NVIDIA Vera CPU beats Intel and AMD in early benchmarks

NVIDIA's new Vera CPU, a custom-designed processor for data centers running advanced AI workloads, has demonstrated highly competitive performance in initial benchmarks, according to a recent independent test. The results indicate the 88-core chip, which uses NVIDIA's in-house "Olympus" core architecture, performs strongly against current server processors from Intel and AMD.

The Vera CPU is built around 88 custom Olympus cores compatible with the Armv9.2 instruction set. A key feature is its second-generation LPDDR5X memory subsystem, which delivers up to 1.2 terabytes per second of bandwidth while consuming less than 30 watts. In memory performance tests, the chip sustained approximately 90% of its peak bandwidth, a record in that specific benchmark, and delivered over four times the memory bandwidth per core compared to traditional x86 server CPUs.

Performance testing, conducted at NVIDIA's headquarters, showed the Vera CPU achieved a geometric mean performance increase of 63% over its predecessor, the Grace CPU. In the same aggregate measure, it outperformed a 64-core AMD EPYC 9575F processor by about 10% and a 128-core Intel Xeon 6980P by approximately 55%. The chip also...