On June 12, 2025, AMD held its Advancing AI event in San Jose, where CEO Lisa Su formally launched the Instinct MI350 series of accelerators. The presentation put its firepower behind two arguments: a large generational leap over the MI300 line, and a feature-by-feature comparison against Nvidia.
Just as telling was the staging. MI350 was the opening act — for the first time, AMD put rack-scale design at the center of the show, with the real target being the Helios rack and MI400 GPUs in 2026.
Specs and Availability
The MI350 family spans the air-cooled MI350X and the liquid-cooled MI355X, built on the CDNA 4 architecture with TSMC’s N3P process. The compute die pairs eight XCDs for 256 compute units, with 185 billion transistors, up 21 percent from the MI300’s 153 billion. The I/O die shrank from four tiles to two, doubling Infinity Fabric width to up to 5.5TB/s bi-sectionally. Each GPU carries up to 288GB of HBM3E with 8TB/s of bandwidth — 60 percent more capacity than Nvidia’s GB200/B200 — plus new FP4 and FP6 data types, with FP6 running at FP4 rates, a differentiator AMD highlighted repeatedly. The MI355X goes the liquid-cooling route at 1,400W, up from the MI300X’s 750W and the MI325X’s 1,000W. The trade-offs: FP64 matrix performance is halved versus the MI300X, and there is no APU variant like the MI300A — this generation is designed purely around AI inference.
On timing, the MI350 series is in production, began shipping in May, and partner servers and cloud instances are expected to arrive through the third quarter of 2025.
The Claims: 4x Compute, 35x Inference
AMD’s headline numbers: up to 4x AI compute and up to 35x faster inference versus the MI300X. The per-model figures are more modest — around 3x on DeepSeek R1, 3.3x on Llama 4 Maverick, up to 4.2x on AI agent and chatbot workloads, and 2.6x to 3.8x on content generation, summarization, and conversational AI. Against Nvidia, AMD claims the MI355X is up to 1.3x faster in like-for-like FP4 inference than the B200/GB200 and up to 1.13x faster in select training workloads; compared with the B200, the MI355X also doubles peak FP64 performance (72 TF on the MI350X, 79 TF on the MI355X).
Su set the tone in her keynote: “With the MI350 series, we’re delivering the largest generational performance leap in the history of Instinct, and we’re already deep in development of MI400 for 2026.”
ROCm 7 and the Enterprise Software Stack
The software side got its own launch: ROCm 7 delivers more than 4x inference and 3x training performance improvement over ROCm 6.0, with better support for industry-standard frameworks, expanded hardware compatibility, and new development tools, drivers, APIs, and libraries. AMD also debuted ROCm Enterprise AI, positioned as an MLOps platform for enterprise AI operations. For a company long criticized for lagging on software ecosystems, whether those two commitments land matters more than any spec sheet.
Platform Details: From 8 GPUs to 128
The system-level numbers matter just as much. A single 8-GPU platform carries 2.3TB of HBM3E; nodes pair two fifth-generation EPYC “Turin” CPUs with seven Infinity Fabric links totaling 1,075 GB/s, on PCIe 5.0 x16. At rack level, direct liquid-cooled configurations scale to 128 MI355X GPUs with 36TB of HBM3E in total, while air-cooled racks hold 64 GPUs with 18TB. The GPUs use OAM modules on OCP-standard UBB boards; scale-out runs over Pollara NICs that support Ultra Ethernet Consortium specifications, with Ultra Accelerator Link (UAL) handling scale-up. AMD’s argument is that the higher power buys density and better total cost of ownership.
Previewing 2026: Helios Racks and MI400
The real bet is next year. Helios is a double-wide rack-scale system built for frontier model training and large-scale inference, integrating MI400 GPUs with Zen 6-based “Venice” EPYC CPUs and Pensando “Vulcano” NICs. The MI400 preview: 20 petaflops of FP8 compute, 432GB of HBM4, 19.6TB/s of memory bandwidth, and 300Gbps of scale-out bandwidth per GPU. Andrew Dieckmann, AMD’s CVP and GM for data center solutions, said ahead of the event that Helios offers 50 percent more memory bandwidth and 50 percent more scale-out bandwidth than rival offerings. Su was blunter: the MI400 “is really designed from the ground up as a rack-level solution.”
For buyers, the practical question for late 2025 is concrete: can MI355X liquid-cooled racks — with their power and cooling demands — deliver better inference cost than Nvidia? AMD’s answer gets tested by real deployments in Q3.
Sources
- AMD announces MI350X and MI355X AI GPUs, claims up to 4X generational gain, 35X faster inference — Tom’s Hardware
- AMD launches Instinct MI350 GPUs, unveils double-wide Helios AI rack-scale system — Data Center Dynamics
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.