CDNA 3
AMD Instinct MI300X
wins
5
The AMD MI300X leads all GPUs on memory capacity (192GB HBM3) and bandwidth (5.3 TB/s), making it exceptional for serving very large language models that don't fit on H100.
Hopper
NVIDIA H200 141GB
wins
1
The H200 upgrades the H100 with 141GB HBM3e memory and 4.8 TB/s bandwidth — a massive boost for memory-bound workloads like large LLM inference.
Performance comparison
Visual bars — winner highlighted
AMD Instinct MI300X
Metric
NVIDIA H200 141GB
FP16 TFLOPS
Bandwidth
VRAM
TDP (lower=better)
TFLOPS/watt
GB/s per watt
TFLOPS/$
FP16 TFLOPS ÷ best on-demand $/hr
GB/s per $
Memory bandwidth ÷ best on-demand $/hr
Full specification table
| Spec | AMD Instinct MI300X | NVIDIA H200 141GB |
|---|---|---|
| Architecture | CDNA 3 | Hopper |
| VRAM | 192 GB HBM3 | 141 GB HBM3e |
| VRAM type | HBM3 | HBM3e |
| Memory bandwidth | 5,300 GB/s | 4,800 GB/s |
| FP16 TFLOPS | 1,307 TFLOPS | 989 TFLOPS |
| TDP | 750 W | 700 W |
| TFLOPS/watt | 1.74 T/W | 1.41 T/W |
| GB/s per watt | 7.07 GB/s/W | 6.86 GB/s/W |
| NVLink | No | Yes |
| Release year | 2023 | 2024 |
| Best cloud price | $1.85/hr | $2.61/hr |
| TFLOPS/$ | 706.5 | 378.9 |
| GB/s per $ | 2865 | 1839 |
Live cloud pricing
On-demand pricing across providers — updated in real time
AMD Instinct MI300X
NVIDIA H200 141GB
Power efficiency analysis
TFLOPS/watt and GB/s/watt — critical for data center TCO
AMD Instinct MI300X
NVIDIA H200 141GB
TFLOPS/watt measures compute efficiency — how much AI throughput you get per watt of power consumed. For data centers with PUE of 1.2–1.5, a 10% improvement in TFLOPS/watt translates directly to lower electricity costs and cooling requirements. GB/s/watt measures memory bandwidth efficiency, which is the binding constraint for memory-bound LLM inference workloads.
Performance-per-dollar
TFLOPS/$ and GB/s/$ at current best cloud prices
AMD Instinct MI300X
NVIDIA H200 141GB
TFLOPS/$ = FP16 TFLOPS ÷ best on-demand $/hr. Higher is better — more compute per dollar. GB/s/$ = memory bandwidth ÷ best on-demand $/hr. For memory-bound inference workloads (most LLM serving), GB/s/$ is often a better predictor of cost-efficiency than TFLOPS/$. Prices are live from the Compute Comparison API and update in real time.
Workload Cost Comparison
Total job cost = best on-demand $/hr × GPU count × hours
| Workload | Scale | AMD Instinct MI300X | NVIDIA H200 141GB |
|---|---|---|---|
| 7B LLM fine-tune | 8× GPU, 24 h | $355▼ | $501 |
| 70B LLM fine-tune | 64× GPU, 72 h | $8.5k▼ | $12k |
| Inference server | 1× GPU, 730 h | $1.4k▼ | $1.9k |
| Image gen batch | 4× GPU, 8 h | $59▼ | $84 |
Based on current best on-demand prices. Spot pricing typically 40–70% lower. Actual cost depends on provider, region, and availability.
When to choose each GPU
Choose AMD Instinct MI300X for:
- Memory-bound LLM inference
- Large model serving
- HPC
Weaknesses / Limitations
- ROCm software ecosystem is less mature than CUDA — some frameworks require extra porting effort
- Fewer cloud providers offer MI300X compared to H100
- 750W TDP — high power draw comparable to H100 SXM
Choose NVIDIA H200 141GB for:
- Very large LLMs
- Memory-bound inference
- Multi-modal models
Weaknesses / Limitations
- Very limited cloud availability — fewer providers than H100
- Premium pricing over H100 with same FP16 TFLOPS
- 700W TDP — same power envelope as H100 with no efficiency gain
Popular GPU comparisons