Intel launches Xeon 6 processors and Gaudi 3 AI accelerators

Intel has formally launched its next Xeon 6 server processors as well as the Gaudi 3 AI accelerators, making some pretty big boasts in the process.

The Xeon 6 6900P is the first in the Xeon family with performance cores (P-cores) designed for compute-intensive workloads and efficient cores (E-cores) for less compute-intensive work at a lower power draw. Intel is mixing and matching P-cores and E-cores up and down the Xeon 6 lineup depending on their target market. The Xeon 6 6700E with E-cores launched in June.

Intel claims Xeon 6 6900P delivers twice the performance of its predecessor thanks to an increased core count, double the memory bandwidth, and AI acceleration capabilities embedded in every core.

In addition to the P-cores, Intel has added AI inference capabilities to the Xeon 6900P lineup to bring AI coprocessors to CPUs. AMD is doing the same thing. The belief is that inferencing has a much lower power demand and therefore can be done on a client PC rather than require server processors like GPUs.

The stats for the Xeon 6900P chips are impressive. Compared to the previous Xeon generation, the maximum core count is doubled to 128 cores by using a chiplet design. So instead of one large piece of silicon, the processor is broken up into three easier-to-manufacture pieces.

The Xeon 6 is the first processor to support new MRDIMM modules from Micron, which improve bandwidth and latency performance. In the case of Xeon 6900P, memory improves by up to 57% to 8,800 MT/s.

The 6900P series supports six Ultra Path Interconnect 2.0 links for CPU-to-CPU transfer speeds of up to 24 GT/s, up to 96 lanes of PCIe 5.0 and CXL 2.0 connectivity, new vector extensions for high-performance computing, and a new matrix extension with 16-bit floating point the sweet spot in AI inferencing.

The higher core count comes with a price, however, and that is power consumption. The thermal design power (TDP) for four of the five processors in the 6900P family is 500 watts, while one processor consumes 400 watts. By contrast, the fifth-generation Xeon had a maximum TDP of 350 watts.

But there is definitely a performance gain. In the benchmark using a 7 billion parameter Llama 2 chatbot, Intel’s 96-core Xeon 6972P is more than three times faster than AMD’s 96-core EPYC 9654 and 128% faster than its previous generation Xeon. In a BERT language processing benchmark, the Xeon 6972P is 4.3 times faster than the AMD Epyc and 2.2 times faster than the previous generation Xeon.

However, it should be noted that the Epyc processor used in the benchmarks has been on the market for two years, and AMD has a new generation of processors due shortly.

Read more about Intel