AMD has celebrated the event ‘advancing with AI’ highlighting the launch of its new accelerators Instinct Mi350. These GPUS are based on CDNA 4 architecture and will be manufactured by TSMC in their 3 Nm process node. Its preferential destination will be HPC high performance systems for artificial intelligence where they will compete with NVIDIA solutions.
The new AMD AI accelerators have 185,000 million transistors and will be marketed in two versions, Instinct MI350X and higher performance, Instinct MI355X. They will be available in liquid and air cooling configurations, are compatible with the latest types of FP6 and FP4 AI data and are equipped with an enormous HBM3E memory capacity.
AMD has also presented Rocm 7, a new review of its well -known open source platform for AI, HPC and heterogeneous computing.
AMD ensures that the series offers 20 pflop computing FP4/FP6, which is a Four -time performance improvement Between generations. As for HBM3E memory, faster data transfer speeds are obtained with stratospheric capacity in both versions of 288 Gbytes. The chips are also equipped with UBB8, a new standard for implementation of AI accelerating solutions, which allows a faster deployment in air or liquid refrigerated nodes.
Instinct Mi350 performance
AMD shared metrics say that these solutions offer 8 TB/s of aggregate memory bandwidth, 79 FP64 TFLOPS, 5 PFLOPS of FP16, 10 PFLOPS of FP8 and 20 PFLOPS of computation FP6/FP4. These figures correspond to the top of the range of range, Instinct MI355X, of 1,400 watts. It should be noted that both versions use the same matrix, but 355x has a higher TDP consumption.
Compared to the previous series MI300, AMD ensures that they offer an increase in up to 35 times In the inference yield with flame 3.1 405b (performance), which is a huge increase. The full ecosystem platform will offer up to 8 GPU of the MI355 series with 2.3 TB of HBM3E memory, 64 TB/s total bandwidth, 0.63 PFLOP of FP64, 81 PFLOP of FP8 and 161 PFLOP of Computer Performance FP6/FP4.
A complete rack With liquid cooling, it will house between 96 and 128 GPUS of the Instinct Mi350 series with up to 36 TB of HBM3E memory, 2.6 Computer Ex -Flops FP4 and 1.3 Computer Ex -Flops FP8. The Rack will use CPUS TURIN EPYC of the company based on the Zen 5 core architecture together with the 400 Pollara interconnection solution.
Compared to their main competition, the Blackwell GB200 of NVIDIA, AMD data show an increase in yield between 20 and 30% in flame and can generate up to 40% more tokens. Besides, AMD solutions are significantly cheaper than Nvidia which will contribute to AMD’s objectives in terms of Price/benefit.
AMD says that the MI350 instinct are already available for their partners since this week for an expected availability from the third quarter of 2025. The company has also presented the next generation, Mi400, already in development and launch planned in 2026.
Sunnyvale’s company has taken the opportunity to give some details about the Instinct MI400, which will be integrated into the Rack Amd Helios platform, and has confirmed that this new generation will be a real beast, since A Rack Helios will have a bandwidth of up to 260 TB/s, a power in FP4 and FP8 of 2.9 efflops 7 1.4 efflops (Exaflops)and will have a total 31 TB memory capacity with a bandwidth of 1.4 pb/s.
Its launch is expected for 2026and in 2027 we will see the launch of a new generation rack with the CPUS AMD EPYC Summer next generation, which will be based on a new architecture. This rack will be equipped with the new MI500 instinct accelerators and with an NIC solution (Network Interface Card) Vulcano of thinking.
