[REVISION HISTORY]

AMD expands AI hardware across data centre and edge

Updated 13 times since CLSTR started tracking revisions of this situation.

What changed

2026-08-28 09:23 UTC → 2026-08-31 10:11 UTC · added ~~removed~~

AMD’s AI push has accelerated through new hardware, software, and partnerships. In early July, the company released the Ryzen AI Halo mini-PC, a $3,999 workstation built around the Ryzen AI Max+ 395 APU. AMD also unveiled its 2nm Zen 6-based EPYC “Venice” server line and the Helios rack-scale platform, which integrates 72 Instinct MI455X GPUs. Anthropic has committed up to $5 billion and 2 GW of MI455X GPUs for Claude model training, with the first gigawatt slated for early 2027. At the Advancing AI 2026 event, AMD expanded its Helios solution, claiming it delivers 30% more inference tokens per dollar than competitors. The company also introduced the Kria AI system-on-module and a Robotics Developer Platform using Ryzen AI Embedded X100 processors. To bolster inference, AMD entered a definitive agreement to acquire Toronto-based startup Taalas, which hard-wires AI model weights into silicon using a structured ASIC approach to reduce compute and memory bottlenecks. This technology is intended for integration across Instinct GPUs, Helios systems, and EPYC CPUs. Recent software updates include Day 0 support for the Qwen 3.8 27B model, enabling local execution on Ryzen AI Max+ 395 and Radeon AI PRO R9700 hardware. Developers can also use the Lemonade hardware-aware inference layer to optimize load distribution. Further infrastructure developments include a joint reference design with Schneider Electric for the Helios platform to assist in deploying high-density AI environments. ~~Additionally, the~~ The launch of ROCm 10 and the ROCm.AI toolkit introduced Hyperloom, ~~a tool~~ an autonomous agentic system designed to automate performance ~~tuning.~~ tuning and reduce optimization time from weeks to hours. AMD ~~reported~~ reports that the Instinct MI355X platform achieved a 3.3x increase in inference throughput for models like GLM-5 and ~~DeepSeek-R1 using these updates.~~ DeepSeek-R1, alongside a 2.4x gain in training efficiency for models such as DeepSeek-V3.

Versions

Only revisions since CLSTR began indexing content versions appear here. Select a version to see what changed compared to the one before it.