput
- 3958 TFLOPS
- NVLink
- 900 GB/s
H200
Frontier training, long-context inference
- VRAM
- 141 GB HBM3e
- Bandwidth
- 4.8 TB/s
- FP8 throughput
- 3958 TFLOPS
- NVLink
- 900 GB/s
GB200 NVL72
Trillion-parameter training
- VRAM
- 13.5 TB HBM3e
- Bandwidth
- 576 TB/s
- FP8 throughput
- 1.44 EFLOPS
- NVLink
- 1.8 TB/s per GPU
B200
Next-gen training
- VRAM
- 192 GB HBM3e
- Bandwidth
- 8 TB/s
- FP8 throughput
- 9 PFLOPS
- NVLink
- 1.8 TB/s
L40S
Inference, rendering, fine-tuning
- VRAM
- 48 GB GDDR6
- Bandwidth
- 864 GB/s
- FP8 throughput
- 1466 TFLOPS
- NVLink
- —
A100
Established training pipelines
- VRAM
- 80 GB HBM2e
- Bandwidth
- 2 TB/s
- FP8 throughput
- —
- NVLink
- 600 GB/s
// Comparison
6 results
| A100 | 80 GB HBM2e | 2 TB/s | — | 600 GB/s | In stock |
| B200 | 192 GB HBM3e | 8 TB/s | 9 PFLOPS | 1.8 TB/s | Waitlist |
| GB200 NVL72 | 13.5 TB HBM3e | 576 TB/s | 1.44 EFLOPS | 1.8 TB/s per GPU | Reserved |
| H100 | 80 GB HBM3 | 3.35 TB/s | 3958 TFLOPS | 900 GB/s | In stock |
| H200 | 141 GB HBM3e | 4.8 TB/s | 3958 TFLOPS | 900 GB/s | In stock |
| L40S | 48 GB GDDR6 | 864 GB/s | 1466 TFLOPS | — | In stock |