Volume 343, Issue 2 · 8 articles from 2 issues
AMD's 'Building AMD Helios' video series shows how its Helios rackscale AI system is engineered, tested and validated.
AMD published a behind-the-scenes video series on engineering its Helios rackscale AI solution.
The first episode covers turning rack-scale architecture into a physical system, including mechanical and electrical design, liquid cooling, serviceability and system integration, while the second covers stress testing and validation ahead of customer delivery.
According to AMD, customers are also testing Helios at their own sites to validate workloads and prepare deployments at scale.
AMD's open, OCP-aligned Helion rack design targets agentic AI with CPU-only or mixed CPU/GPU configurations.
AMD unveiled Helion, an open rack-scale design for agentic AI workloads, which involve repeated model calls and multi-step tool use.
The design is aligned with Open Compute Project specifications, which AMD plans to publish, and offers a CPU-only hyperscale configuration with up to 32,768 cores on sixth-generation EPYC processors, plus an enterprise configuration that mixes EPYC nodes with Instinct MI350P GPU nodes.
It supports air and liquid cooling and includes a token router that directs agents' model requests across frontier and open models.
Cisco, Dell, HPE, Lenovo, Supermicro, QCT, Wiwynn and Foxconn are supporting the platform.
AMD's next Ryzen AI Halo developer platform offers up to 192GB unified memory for running 300B-parameter models locally.
AMD introduced the next Ryzen AI Halo Developer Platform, powered by the Ryzen AI Max+ PRO 495 processor with up to 16 Zen 5 CPU cores, 40 RDNA 3.5 GPU compute units, an NPU of up to 55 TOPS and up to 192GB of LPDDR5x unified memory.
AMD says it can run models of 300 billion parameters and larger locally.
The compact desktop system targets developers building agentic workflows that need CPU orchestration, accelerated compute and large memory.
AMD's FastFlow-SDK is a single platform for running, customizing and optimizing local AI down to kernels.
AMD introduced FastFlow-SDK, a platform for local AI software development on AMD hardware.
AMD says most tools only offer point-and-run inference through an OpenAI-compatible endpoint and block access when developers need to change how models execute or build custom operators.
FastFlow-SDK lets developers start by running a supported model and go further down to operators and kernels such as GEMM, attention and RoPE, within one ecosystem and without switching licenses, vendors or toolchains.
At the OCP 2026 keynote, Lisa Su showed AMD's open full-stack AI portfolio and the 6th Gen EPYC Verano CPU.
During the opening keynote of the 2026 OCP Global Summit in San Jose, AMD chair and CEO Lisa Su outlined how AMD's open, full-stack infrastructure supports agentic AI from enterprise systems to gigawatt-scale deployments.
She displayed the 6th Gen EPYC Verano data center CPU, the first EPYC with an LPDDR memory interface, due in 2027.
AMD emphasized working with ecosystem partners to champion open standards throughout the four-day event.
AMD's ROCm blog shows DP8+EP8 parallelism beats default tensor parallelism for long-context GLM-5.3 on MI300X.
This ROCm blog post examines serving GLM-5.3-FP8, a Mixture-of-Experts model with 744 billion total and about 39 billion active parameters and a context limit over one million tokens, on a single MI300X server.
It explains that default tensor parallelism splits every layer across eight GPUs and adds communication overhead that grows with sequence length.
The authors show a data- and expert-parallel DP8+EP8 layout sustains more throughput for long prompts, long outputs and high concurrency.
AMD's Helion reference rack for agentic AI pairs up to 32,768 Zen 6 cores with optional Instinct MI350P GPU blades.
At the OCP Summit in San Jose, AMD revealed Helion, an OCP-compliant reference rack for agentic AI that is distinct from its Helios training system.
A CPU-only hyperscale configuration provides dedicated capacity for agent execution and connects to separate inference infrastructure, while the enterprise version combines agent execution, GPU inference and supporting workloads in one rack.
It offers up to 32,768 Zen 6 cores from 256-core Venice EPYC chips, up to 16 Instinct MI350P GPUs in 5U eight-GPU blades and up to 102Tbps of networking, with blades in dense, balanced and 5GHz performance variants.
Partners will ship systems in November.
AMD introduced Helion, an open OCP-based rack blueprint for agentic AI with configurable CPU and GPU compute.
AMD introduced the AMD Helion agentic AI rack, an open blueprint for agentic AI infrastructure based on an Open Compute Project spec design.
Building on the Helios rack-scale approach, Helion lets customers tailor CPU core density, performance and power to agent workloads, with configurations ranging from CPU-only for hyperscale deployments to integrated CPU and GPU for enterprises.
Helion also includes an integrated token router that directs agents' model calls across frontier and open-source models.