Jensen Huang, CEO of Nvidia, has introduced the corporate might be making a full transition to Intel processors (opens in new tab) for its upcoming DGX H100 unit and supercomputer initiatives sooner or later. Nvidia might be utilizing Intel’s forthcoming Rapids Sapphire Xeon processor lineup as a complete substitute for AMD’s Zen 3 EPYC CPU, which Nvidia has been utilizing extensively for years.
Huang says the first cause for switching CPU manufacturers was the distinctive single-threaded efficiency Sapphire Rapids presents over the competitors. It makes a whole lot of sense, contemplating Sapphire Rapids Xeon Scalable processors are already transport to finish clients. AMD, in the meantime, has to do with its present Zen 3 and Zen 3 V-Cache EPYC processors, as the corporate has nonetheless not introduced a launch date for its upcoming Genoa Zen 4-based server processors, which would be the direct countermeasure to Sapphire Rapids.
We nonetheless haven’t got full particulars on Sapphire Rapids but, however we already know extra info based mostly on leaks over the previous 12 months. Sapphire Rapids is Intel’s next-generation server structure and can use the identical Golden Cove efficiency cores as Intel’s desktop Alder Lake structure. Sapphire Rapids is the server model of Alder Lake.
Because of this, Sapphire Rapids ought to have an analogous IPC uplift as Intel’s Alder Lake structure. Sapphire Rapids may also assist the most recent reminiscence and storage applied sciences, together with DDR5, HBM2E, and PCIe Gen 5, bringing it updated with the most recent and most exceptional applied sciences.
Core counts ought to peak at 56 cores on the very best trim ranges, with 80 PCIe Gen 5 lanes. The one downside with Intel’s core specs is that AMD far outperforms Intel in core rely, with its top-of-the-line chips, such because the EPYC 7773X, characteristic 128 cores, which is greater than double the core rely of Sapphire Rapids.
It ought to make Sapphire Rapids inferior in multi-threaded efficiency in comparison with Zen 3 EPYC, and in keeping with leaked benchmark outcomes of Sapphire Rapids, this seems to be true. Nonetheless, for Nvidia, it appears this might be a non-issue as it’s focusing virtually totally on single-threaded efficiency for its H100 DGX items.
Sapphire Rapids may also assist a number of different newer applied sciences Intel’s earlier Ice Lake server structure didn’t. A few of these embody the brand new AMX AVX512_BF16 instruction set for deep studying functions and a particular Knowledge Streaming Accelerator that offloads all storage calls from the CPU to a devoted chip to lower CPU utilization.
For Nvidia, Sapphire Rapids will present its DGX items with far increased single-threaded efficiency plus increased reminiscence bandwidth and extra PCIe bandwidth in comparison with AMD’s present Zen 3 choices.
DGX H100 caters to AI-intensive functions particularly, with every DGX unit that includes 8 of Nvidia’s model new Hopper H100 GPUs with a efficiency output of 32 petaFlops. For extra particulars, try our protection.