AMD launched its latest Instinct AI GPUs, the MI350X and MI355X. The MI350 series follows the MI325 series, which helped drive AMD to a nearly 40% share of the server CPU segment in the first half of this year. The chipmaker touted the MI350’s performance gains while also highlighting its expanding “Open AI ecosystem” with ROCm7 and Helios, its AI rack. In the fierce competition with Intel for AI data center dominance, AMD projected strength.
“AMD is driving AI innovation at an unprecedented pace, highlighted by the launch of our AMD Instinct MI350 series accelerators, advances in our next generation AMD ‘Helios’ rack-scale solutions, and growing momentum for our ROCm open software stack,” said Dr. Lisa Su, AMD chair and CEO, in a statement. “We are entering the next phase of AI, driven by open standards, shared innovation, and AMD’s expanding leadership across a broad ecosystem of hardware and software partners who are collaborating to define the future of AI.”
Credit: AMD
The Instinct MI350 series is a big improvement over the MI325 series in transistor counts and memory bandwidth. That translates to significantly better performance. According to AMD, the MI350 series represents a 4x AI compute increase. The chipmaker also reports a 35x inferencing increase.
The new MI350 series is built on the CDNA 4 architecture with TSMC’s newest 3nm node. It has 185 billion transistors. As Wccftech points out, the two versions of the GPU have different power ratings despite having the same die. The MI350X has a TBP of up to 1,000W, while the MI355X has a whopping 1,400W TBP. AMD offers fan and liquid cooling support.
The Instinct MI350X and MI355X are similar regarding high-bandwidth memory (HBM). They have 288GB of HBM3E memory with 8TBps bandwidth, compared with the 256GB of HBM3E memory and 6TBps bandwidth on older MI325X GPUs.
AMD also released floating-point precision scores. Individually, the MI350X ranges from 72 TFLOPS in FP64 to 18.45 PFLOPS in FP4 with structured sparsity. The MI355X scored 78.6 TFLOPS in FP64 tests and 147.6 PFLOPS in FP4.
AMD Helios AI Rack.
Credit: AMD
The processors typically appear in sets of eight in the Instinct MI350 series platforms. AMD also provided platform performance, with the MI350X platform delivering 577 TFLOPS in FP64 and the MI355X scoring 628.8 TFLOPS. For FP4 performance, the MI350X provided 147.6 PFLOPS, while the MI355X platform provided 161 PFLOPS.
AMD also discussed its Helios AI rack infrastructure, which will arrive sometime in 2026. It will support the MI350 series platforms, as well as the MI400 series, which AMD teased. The MI400, which is also landing in 2026, is expected to have 432GB of HBM4 memory and 19.6TBps memory bandwidth. AMD hasn’t elaborated on pricing just yet.