AMD To Refresh Instinct MI300 Series With MI350 AI Accelerator Using 4nm Node This Year

AMD seems to be preparing a 4nm refresh of its MI300 AI Accelerators known as the MI350 which is planned for later this year.

AMD's current MI300 lineup consists of the AI-optimized MI300X & the compute-optimized MI300A accelerators but it looks like the company is planning to expand its portfolio. We recently saw the emergence of the MI388X which might be an export-compliant variant for China but AMD did state that it was prevented from shipments. The MI388X was likely going to be another CDNA 3 offering utilizing a 5nm and 6nm process technology but it looks like AMD has a proper refresh planned for its Instinct family for later this year.

According to a report from TrendForce, it looks like AMD might be launching a new part known as the Instinct MI350 which will utilize a refreshed CDNA 3 architecture utilizing TSMC's 4nm process node. While details on the Instinct MI350 are slim, it was recently teased by AMD itself that they'll be offering higher HBM3E capacities in future refreshes of the Instinct MI300 series. So higher HBM capacities coupled with a fine-tuned architecture on the 4nm node can lead to some decent gains.

Furthermore, TrendForce notes that the extension of export controls now includes not only the previously restricted AI chips from NVIDIA and AMD, such as the NVIDIA A100/H100, AMD MI250/300 series, NVIDIA A800, H800, L40, L40S, and RTX4090, but also their next-generation successors like NVIDIA's H200, B100, B200, GB200, and AMD's MI350 series. In response, HPC manufacturers have quickly developed products that comply with the new TPP and PD standards, such as NVIDIA's adjusted H20/L20/L2, which remain eligible for export.

TrendForce

Videocardz was also able to spot a listing from AMD Singapore which confirms the Instinct MI350 accelerator lineup. The product has already been submitted for silicon readiness & optimizations.

It should be remembered that AMD will be competing against both NVIDIA & Intel in the AI space. The Blackwell B100 GPUs are in production and B100/B200 will be rolling out to customers soon. Meanwhile, Intel also announced its Gaudi 3 accelerators which offer up to 50% faster AI compute versus the NVIDIA H100 GPUs. So the space is heating up. In recent MLPerf benchmarks, NVIDIA & Intel were the only ones to submit their AI performance benchmarks meanwhile AMD missed the spotlight as it didn't submit any numbers.

TrendForce has also shared the full list of products that are affected by the latest version of the US export controls against China. These include several current and upcoming GPUs including AMD's Instinct MI388X & MI350 series.

AMD has also confirmed its next-gen MI400 AI accelerator which should be released in 2025 and feature a more capable architecture that is tuned for the AI-era. AMD is also working on its ROCm software suite and has made certain blocks open source to fine-tune its performance for AI work-loads.