Surhat

Healthy and General

Next Year’s AURORA Computer

3 min read
Next Year's AURORA Computer

Next Year’s AURORA Computer: The two products announced from this series are the latest Intel Xeon Max Series super computer processors, and the data center GPU, the Intel GPU Max Series. Next Year’s AURORA Computer

Intel Xeon Max Series

The Xeon Max CPU is a high memory bandwidth x86-bas processor intend to accelerate many HPC workloads without having to change programs. The processor offers up to 56 performance cores divid into 4 tiles and connected using Intel’s integrated multi-die interconnect bridge (EMIB) technology, with a 350-watt envelope. This processor has an in-package high bandwidth memory (high bandwidth memory/HBM) of 64GB. Apart from that, it is also support with PCI Express 5.0 and CXL1.1 I/O.

Capacity is just over 1GB per core, and is claim to be enough to handle most of the typical HPC workload. Compar to its competitors, the Intel Xeon Max Series is claimed to be able to provide performance up to 4.8x better than competitors on real HPC workloads.

Some of the improvements offered include:

Up to 68% less power consumption than AMD Milan-X cluster for the same HCPG performance.
The AMX extension boosts AI performance and provides 8x higher peak throughput than AVX-512 for INT8 with INT32 stack operation

Provides flexibility to work with different HBM and DDR memory configurations.
Workload benchmarks:

Climate modeling: 2.4x faster than AMD Milan-X on MPAS-A using HBM only.
Molecular Dynamics: In DeePMD, a 2.8x increase in performance compared to competing products with DDR memory.

Intel GPU Max Series

As for the Intel GPU Max Series, this is the Intel graphics processor with the highest density. It contains more than 100 billion transistors in a 47-tile package with a very high memory bandwidth of up to 128 GB. Equipped with an open software ecosystem – oneAPI and Intel’s AI Tools, provides a variety of single programming capabilities, and can enable advanced features of Intel Max Series products.

Intel Max Series GPUs provide up to 128 Xe-HPC cores, the latest architecture targeting the most demanding compute workloads.

Some of the other advantages offered include:

408MB L2 cache– the highest in the industry – and 64MB L1 cache for increased throughput and performance
The only HPC/AI GPU with native ray tracing acceleration, designed to accelerate scientific visualization and animation
Workload benchmarks:
Finance: 2.4x greater performance achievement compared to NVIDIA’s A100 for Riskfuel credit option pricing calculations.
Physics: 1.5x improvement over A100 for NekRS virtual reactor simulation.

The Max Series GPUs will be available in several form factors to address different customer needs:

GPU Max Series 1100:300-watt double-wide PCIe card with 56 Xe cores and 48GB of HBM2e memory. Multiple cards can be connect via the Intel Xe Link bridge.

GPU Max Series 1350: 450-watt OAM module with 112 Xe cores and 96GB HBM

GPU Max Series 1550: A maximum-performance 600-watt OAM module from Intel with 128 Xecore and 128GB HBM.
Beyond individual cards and modules, Intel will offer Intel Data Center GPU Max Series subsystems with a carrier x4 GPU OAM board, and Intel Xe Link to support high-performance multi-GPU communications within these subsystems.

Coming in 2023, Used on the AURORA Supercomputer at Argonne National Laboratory

As for availability. The Intel Max Series products are schedul to launch in January 2023. In addition, both the Intel Xeon Max series processors and the Intel GPU Max Series will be used in the AURORA supercomputer being built at Argonne National Laboratory.

It is expected to be the first supercomputer with a peak double-precision computing performance of up to 2 exaflops. Aurora will also be the first supercomputer to demonstrate the power of a Max Series GPU and CPU pair in a single system, with more than 10,000 blades, each featuring 6 Max Series GPUs and 2 Xeon Max CPUs.