5th-Gen Xeon Goes to 64 Cores

Intel’s fifth-generation Xeon family, codenamed Emerald Rapids, refreshes last year’s Xeon, raising the number of cores in the flagship model from 60 to 64 while maintaining socket compatibility. Little has changed in the microarchitecture, but performance receives an extra bump from additional L3 cache and support for higher DRAM bandwidth. Net performance improvements are modest.

The company launched 32 new models with unchanged or higher clock speeds and, with a few exceptions, equal or lower pricing. The number of models is reduced from the fourth generation; the silicon process remains unchanged on Intel 7. According to our analysis, performance per watt and per dollar also rose by an average of 9% and 10%, respectively. The shared L3 cache (or last-level cache) grew by up to 3×, although that multiplier varies by model; DRAM bandwidth grew by an amount that also varies by model, with some remaining unchanged. On certain workloads and models, these changes are likely to yield more performance impact than the clock speeds.

The Xeon brand describes Intel processors intended for data-center use. As such, they contain CPUs and various hardware accelerators. Since serious graphics and artificial intelligence (AI) workloads typically run on dedicated add-on cards, Xeon doesn’t include an integrated GPU or neural processing unit (NPU); AVX and AMX extensions handle inference workloads in the CPU.