Intel Delivers 10,000 Aurora Supercomputer Blades, Benchmarks Against Nvidia and AMD

With two exaflops of efficiency, the Intel-powered Aurora supercomputer is anticipated to beat the AMD-powered Frontier supercomputer, presently the quickest on the earth, and topped the Prime 500 listing of quickest supercomputers. Nonetheless, on account of Intel’s ongoing delays in offering the {hardware}, Aurora was unable to make the listing introduced as we speak because it has not but submitted a benchmark to the Prime 500 committee. Intel as we speak shared new particulars concerning the system and introduced at its ISC convention that it has supplied greater than 10,000 working wings for the Aurora supercomputer; actual blades required for full deployment. We’ll cowl the small print beneath.

However Intel says the system might be totally operational later this yr, and has shared head-to-head benchmarks with Aurora in opposition to AMD and Nvidia-powered supercomputers, with a 2x efficiency benefit over AMD’s MI250X GPUs and a 20% over Nvidia’s H100. Claims revenue. GPUs.

image one associated to 2

Intel says it provides silicon to the Argonne Management Computing Facility (ALCF) for greater than 10,000 blades, each fourth-generation Sapphire Rapids Xeon chips and Ponte Vecchio GPUs.

Nonetheless, the Aurora is designed to work with Intel’s continuously delayed HBM-equipped Sapphire Rapids “Xeon Max” chips. On account of these delays, Intel initially began transport non-HBM Sapphire Rapids chips to ALCF, and the ability started populating the Aurora with customary non-HBM Sapphire Rapids chips as a brief measure.

Intel is now offering the ALCF with quicker HBM-equipped Xeon Max chips, however not all the 10,000 blades it claims to be delivered have Max chips. We requested Intel and firm representatives confirmed that not all wings are geared up with the last word Xeon Max silicon. The corporate tells us that about 75% of the blades comprise the newest Xeon Max revision of silicon. Presumably, that is the bottleneck that stops the system from providing a benchmark for the Top500 listing.

The system consists of 166 cupboards, 64 blades per rack, for a complete of 10,624 blades, so the greater than 10,000 blades delivered are adequate for the system to function – not at full efficiency.

Intel has additionally shared extra specs for the Aurora supercomputer, together with detailed specs you possibly can see within the slide above. With 21,248 CPUs and 63,744 Ponte Vecchio GPUs, the Aurora will both catch or exceed two exaflops of efficiency when totally on-line earlier than the tip of the yr. The system additionally has 10.9 petabytes (PB) of DDR5 reminiscence, 1.36 PB of HBM hooked up to CPUs, 8.16 PB of GPU reminiscence, and 230 PB of storage capability offering 31 TB/s of bandwidth (different fascinating particulars, scroll up).

Intel additionally introduced that Aurora will start operating productive AI workloads throughout a spread of workloads. The ‘Aurora GPT’ main language mannequin might be science pushed and could have 1 trillion parameters with Megatron and DeepSpeed ​​foundations. Intel supplied the next abstract of the mission:

“These generative AI fashions for science might be skilled on normal textual content, code, scientific texts and structured scientific knowledge from biology, chemistry, supplies science, physics, medication and different sources. The ensuing fashions (with as much as 1 trillion parameters) techniques biology, polymer chemistry and vitality supplies can be utilized in a wide range of scientific purposes, from the design of molecules and supplies to the synthesis of data from tens of millions of sources to suggest new and fascinating experiments within the fields of local weather science and cosmology. It may be used to speed up the identification of organic processes associated to most cancers and different illnesses and to suggest targets for drug design.”

image one associated to 4