Intel announced the release of new server processors of the Xeon 6 series (Granite Rapids), which have exclusively P-cores. The company also announced the release of a specialized AI accelerator, Gaudi 3.
Image source: Intel
Granite Rapids are manufactured using Intel 3 (5 nm) process technology. The series includes five models with a number of cores from 72 to 128, a base frequency of 2.0 to 2.7 GHz and a maximum frequency of 3.9 GHz (on one core), as well as from 3.2 to 3.7 GHz on all kernels. The processors received from 432 to 504 MB of L3 cache memory and have a stated TDP of 400 to 500 W.
The chips support both single- and dual-processor builds, have support for 12-channel DDR5-6400 and MRDIMM-8800 RAM, and are also equipped with support for 96 PCIe lanes.
Intel says the Xeon 6 differs from its predecessors with an increased number of cores, double the memory bandwidth and AI acceleration capabilities built into each core. These chips are designed to meet AI performance demands from edge systems to data centers and cloud environments.
Intel says the new Xeon 6 is more than twice as fast as Epyc Genoa processors (up to 96 Zen 4 cores) across a wide range of compute workloads and more than five times faster at neural network workloads.
In turn, specialized AI accelerators Gaudi 3 are specially optimized for working with generative models. They use 64 tensor processing units (TPCs) and eight matrix multiplication engines (MMEs) to accelerate deep neural network computations. Gaudi 3 accelerators also received 128 GB of on-board HBM2 memory and support up to 24 200 Gbit Ethernet ports for scalable networks. Gaudi 3 claims seamless compatibility with the PyTorch framework and advanced Hugging Face transform and diffusion models.
Intel says the new Gaudi 3 AI accelerators deliver up to 20 percent more throughput and 2x better price-performance than the H100 to power the LLaMa 2 70B model.