Intel Announces Xeon Max Server Processors and Data Center GPUs – Computers – News

Intel announced Xeon Max, server processors known internally as Sapphire Rapids HBM. In addition, the company announces the Data Center GPU Max series, code-named Ponte Vecchio GPU server. Deliveries of both products will begin in January.

Of Xeon Max-cpu-seriesformerly known as Sapphire Rapids HBM, it has up to 56 performancecores, which are spread over four chiplet-like tiles. Processors don’t get efficient E-cores. Those tiles are interconnected with what Intel integrated multi-die interconnect bridge called, or emib. It is a kind of small silicon bridges that act as an interconnect. This CPU series has TDP up to 350W.

The CPU also gets 64GB of HBM2e memory spread across four 16GB stacks. This gives each core more than 1GB of HBM2e memory at its disposal. Xeon Max also supports PCIe 5.0 and the CXL 1.1 interconnect standard for I / O. Intel says the new CPU uses 68% less power than an AMD Milan-X cluster for the same performance, or is 2.4 times faster when building climate models than such an AMD cluster when using the HBM memory.

The Data Center GPU Max series, formerly known as Ponte Vecchio, consists of three different boards that can be equipped with up to 128 Xe-HPC cores. The cards are also made up of several floor tiles and get up to 408MB of L2 cache and up to 64MB of L1 cache. Data Center GPU cards also get native ray-tracing acceleration, which Intel believes can speed up scientific visualizations and animations. In some financial benchmarks, the cards would be 2.4 times faster than Nvidia’s A100 cards; in some physical simulations, Intel cards would be 1.5 times faster.

Intel’s most powerful data center GPU is the Max Series 1550, a 600W card with 128 Xe cores, 128GB of HBM2e memory, and 128 ray tracing units. The Max 1350 series has a TDP of 450W, 112 Xe cores, and 96GB of memory. The 1100 GPU in turn has a TDP of 300W, 56 Xe cores, and 48GB of memory. The boards can be linked together using Intel Xe Link bridges.

The Aurora supercomputer, among others, must be equipped with Xeon Max products. It should be the first supercomputer capable of achieving more than two exaflops of computing power. Incidentally, this supercomputer has been waiting for Intel’s server products for years; Aurora actually did to be delivered in 2021. This has become last year it was postponed to 2022 due to the delay of the products announced today.

So Aurora won’t hit the two exaflops until next year, with the release of the Xeon Max and Data Center GPU Max products. The supercomputer will have more than 10,000 blades. Each blade contains six Max Series GPUs and two Xeon Max CPUs. Several US labs and Kyoto University are also getting supercomputers with the products announced today. The successor to the Max Series GPU is expected to appear in 2024. Those GPUs now have the code name Ponte di Rialto.