Nvidia details 88-core Vera CPU with six-chiplet Olympus design
TL;DR
- Nvidia's Vera server CPU packs 88 new Olympus cores across six dies on one interposer, fed by eight 128-bit LPDDR5X memory controllers.
- Nvidia claims up to 30x total throughput vs Grace Blackwell in select scenarios, with 1.8x on agentic workloads and 1.5x on data processing.
- Vera talks to Nvidia's next-generation GPU over NVLink-C2C, exposes 96 lanes of PCIe and CXL, and prioritizes single-thread performance for agentic workloads.
Nvidia used its Hot Chips 2026 session on August 24 to disclose the internals of Vera, its next-generation server CPU: 88 cores of a new Olympus architecture split across six dies on a single interposer, fed by eight 128-bit LPDDR5X memory controllers.
The chip pairs with Nvidia's next-generation GPU over NVLink-C2C and exposes 96 lanes of PCIe and CXL. The company's headline claim is "up to 30 times the total throughput of Grace Blackwell." ServeTheHome's more grounded benchmark line reads: "Agentic workloads are close to 1.8x, while data processing workloads are closer to 1.5x." Those figures come from Nvidia's own workloads, not third-party production traces.
The design bet is on single-thread performance rather than raw core count. Olympus runs a 10-wide decode front-end and uses what Nvidia calls "statically partitioned spatial multithreading" as an alternative to conventional SMT. Agentic AI pipelines are the stated target.
Ryan Smith wrote the ServeTheHome breakdown live from the session. Process node, socket TDP, and shipping dates were not disclosed. Vera arrives the same week our chips tracker logged IBM's mainframe chip running Z and Arm workloads on the same core and Waymo's custom 5nm chip for its 13-camera Ojai robotaxi, a reminder that Arm-based server silicon is filling out fast.
Originally reported by servethehome.com
Read the original article →Original headline: Nvidia Details 88-Core Vera CPU at Hot Chips With Six-Chiplet Design and 30x Grace Blackwell Throughput