AMD CEO Lisa Su officially announced at the AMD Advancing AI annual conference that the first Helios rack-scale AI solution has entered full-scale mass production. The roster of Instinct partner ecosystem displayed on the event venue backdrop includes multiple Taiwan publicly listed companies, with TSMC accounting for 14 of them.
Helios AI rack specs: 15% more performance than NVIDIA NVL72
According to an AMD official blog post, the key technical specifications of the Helios rack system are as follows: based on the AMD Instinct MI455X GPU, with 72 GPUs integrated into a single rack. Using the UALoE scalable architecture, it provides 260TB/s of vertical scaling bandwidth, paired with AMD EPYC “Venice” processors and AMD Pensando networking chips. The full system delivers 2.9 Exaflops FP4 compute performance and 31TB of HBM4 memory capacity.
AMD vs. NVIDIA Vera Rubin NVL72 comparison data: Helios offers 15% more AI compute performance
AMD claims that Helios offers 15% more AI compute performance, 50% more HBM memory capacity, and 50% more horizontal scaling bandwidth than NVIDIA’s Vera Rubin NVL72. In DeepSeek-V4-Flash model tests, under high-interactivity scenarios, the MI455X token throughput can reach up to 34 times that of the previous-generation MI355X, while the cost per million tokens can be reduced by up to 18 times.
14 Taiwan supply-chain companies listed at the AMD Advancing AI conference
Based on reports from Ettoday and the Central News Agency, the Taiwan publicly listed company partner roster revealed on the AMD venue backdrop is as follows:
TSMC (2330): responsible for AMD MI455X GPUs’ 2-nanometer and 3-nanometer advanced process technologies
Gigabyte (2376), ASUS (2357), ASRock (3515, parent company: Pegatron): server/motherboard suppliers
Wiwynn (6669), Quanta (2382, parent company: Quanta Group), Wistron (3231): ODM server design and manufacturing
Foxconn (2317), Compal Electronics (2356), ASUS?—note: 仁寶 (2324), and Wistron?—sorry—note: 和碩 (4938): EMS/ODM manufacturers
MSI (2377), Inventec (3706), Advantech (2395): server and industrial PC suppliers
According to a report by The工商時報, infrastructure cooperation partners also include Sanmina, a U.S. EMS giant (新美亞).
Lisa Su: AI accelerator market to reach $1.4 trillion by 2030
Based on Lisa Su’s public remarks during the meeting, she estimates that the AI accelerator market size will reach about $1.4 trillion by 2030. The driver is the stage-by-stage jump in compute demand brought by agentic AI—when users ask AI agents to complete tasks, the system needs dozens of reasoning steps, tool calls, and data-access steps, consuming large amounts of GPU compute power.
So far, Helios has been adopted by major cloud providers including OpenAI, Meta, Microsoft, and Oracle. The software layer is connected through the open ROCm platform, supporting mainstream frameworks such as PyTorch, TensorFlow, and JAX.
FAQ
What are the core specs and performance of the AMD Helios rack?
According to AMD’s official blog, a single Helios rack integrates 72 MI455X GPUs, delivering 2.9 Exaflops FP4 compute performance and 31TB of HBM4 memory. AMD data shows that compared with NVIDIA’s Vera Rubin NVL72, it offers 15% more AI compute performance, 50% more HBM memory capacity, and 50% more horizontal scaling bandwidth.
Which 14 Taiwan supply-chain vendors has AMD Helios partnered with?
According to reports from Ettoday and the Central News Agency, the 14 Taiwan publicly listed companies include TSMC (2330), Gigabyte (2376), ASUS (2357), ASRock (3515), Wiwynn (6669), Quanta (2382), Wistron (3231), Foxconn (2317), Compal Electronics (2356), Inventec (3706), Compal?—sorry—note: 仁寶 (2324), and Pegatron (4938), MSI (2377), and Advantech (2395).
What is Lisa Su’s forecast for the AI accelerator market?
Based on Lisa Su’s public statements at the AMD Advancing AI conference, she estimates that the AI accelerator market size will reach about $1.4 trillion by 2030, driven mainly by the explosive compute demand of agentic AI.
