Hugging Face on AMD Instinct MI455X: First Transformers Results
Introduction
At Advancing AI 2026, AMD introduced the AMD Instinct MI455X, its newest datacenter GPU for AI. The headline specification is hard to miss: 432 GB of high-bandwidth memory on a single GPU.
For Hugging Face, the MI455 marks the next step in our close collaboration with AMD, guided by a simple goal: make the Hugging Face ecosystem work across AMD hardware and take full advantage of each platform. Transformers already runs daily CI on MI300 infrastructure, and this year we’re expanding coverage to Ryzen and Radeon systems. At the kernel level, we’re also working with the AITER team to bring optimized AMD kernels to the Hugging Face Hub, where they can be used directly in Transformers workflows.
We’re excited to start testing MI455! We got early access and used that time to test the Transformers library. Here are our preliminary findings.
A First Look at the MI455
The most immediate change from the previous generation is memory capacity. Each card comes with 432 GB of ultra-fast HBM4 and up to 23.3 TB/s of memory bandwidth. To put that in perspective, the previous MI300 generation had 192 GB per card. The four-accelerator system we tested therefore exposes about 1.7 TB of high-bandwidth memory in one machine.
It means larger models can stay on fewer devices, long-context workloads have more room for their key-value cache, and serving systems can keep more requests active before memory becomes the limiting factor.
To make that concrete, we ran a preliminary capacity study with a fixed model and context configuration. We increased the number of concurrent requests on a single accelerator until the workload ran out of memory. The 192 GB MI300 reached its memory limit first, while the 432 GB MI455X supported over three times more concurrent requests before reaching its own limit, thanks to its roughly 3× larger KV cache capacity after loading the 64 GB Qwen3-32B BF16 model.
We will follow-up with speed comparison as we progress with our tests.
MI455 on Transformers
When we bring up a new accelerator in Transformers, our main signal is the test suite for a curated set of 24 important model architectures. That list spans the major patterns and modalities in the library like encoders, decoders, vision models, audio models, multimodal systems, and modern large language models. It provides a strong proxy for the health of the broader ecosystem.
After our fixes, the MI455 success rate across these 24 models is approximately 99.5%, depending on a handful of flaky tests. Relative to the current CI infrastructure, this is in the same range as the 99.4% measured on MI300 and the 99.1% measured on NVIDIA A10 in daily CI.
Reaching this level required work across the stack. Together with AMD, we enabled a stable Flash Attention path, added torchcodec support for audio and video in multimodal models, and resolved output-comparison issues uncovered during testing.
What comes next
Next, we plan to add MI455 to the Transformers CI, continue bringing optimized AITER kernels to the Hugging Face Kernel Hub, and expand MI455 validation across more Hugging Face libraries. Thanks again to the AMD team for giving us early access and the close collaboration needed to have those results on day one!
