AMD Unveils MI455X GPU, Venice CPU, and Helios Rack; Forecasts $1.4 Trillion AI Accelerator Market by 2030
智东西 · wechat · 2026-07-24
At AMD Advancing AI, CEO Lisa Su announced multiple products:
- MI455X GPU: World's first 2nm GPU, 320B transistors, 432GB HBM4, 40 PFLOPS FP4, 34x throughput vs MI355X, 1/18 token cost.
- Venice CPU: First 2nm server CPU, 512 threads, PCIe 6.0, 2.2x performance vs Nvidia Vera.
- Helios rack: 72 MI455X + 18 Venice CPUs, 30% more tokens per dollar vs Nvidia rack.
- MI350P PCIe GPU: Air-cooled, 144GB HBM3E, 4.2x tokens per dollar vs Nvidia RTX Pro 4000.
- Roadmap: MI500 series in 2026, MI600 in 2028; Helios 500/600 in 2026/2028.
- Market forecast: AI accelerator market to reach $1.4T by 2030, GPU dominant.
- Partnership: With Cerebras for decoupled AI inference, 5x tokens per watt per second.
- China ecosystem: ByteDance, Tencent, Alibaba, Moonshot AI, MiniMax, DeepSeek featured.
More from Embodied
- U.S. Moves to Rebuild Domestic Robotics Supply Chain as Strategic Battleground — Rewkang · 2026-07-24
- South Park Commons and Viam Robotics to Host Robotics Hack Day — ditzikow · 2026-07-24
- Neuralink says trial participants are learning to steer a wheelchair with their minds — shivon · 2026-07-24
- A $7 hand and a 3D-printed gripper that can swap everyday tools — shuyanzh36 · 2026-07-24
- Android NPU demo generates Stable Diffusion 2.1 images locally in 6.1 seconds — nntb · 2026-07-24
- RoboStrategy teases a 6,000-word memo on Standard Bots and robot-arm architecture — Rewkang · 2026-07-24