Edited by 
Short answer
The AMD Instinct MI355X and NVIDIA B200 both have 8 TB/s of HBM3e bandwidth, but the MI355X carries 288 GB against the B200's 180 GB, about 1.6 times as much. MI355X also has about 10% more dense FP16, FP8 and FP4 throughput, more than twice the FP6 and FP64 throughput, and draws 1,400 W against 1,000 W. In MLPerf Inference 6.0, AMD reported single-node results from roughly equal to about 19% ahead of B200, depending on the test. B200's advantages are CUDA software and NVLink scaling to 72 GPUs in GB200 NVL72, where MI355X tops out at 8 GPUs per node.
MI355X vs B200 specifications
| Specification | AMD Instinct MI355X | NVIDIA B200 (HGX) | Difference |
|---|---|---|---|
| Architecture | CDNA 4 | Blackwell | |
| Process | TSMC N3P compute, N6 base dies | TSMC 4NP | |
| Transistors | 185 billion | 208 billion | |
| Memory | 288 GB HBM3e | 180 GB HBM3e | 1.6× MI355X |
| Memory bandwidth | 8 TB/s | 8 TB/s | equal |
| FP16 / BF16 (dense) | 2.5 PFLOPS | 2.25 PFLOPS | 1.11× MI355X |
| FP8 (dense) | 5.0 PFLOPS | 4.5 PFLOPS | 1.11× MI355X |
| FP6 (dense) | 10.1 PFLOPS | 4.5 PFLOPS | 2.2× MI355X |
| FP4 (dense) | 10.1 PFLOPS | 9 PFLOPS | 1.12× MI355X |
| FP64 | 78.6 TFLOPS | 37 TFLOPS | 2.1× MI355X |
| Power | 1,400 W (liquid cooled) | 1,000 W | 1.4× MI355X |
| GPU-to-GPU links | 1,075 GB/s Infinity Fabric | 1,800 GB/s NVLink 5 | 1.7× B200 |
| Largest scale-up domain | 8 GPUs | 8 (HGX) or 72 (GB200 NVL72) | |
| Software | ROCm | CUDA | |
| Launched | June 2025 | 2024 |
All throughput figures are per GPU and dense, without 2:4 sparsity. See every number side by side with 26 other chips in the AI GPU specs database.
Memory: the MI355X's biggest advantage
Both chips read memory at 8 TB/s, so a model that fits on one GPU generates tokens at a similar speed on either. The difference is how much fits. With 288 GB, one MI355X can hold:
- A 70B model at FP16 (about 141 GB) with plenty left over for the KV cache, where a B200 is tighter.
- Llama 3.1 405B at FP4 (about 203 GB of weights) on one GPU. A B200 needs two.
- Llama 3.1 405B at FP8 (about 405 GB) on two GPUs. B200 needs three.
Fewer GPUs per model means less splitting, less cross-GPU traffic and more room for concurrent users. Check any model in the VRAM calculator.
Compute: close, with two exceptions
For the formats most AI work uses, FP16, BF16, FP8 and FP4, the MI355X is about 11% ahead on paper. The two large gaps are:
- FP6: MI355X runs FP6 at its FP4 rate, 10.1 PFLOPS. B200 runs FP6 at its FP8 rate, 4.5 PFLOPS. This matters if you serve models quantized to 6 bits, a middle ground between FP8 accuracy and FP4 size.
- FP64: MI355X has 78.6 TFLOPS against 37 for B200, which suits scientific computing and simulation.
Peak numbers are ceilings. How close real work gets depends on kernels and libraries, which is where software comes in.
MLPerf Inference 6.0 results
In April 2026 AMD published its MLPerf Inference 6.0 results and compared them with NVIDIA's submissions. These ratios are AMD's own calculations from results verified by MLCommons, and single-node means 8 GPUs.
| Benchmark (single node) | MI355X vs B200 | MI355X vs B300 |
|---|---|---|
| Llama 2 70B Offline | tied | 92% |
| Llama 2 70B Server | 97% | 93% |
| Llama 2 70B Interactive | 119% | 104% |
| gpt-oss-120b Offline | 111% | 91% |
| gpt-oss-120b Server | 115% | 82% |
The picture: MI355X roughly matches or beats B200 on single-node inference and trails the newer B300. AMD also showed MI355X clusters passing 1 million tokens per second on Llama 2 70B across 87 GPUs, with about 93% scaling efficiency.
Scaling: NVLink vs Infinity Fabric
An MI355X node links 8 GPUs with Infinity Fabric. Beyond that, traffic goes over the network. B200 also comes in 8-GPU HGX servers, but its sibling GB200 joins 72 GPUs in one NVLink domain in GB200 NVL72 racks. For very large mixture-of-experts models that spread experts across dozens of GPUs, that larger domain is a real advantage. AMD's answer is the MI455X in its Helios rack, which also links 72 GPUs and began volume deployment in the second half of 2026.
Software: CUDA vs ROCm
CUDA has nearly two decades of libraries, tools and tuned kernels, and new models and research code usually run on NVIDIA first. ROCm has improved quickly: PyTorch, vLLM and SGLang support Instinct GPUs, and popular open models typically run on day one. The gap shows most in custom kernels, less common models and smaller teams without time to tune. For standard serving of popular open models, it is much narrower than it was in 2023.
Which should you choose?
- Choose MI355X to serve large models with fewer GPUs, for FP6 or FP64 work, or when availability and price are better, which is often the case for AMD in the cloud.
- Choose B200 when your stack depends on CUDA-only tools, when you need the 72-GPU NVLink domain of GB200 NVL72, or when your team has no time to validate a second platform.
- Looking ahead? The next round is AMD MI455X against NVIDIA Rubin. Read NVIDIA Rubin explained and compare both in the specs database.
Frequently asked questions
Is the MI355X faster than the B200?
On paper, the MI355X has about 11% more dense FP16, FP8 and FP4 throughput, the same memory bandwidth and 60% more memory. In AMD's MLPerf Inference 6.0 comparison it ranged from roughly equal to 19% faster than B200 on single-node tests. Results depend on the model, batch size and software.
What is the difference between MI355X and MI350X?
They use the same chip and the same 288 GB of HBM3e. The MI350X runs at up to 1,000 W with air cooling and somewhat lower clocks. The MI355X runs at up to 1,400 W with liquid cooling for higher throughput.
Does the MI355X support FP4?
Yes. CDNA 4 added FP4 and FP6 matrix support in the OCP microscaling formats, MXFP4 and MXFP6, at about 10 PFLOPS dense. NVIDIA's B200 supports MXFP4 and its own NVFP4.
Can MI355X run CUDA code?
Not directly. Code written for CUDA has to be ported to HIP, AMD's similar programming interface, which AMD's HIPIFY tools partly automate. Code written in PyTorch or served through vLLM usually runs without changes.
