AMD Instinct MI350X
Air-cooled AI accelerator using CDNA4 architecture and 3nm process.
The AMD Instinct MI350X is an artificial intelligence accelerator designed for data centres. It uses the CDNA4 microarchitecture and a 3 nanometre manufacturing process. The chip features 288 gigabytes of HBM3E memory and supports low-precision formats like FP6. It serves as a competitor to NVIDIA Blackwell platforms in training and inference tasks.
- Uses 3nm N3P process for compute dies
- Features 288GB of HBM3E memory
- Supports FP6 at same rate as FP4
- Air-cooled UBB8 form factor
- 32 enabled compute units per die
What it is
The AMD Instinct MI350X is a graphics processing unit built for artificial intelligence workloads. It belongs to the MI350 series, which AMD launched to compete with NVIDIA B200 generation hardware. The chip relies on the CDNA4 architecture and is manufactured using a 3 nanometre process node. It includes 288 gigabytes of high bandwidth memory. This model is air-cooled, distinguishing it from the liquid-cooled MI355X variant. The company positions it for both AI training and inference in data centre environments.
How it came about
AMD introduced the MI350 series at its Advancing AI 2025 event. Alan Smith, the chief architect, explained that the design evolved from previous Graphics Core Next architectures. The team focused on optimising distributed computing algorithms for high performance. Unlike the prior MI300 series, which balanced high performance computing and AI, the MI350 targets AI specifically. This shift allowed changes to the compute unit layout. The launch followed the MI300 series, which saw adoption in supercomputers like El Capitan and cloud platforms such as Microsoft Azure.
Inside the chip
The chip uses a chiplet design with compute dies on the N3P process and an input output die on N6. Each compute die has 36 physical compute units, but only 32 are enabled. This power of two configuration helps tile tensors efficiently. The local data store capacity increased to 160 kilobytes, doubling the bandwidth to the register file. The matrix engine supports FP4, FP6, and FP8 data types. FP6 runs at the same throughput as FP4. The design removes hardware acceleration for TF32, relying on BF16 or software emulation instead.
How it was made
TSMC manufactures the compute complexes using its 3 nanometre N3P technology. The input output components use the older N6 process. This mix allows flexibility in the chiplet architecture. The physical layout includes spare compute units for harvesting during production. This strategy improves yields for high volume manufacturing. The final package fits the UBB8 form factor. This standard enables compatibility with eight GPU systems similar to NVIDIA HGX platforms.
Where it went
Data centres use this accelerator for large language model training and inference. AMD claims significant performance gains over the previous MI300X for models like Llama 3 70B. The chip supports rack scale infrastructure with up to 128 GPUs per rack. It connects via UltraEthernet Consortium and UALink standards. These systems provide up to 36 terabytes of memory per rack. The air cooled design suits lower density rack deployments compared to liquid cooled alternatives.
Versions and family
The MI350 series includes two main models: the MI350X and the MI355X. The MI350X is air cooled, while the MI355X uses liquid cooling and runs at 1400 watts. The MI355X offers higher performance. Both share the same CDNA4 architecture and memory specifications. The MI350 series replaces the MI300 line. The MI325X served as an intermediate update to the MI300X, similar to how NVIDIA updated its H100 to the H200.
Why it matters
AMD aims to challenge NVIDIA dominance in the AI accelerator market. The company emphasises value based sales and lower cost per token. By matching specifications of NVIDIA Blackwell parts, AMD seeks to offer a competitive alternative. The focus on FP6 performance aims to lead in both inference and potential training applications. This release continues AMD annual cadence for AI focused GPU products.
Questions people ask about AMD Instinct MI350X
What is the difference between MI350X and MI355X?
The MI350X is air cooled, while the MI355X is liquid cooled and runs at 1400 watts. The MI355X is the faster of the two models, though both share the same underlying architecture and memory specs.
Why does the chip have fewer enabled cores than physical ones?
Each die has 36 physical compute units, but only 32 are enabled. This power of two number helps tile tensors efficiently. The extra units allow for harvesting during manufacturing to improve yields.
Does it support TF32 data format?
Hardware acceleration for TF32 was removed. The chip uses BF16 for higher throughput instead. Users can emulate TF32 via software using BF16 or cast it to FP32, but there is no dedicated hardware path.
Good reading and sources
There is no English Wikipedia article on AMD Instinct MI350X yet, so the facts on this page come from the sources listed above: servethehome.com/amd-instinct-mi350-launch-event-coverage/, extremetech.com/computing/amd-launches-instinct-mi350x-and-mi355x-ai-g, chipsandcheese.com/p/amds-freshly-baked-mi350-an-interview. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.