Today, Intel announced the world’s first x86 CPUs with HBM memory: the Intel Xeon CPU Max series. This is a product family we used to call “Sapphire Rapids” and it will consist of 56 performance cores (112 threads) with a power consumption of 350W TDP. It features an EMIB-based design and is divided into four groups. But the most interesting thing is that it also uses 64GB of HBM2e memory, divided into 4 16GB clusters, with a total of 1TB/s memory bandwidth, and each core has more than 1GB of HBM.
Coincidentally, this is also the CPU that will power Argonne National Laboratory’s Aurora supercomputer. These will also be sent to Los Alamos National Laboratory and Kyoto University. Intel also said that the integration of HBM memory will not require changes to the operating system and program code, and can be seamlessly upgraded for end users.
“To ensure no HPC workloads are left behind, we need a solution that maximizes bandwidth, maximizes compute, maximizes developer productivity, and ultimately maximizes impact. Intel Max Series The product brings high-bandwidth memory to a wider market, while also bringing oneAPI to make it easy to share code between CPUs and GPUs and solve the world’s biggest challenges faster.” Intel Corporate Vice President and Jeff McVeigh, general manager of the Supercomputing Group, said so.
The 56 cores are distributed across 4 tiles and are connected using Intel’s Multi-Chip Interconnect Bridge (EMIB), paired with a 64GB HBM. For the interface bus, the platform will feature PCIe 5.0 and CXL 1.1 I/O.
At the same HCPG performance, the Xeon Max uses 68% less power than the AMD Milan-X cluster. The AMX extension improves AI performance and provides 8x higher peak throughput than AVX-512 for cumulative INT8 versus INT32 operations, and provides flexible compatibility to run in HBM and DDR memory configurations.
Intel claims that the Xeon Max offers up to 5x better performance in some workloads compared to older Intel Xeon 8380 series processors or the AMD EPYC 7773X. It’s worth noting that AMD will announce its Genoa-based CPUs tomorrow, so the TCO analysis comparison will begin at that point.
Intel’s new Xeon Max CPUs also contains 20 accelerator engines for AVX-512, AMX, DSA, and Intel DL Boost workloads. In MLPerf DeepCAM training, Intel outperforms AMD’s 7763 by 3.6x and NVIDIA’s A100 by 1.2x.
The new Max CPU lineup will land in 2023 to counter AMD’s Genoa. AMD has previously been rumored to be considering HBM versions of its upcoming Genoa CPUs as well, but if they don’t — then that would give Intel a unique advantage in workloads with limited memory bandwidth.
Intel Xeon Max CPUs will debut in the Aurora supercomputers currently under construction at Argonne National Laboratory (these computers started shipping to them some time ago). Aurora is expected to be the first supercomputer to exceed 2 exaflops of peak double-precision computing performance. Aurora will also be the first to demonstrate the power of Max-series GPU and CPU pairing in a single system, with over 10,000 blades, each containing six Max-series GPUs and two Xeon Max CPUs.
