Intel on the eve of the event Supercomputing 22 has introduced Xeon Max data center class CPUs. It should be noted that earlier, these processors were referred to as Sapphire Rapids HBM. In addition, the blue team has unveiled the Intel Data Center GPU Max series graphics processors. Previously, the aforementioned GPUs were simply called Ponte Vecchio. Intel says its new processors are for the most complex supercomputer processing.
Xeon Max series CPUs
For decades, processors based on the x86 architecture have been used for a variety of technical processes; Therefore, today they can be used for a variety of processing tasks. Despite significant improvements in the performance of CPU cores in recent years, today’s processors face two major limitations in processing artificial intelligence and high-performance computing (HPC) tasks: one is memory bandwidth and the other is Parallelism. Intel promises that its Xeon Max series processors will bypass both limitations.
According to Toms HardoverXeon Max series processors are equipped with up to 56 powerful cores based on the Golden Cove microarchitecture. These cores are located in the form of four chiplets on a single platform, and Intel communicates between the chiplets through the exclusive EMIB technology. According to Intel, the new data center-class chips have special processing engines focused on artificial intelligence and HPC and use 64 GB of on-board HBM2E memory.
Xeon Max series processors have up to 56 high-power Golden Cove cores
New Intel processors, like other Sapphire Rapids series products, support eight DDR5 memory channels and PCIe Gen 5 interface with CXL 1.1 protocol. Intel’s improved cores use AVX-512 and Deep Learning Boost technology, and even the AMX matrix accelerator. New kernels also support DSA. This technique removes the pressure caused by data transfer from the processor.
Using the HBM2E’s 64GB of onboard memory (including four 16GB slots) leads to a maximum bandwidth of one terabyte, which equates to approximately 1.14GB of memory and 18.28GB/s of bandwidth per core. To better understand these numbers, it is better to know that the 56-core Sapphire Rapids processor, which has eight DDR5-4800 modules, achieves a bandwidth of 307.2 Gbps, or better said, 5.484 GB/s per core.
Xeon Max series processors can use HBM2E memory in different ways: used as system memory, used as a high-speed cache for the DDR5 memory subsystem, and used as part of integrated HBM memory. The last method requires software optimizations; But the first two methods do not need to change the code.
Depending on the processor type, Intel says the Xeon Max can perform about 3.0 to 5.3 times more powerful than the Xeon Scalable 8380. The Xeon Scalable 8380 processor is now available in the market. Thanks to its HBM2E memory, Intel’s new generation processors are up to 2.8 times more powerful than AMD EPYC 7773X in processes such as 3D model development for molecules. This AMD processor relies on 3D V-Cache.
Using HBM2E memory has another advantage for Intel, which is reducing the amount of data transfer between CPU and GPU; A topic that is very important for a number of HPC processes.
Intel GPU Max series graphics processors
Intel has said that its new graphics processors in the field of data centers are produced based on a version of the Ponte Vecchio architecture; An architecture that was first introduced in 2019. Intel has been quietly developing Ponte Vecchio for a long time until it released detailed information about it in 2020 and 2021.
Ponte Vecchio is the most complex processor ever built; Because excluding memory, it has more than 100 billion transistors in 47 processing tiles (including eight HBM2E tiles). Ponte Vecchio’s architecture extensively uses Intel’s modern chip packaging technologies such as EMIB. Several companies produce numerous Ponte Vecchio chiplets with various technologies.
Intel says it builds Data Center GPU Max GPUs using a modified Xe-HPC architecture that is specifically optimized for AI processing and the HPC domain. By using this architecture, new processors support a large number of formats and instructions.
|Technical specifications of Intel’s new data center class graphics
|Data Center Max 1100
|Data Center Max 1350
|Data Center Max 1550
|Processing tiles (chiplets)
|Xe HPC cores
|Ray interceptor cores
|512 bit vector engines
|4,096 bit matrix engines
|Cache L2 Rambo
Compared to Xe-HPG, the Xe-HPC architecture has a much larger number of subsystems for memory and cache and uses different cores. The Xe-HPC architecture does not have a texture unit; For this reason, it cannot process graphic content with traditional methods. Surprisingly, Xe-HPG supports ray tracing technology for data visualization on a supercomputer.
One of the important features of the Xe-HPC architecture is the XMX technique, which helps Intel’s new GPUs achieve 419 teraflops in the TF32 format. Preliminary tests show that Ponte Vecchio’s most powerful processor doesn’t come close to Nvidia H100 in most scenarios, and it doesn’t outperform AMD Instinct MI250X in almost all processes. Intel says the Data Center GPU Max 1550 is up to 2.4x faster in Riskfuel simulations and up to 1.5x faster in NekRS simulations than Nvidia’s A100.
Intel plans to release three new products based on Ponte Vecchio. The most powerful Data Center model is called GPU Max 1550 and includes 128 Xe-HPC cores and 128 GB of HBM2E memory and 600 watts of power. The second model, the Data Center GPU Max 1350, uses 112 Xe-HPC cores with 96 GB of memory and 450 watts of power. The third and weakest model is the Data Center GPU Max 1100. This processor has 56 He-HPC cores, 56 GB of HBM2E memory and 300 watts of power.
Rialto Bridge; The next generation of Intel data center graphics
In addition to the unveiling of the Data Center GPU Max series graphics, Intel released preliminary details of the Rialto Bridge series graphics; Graphics coming in 2024. Intel says products in this family will use improved Xe-HPC cores and will continue to be based on the Ponte Vecchio architecture. The use of new cores makes the power consumption in Rialto Bridge reach a maximum of 800 watts. Needless to say, models of these processors are produced with less power.
Intel says Argonne National Laboratory in the US is one of the first customers to receive the new Intel Xeon Max series and Intel Data Center GPU Max products. This research laboratory is engaged in the production of a supercomputer with a power of more than two exaflops. Intel’s business partners will launch Xeon Max and Data Center GPU Max-based devices for the server market early next year.