A rack of 256 Vera CPUs
Pegatron said its NVIDIA Vera CPU Rack is built on the NVIDIA MGX modular reference architecture and holds 256 liquid-cooled Vera CPUs, more than 22,500 cores and more than 45,000 threads, with up to 300 TB/s of aggregate LPDDR5X memory bandwidth. The company said the rack supports 22,500 sandboxes and delivers 80 percent faster agentic CPU performance, aimed at agent execution, reinforcement learning evaluations and data-processing pipelines. NVIDIA Spectrum-X Ethernet and NVIDIA Quantum-X800 InfiniBand handle the networking.
Rubin racks and a 2U system
The RA4803-72N3 is a fully liquid-cooled NVIDIA Vera Rubin NVL72 rack with 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs, which the company lists for frontier-scale training, reinforcement learning and large-scale agentic inference. The AS210-2T1-8H3 is a 2U fully liquid-cooled NVIDIA HGX Rubin NVL8 system with an eight-GPU Rubin baseboard, a choice of Intel Xeon 6 or AMD EPYC 9006 processors and rack-scale configurations of up to 72 GPUs. The release also names NVIDIA Groq 3 LPX as an extension of the Vera Rubin platform for latency-sensitive inference.
AMD servers and storage
Pegatron will introduce the MS310-2A2E, a 2OU two-node server on AMD EPYC 9006 Series processors, powered through a 48VDC busbar clip and aligned with DC-MHS, OCP 3.0 NIC and DC-SCM 2.1. It will also show the GS203-2A2 and GS205-2A2E dual-socket servers on the same AMD processors, and the SS201-1A1 storage platform, which supports up to 32 E3.S NVMe SSDs.
At the show
The systems are on display at Booth #5416 at SC26, which runs November 16 to 19 at McCormick Place in Chicago. Gary Cheng, president and co-CEO, said GPUs remain essential for AI acceleration while CPUs are taking a more critical role in agent execution, reinforcement learning and data processing.