AMD unveils largest data-center accelerator to challenge Nvidia

AMD unveiled its largest data-center accelerator to date, expanding its Instinct family. The device pairs AMD’s GPU compute with increased on-package memory and a chiplet-based design aimed at hyperscalers, cloud providers and large enterprises.

AMD described the product as built for large-scale generative AI models and other machine-learning workloads. The architecture combines high-performance GPU compute with substantial on-package memory to reduce data movement and improve throughput for memory-intensive tasks. AMD targeted both training and inference use cases.

The company stated the accelerator will integrate with AMD’s software stack to support common AI frameworks and provide tools for model development and deployment. AMD highlighted power efficiency and memory capacity as key differentiators and noted larger on-board memory can reduce the need for distributed-memory training and simplify model scaling.

OEMs and cloud providers will offer the product; AMD expects sample systems to appear in partner roadmaps and early trials are underway. The company said broader availability is planned later this year, with early shipments to select partners in the near term. Pricing was not disclosed.

The design uses a modular chiplet package that lets AMD scale compute and memory independently. The company noted this approach can improve manufacturing flexibility and cost efficiency. AMD pointed to software advances intended to ease migration and allow developers to run existing frameworks on the new hardware with minimal changes.

Industry observers will look at cost-per-performance and power-per-performance figures once independent benchmarks are published. The product targets Nvidia’s data-center accelerator market by emphasizing memory capacity and integration to reduce system complexity for very large models.

AMD has increased investment in data-center GPUs and high-performance computing over several years, shifting from discrete GPU designs to more integrated, chiplet-based approaches. The new accelerator is the company’s largest effort to date to expand its presence in the AI accelerator market.