Looking forward to competing with NVIDIA, AMD and other competitors in the artificial intelligence market, Intel has shared some details about the upcoming Gaudi 3, which will represent the last model before Intel merges its Gaudi and GPU series into a single product Falcon Shores Guadi Artificial Intelligence Accelerator.
Intel said that Gaudi 3, built on a 5-nanometer process, has four times the performance of Gaudi 2 in terms of BF16 workloads, twice the network performance of Gaudi 2, and 1.5 times the HBM capacity of Gaudi 2. Gaudi 3 also moved to a design with a graphics unit with two compute clusters, rather than the single-chip solution Intel used for Gaudi 2.
Although the various graphics units of Habana Gaudi IP and Xe GPU IP are merged, the next Falcon Shores will look and function as a single GPU through the OneAPI programming interface. Falcon Shores will use HBM3 memory and Ethernet switching and support the CXL programming model. Additionally, applications tuned for Gaudi accelerators and Xeon Max GPUs will be forward-compatible with Falcon Shores, providing customers with code continuity between the two distinct GPU and Gaudi families.
Additionally, Intel’s data center GPU Max series is now shipping to customers, with Supermicro offering systems with eight OAM-spec GPUs, and Dell and Lenovo offering four OAM GPU servers. And Intel’s benchmark compares the OAM form factor’s Max 1550 (a 600W GPU) to Nvidia’s PCIe form factor H100 (a 350W competitor). Therefore, these benchmarks are not a case of comparing performance. Intel says the reason for the benchmark differences is the difficulty in getting the OAM form factor H100 GPU.
Intel further pointed out that now we are waiting for the Aurora supercomputer submitted by the Argonne National Laboratory in the United States to participate in the Top500 auction to see if Intel can replace Frontier, which has the computing power provided by AMD products, to further become the fastest in the world. supercomputer.
(Source of first image: Provided by Intel)
Follow TechNews via Google News here
New scientific and technological knowledge, updated from time to time