AMD launches Instinct MI455X GPUs and Helios rack-scale AI platform

AMD unveiled its next-generation AI accelerator, the Instinct MI455X GPU, alongside the Helios rack-scale system built around it, positioning the pairing as a direct competitor to NVIDIA's Vera Rubin platform for large-scale AI inference and training.

What's new

The Instinct MI455X is built on 5th Gen AMD CDNA architecture and anchors AMD Helios, a fully co-designed rack that pairs 72 GPUs with 6th Gen AMD EPYC "Venice" server CPUs and AMD Pensando networking over a UALoE fabric. AMD says the rack delivers 2.9 exaflops of dense FP4 compute, 1.4 exaflops of FP8 compute, 31 TB of HBM4 memory, 1.7 PB/s of aggregate HBM bandwidth, 260 TB/s of scale-up bandwidth, and 43 TB/s of scale-out bandwidth, connecting all 72 GPUs in a single scale-up domain.

On performance, AMD says the MI455X delivers up to 34x higher token throughput and up to 18x lower token cost than its predecessor, the MI355X, when serving DeepSeek-V4-Flash. Measured against NVIDIA's Vera Rubin NVL72 rack, AMD says Helios is designed to deliver up to 15% more AI compute, 50% more HBM capacity, and 50% more scale-out bandwidth. On the Kimi K2 Thinking model with a 32K input and 8K output sequence, AMD's modeled throughput per GPU is up to 15% higher at low interactivity, 12% higher at medium interactivity, and 10% higher at high interactivity than modeled Vera Rubin NVL72 performance.

AMD ROCm software ties the rack to production workloads, with native support for PyTorch, TensorFlow, and JAX, plus tooling for deployment, observability, and lifecycle management. The announcement, made at AMD's Advancing AI 2026 event, also included the more cost-optimized Instinct MI350P GPU and a new Kria AI robotics platform built on Ryzen AI Embedded processors.

Context

The launch follows AMD's MI400-series roadmap disclosed earlier in 2026 and comes as AMD works to convert marquee compute commitments — including its multi-gigawatt GPU supply deal with OpenAI and a separate strategic partnership with Anthropic to deploy up to 2 gigawatts of Instinct MI450-series GPUs — into shipping, benchmarked hardware. AMD says Helios systems will reach customers through OEM partners including Bull, HPE, Lenovo, and Supermicro, with OpenAI expecting to begin Helios deployment in Q4 2026 and accelerating through 2027. AMD has also flagged MI500-series GPUs for 2027 and MI600-series for 2028, signaling an annual cadence intended to keep pace with NVIDIA's own yearly architecture updates.

Why it matters

AMD's pitch has shifted from selling individual accelerator chips to selling a fully co-designed rack — compute, memory, networking, and software bundled and benchmarked as one system, the same unit of sale NVIDIA has pushed with Vera Rubin NVL72. That framing matters because AI infrastructure buyers increasingly evaluate cost and performance at the rack or data-center level rather than per chip, and AMD's head-to-head comparisons against NVIDIA's newest platform are a direct bid to be treated as an equivalent, not just a cheaper alternative. Whether AMD can hold that positioning depends on whether Helios ships in volume on the timeline AMD described and whether the performance claims hold up against independent benchmarks once customers like OpenAI put the racks into production later this year.