Intel introduced the Xeon Max, a line of processors that come with high-speed onboard HBM2e memory, and according to the company, these are the world’s first CPUs with onboard HBM. Previously, these chips appeared under the code name Saphir Rapids HBM. The processors offer up to 56 cores with support for 112 virtual threads and have a TDP of 350 W and are designed for high-performance server systems.
Xeon Max processors use the EMIB interface, which combines four dies with compute cores and 64GB of high-speed HBM2e memory co-located on the same substrate, divided into four clusters of 16GB each. Total throughput is around 1 TB/s. This means that more than 1 GB of HBM2e memory is available per processor core.
Key features of Xeon Max are support for PCIe 5.0 and CXL 1.1 interfaces. HBM2e storage can be used both as additional cache and additional RAM. In addition, a server with Xeon Max cannot be equipped with RAM modules at all – the system relies solely on HBM.
The company claims that Xeon Max power consumption is 68% lower than AMD Milan-X processors for the same performance. Support for the new AMX instructions on Xeon Max accelerates AI tasks and delivers up to eight times the peak performance of AVX-512 instructions in INT8 and INT32 operations.
Intel claims that the Xeon Max is up to five times faster than the Intel Xeon 8380 or AMD EPYC 7773X (Milan-X with 3D V-Cache technology) on some operations. The presentation also compares the Xeon Max to the AMD EPYC 7763 model, against which the new Intel shows up to 3.6 times better performance.
Compared to the NVIDIA A100 Server Accelerator, the Xeon Max is up to 1.2x faster than the competition in the MLPerf DeepCAM test, which addresses computing that accelerates and enhances AI supercomputer simulations.
The Xeon Max series of server processors will be released in January 2023. The main competitors of Xeon Max will be the new AMD EPYC Genoa processors. Their announcement is expected for tomorrow, November 10th. According to rumors, these processors will be released, including in the form of solutions equipped with HBM memory.