The new 3rd generation Gaudi chipsetdesigned for highly efficient and scalable business environments, comes hand in hand with significant performance improvements that Intel summarizes in twice the performance capacity in AI computing in FP8, up to 4 times more performance in BF16 operations, twice the width of network bandwidth and up to 1.5 times more memory bandwidth.
Going into the details of this new unit, which can be parallelized on the same server, we find 64 Tensor units, 8 mathematical matrices, up to 128GB of integrated HBM memory with up to 3.7TB/s bandwidth. It is also supported by 96MB of SRAM memory with 12.8TB/s of bandwidth.
Connectivity includes important new features such as 24 200GbE RoCE links completely open standard unlike the previous generation that used infiniband which is a proprietary solution from Intel. In this way the manufacturer opens up to a completely open solution.
This chipset also includes 16 PCI Express 5.0 links for host connectivity to the server, making it one of the first expansion cards with this link capacity.
Its scalability allows you to generate clusters of up to 1024 nodes with a total of 8192 Gaudi 3 accelerators making it one of the most scalable and massive solutions on the market in a clear commitment by Intel to fight for this market segment with NVIDIA, which is today the undisputed leader in dedicated AI accelerators.
In fact, the brand claims that its Gaudi 3 solutions can be Up to 40% faster on configurations of up to 8192 accelerators simultaneous in languages like GPT3-175B or up to 15% faster than the same NVIDIA H100 in LLAMA2-70B. In inference, it promises even more performance, with up to 1.5 times more speed than NVIDIA’s most advanced solutions to date based on H-range accelerators.
These new accelerators are accompanied today by the launch of the new sixth generation Xeon processors, with immediate availability for the Xeon 6700E models based on Crestmont architecture E-Core.
End of Article. Tell us something in the Comments!
Add Comment