GPU comparison
H100 vs MI300X.
The MI300X is ahead on memory (192 GB vs 80 GB) and speed (at least 32% more FP16 throughput and at least 58% more memory bandwidth). The H100 makes sense when your model fits in 80 GB and its hourly price is lower.
Rent either on Runcrate
- H100Cheaperfrom $3.15/hr
$0.039 per GB/hrper TFLOPS: varies by card
- MI300Xfrom $4.85/hr per GPU
on an 8-GPU machine
$0.025 per GB/hr$0.37 per 100 TFLOPS/hr
Live hourly price, then per GB of memory and per 100 FP16 TFLOPS.
Which is better for
Which to pick for each job. Memory fit and throughput, from the datasheets.
Hold a 70B model in 8-bit weights
70 GB of 8-bit weights
H100: 1 GPU, 10 GB spare
MI300X: 1 GPU, 122 GB spare
MI300X. Same GPU count, at least 58% more memory bandwidth.
Serve an 8B model in 16-bit weights
16 GB of 16-bit weights
H100: 1 GPU, 64 GB spare
MI300X: 1 GPU, 176 GB spare
MI300X. Same GPU count, at least 58% more memory bandwidth.
Fine-tune a 70B model with QLoRA
About 46 GB with 4-bit weights
H100: 1 GPU, 34 GB spare
MI300X: 1 GPU, 146 GB spare
MI300X. Same GPU count, at least 32% more FP16 throughput.
Train on a full node
FP16 tensor throughput per node
H100: 8x, up to 7,912 TFLOPS
MI300X: 8x, 10,459 TFLOPS
MI300X. at least 32% more FP16 throughput per GPU.
Throughput per watt
FP16 TFLOPS per 100 W of TDP
H100: Varies by card
MI300X: 174 TFLOPS per 100 W
Either. Varies by card.
Compare another pair
Hopper
NVIDIA H100
- Memory
- 80 GB HBM3
- FP16 tensor
- up to 989 TFLOPS
- Bandwidth
- up to 3.4 TB/s
CDNA 3
AMD MI300X
on an 8-GPU machine
- Memory
- 192 GB HBM3
- FP16 tensor
- 1307.4 TFLOPS
- Bandwidth
- 5.3 TB/s
At a glance
Spec by spec. The longer bar is the better number.
Full specs
The complete spec sheet.
| Spec | NVIDIA H100 | AMD MI300X |
|---|---|---|
| Architecture | Hopper | CDNA 3 |
| Memory | 80 GB | 192 GB |
| Memory type | HBM3 | HBM3 |
| Memory bandwidth | up to 3.4 TB/s | 5.3 TB/s |
| FP32 | up to 67 TFLOPS | 163 TFLOPS |
| FP16 tensor | up to 989 TFLOPS | 1307.4 TFLOPS |
| INT8 | up to 1979 TOPS | 2610 TOPS |
| TDP | up to 700 W | 750 W |
| GPUs per node | 8 | 8 |
| Hourly price (live) | from $3.15/hr | from $4.85/hr per GPU on an 8-GPU machine |
Questions
H100 or MI300X. What buyers ask first.
Other comparisons