News stories tagged with #FP8
NVIDIA Unveils Vera-CPU Rack at GTC 2026: New Benchmark for CPU-Only AI Infrastructure
At GTC 2026, NVIDIA unveiled the Vera-CPU Rack, a system designed for CPU-only inference featuring 256 Vera CPUs based on a custom Arm v9.2-A core called Olympus. Each CPU includes 88 cores with SMT, FP8 support, and up to 1.5 TB of RAM. The rack provides 400 TB of LPDDR memory and a total memory bandwidth of 300 TB/s, with individual chips achieving 1.2 TB/s. Optimized for agentic workloads, reinforcement learning, and AI training, the solution is being co-developed with partners like HPE, with the Cray GXC240 supporting up to 640 Vera CPUs per rack.