Skip to main content

Nvidia Switches Gears, Chooses Sapphire Rapids Over AMD EPYC

DGX H100
DGX H100 (Image credit: Nvidia)

Jensen Huang, CEO of Nvidia, has announced the company will be making a complete transition to Intel processors (opens in new tab) for its upcoming DGX H100 unit and supercomputer projects in the future. Nvidia will be using Intel's forthcoming Rapids Sapphire Xeon processor lineup as a total replacement for AMD's Zen 3 EPYC CPU, which Nvidia has been using extensively for years.

Huang says the primary reason for switching CPU brands was the exceptional single-threaded performance Sapphire Rapids offers over the competition. It makes a lot of sense, considering Sapphire Rapids Xeon Scalable processors are already shipping to end customers. AMD, meanwhile, has to do with its current Zen 3 and Zen 3 V-Cache EPYC processors. The company has still not announced a release date for its upcoming Genoa Zen 4-based server processors, which will be the direct countermeasure to Sapphire Rapids.

We don't have full details on Sapphire Rapids yet, but we already know more information based on leaks over the past year. Sapphire Rapids is Intel's next-generation server architecture and will use the same Golden Cove performance cores as Intel's desktop Alder Lake architecture. Sapphire Rapids is the server version of Alder Lake.

As a result, Sapphire Rapids should have a similar IPC uplift as Intel's Alder Lake architecture. Sapphire Rapids will also support the latest memory and storage technologies, including DDR5, HBM2E, and PCIe Gen 5, bringing it up to date with the latest and most remarkable technologies. 

Core counts should peak at 56 cores on the highest trim levels, with 80 PCIe Gen 5 lanes. The only problem with Intel's core specifications is that AMD far outperforms Intel in core count, with its top-of-the-line chips, such as the EPYC 7773X, feature 64 cores, eight more than Sapphire Rapids. 

It should make Sapphire Rapids inferior in multi-threaded performance compared to Zen 3 EPYC, and according to leaked benchmark results of Sapphire Rapids, this appears to be true. However, for Nvidia, it seems this will be a non-issue as it is focusing almost entirely on single-threaded performance for its H100 DGX units.

Sapphire Rapids will also support several other newer technologies Intel's previous Ice Lake server architecture did not. Some of these include the new AMX AVX512_BF16 instruction set for deep learning applications and a special Data Streaming Accelerator that offloads all storage calls from the CPU to a dedicated chip to decrease CPU utilization.

For Nvidia, Sapphire Rapids will provide its DGX units with far higher single-threaded performance plus higher memory bandwidth and more PCIe bandwidth compared to AMD's current Zen 3 offerings. 

DGX H100 caters to AI-intensive applications in particular, with each DGX unit featuring 8 of Nvidia's brand new Hopper H100 GPUs with a performance output of 32 petaFlops. For more details, check out our coverage.

Aaron Klotz
Aaron Klotz

Aaron Klotz is a freelance writer for Tom’s Hardware US, covering news topics related to computer hardware such as CPUs, and graphics cards.

  • escksu
    I am not surprised. Adding HBM to the CPU creates a massive performance boost.

    Although we still do not have exact benchmark numbers, but we do know up to 64GB of HBM2e and possibily 1.6TB/s of bandwidth. This is way faster than way bigger than AMD's vcache.
    Reply
  • ccoonansr
    Admin said:
    Nvidia has switched from AMD Zen 3 EPYC chips to Intel's new Sapphire Rapids Xeon CPUs for its new Hopper DGX H100 AI unit.

    Nvidia Switches Gears, Chooses Sapphire Rapids for DGX H100 : Read more
    In the article it says 7773x has 128 cores. This is a mis print. It has 64 cores and 128 threads. Which is 8 more cores than intel. Not double.
    Reply
  • KananX
    So their server GPUs behave exactly like their current GeForce GPUs, hungry for single threaded CPU performance, too dependent on single cores. It’s a bit awkward.
    Reply
  • vertuallinsanity
    This story is legitimate foreshadowing.

    (They won't let me post on TPU anymore):
    Reply
  • saltweaver
    Is Saphirre R archtecture similar to Metero lake, in other words, will Meteor Lake use HBM2E?
    Reply
  • JayNor
    The SPR has dual avx512 SIMD per core+ the AMX tiled matrix operation unit per core + the enhanced IO feeding them: DSA + 8channels DDR5 + 80 lanes pcie5 +CXL. Alder Lake has none of this, so not really going to be close to SPR IPC, especially for the HPC / AI applications that NVDA is targeting.
    Reply
  • Makaveli
    vertuallinsanity said:
    This story is legitimate foreshadowing.

    (They won't let me post on TPU anymore):

    lol so you came here to troll instead nice!!
    Reply
  • dalek1234
    "single threaded performance". I call total BS.

    A CPU with 54 (or whatever the number is) cores, and you're planning on using just one core? Jensen must think we are all stupid. The truth is most likely AMD. Nvidia doesn't want help AMD grow by giving them money.

    The problem I see with this reasoning is "power-consumption". The biggest cost for a lifetime of a supercomputer is electricity. I wonder how many customer's will walk away from DGX H100 AI becasuse it has "Intel, power hungry, inside.

    I have another theory about Nvidia choosing Intel: AMD not willing to sell CPUs to Nvidia, because AMD needs them for their own CPU + Accelerator Cards offerings for upcoming supercomputers, and doesn't have any, or many, to spare.
    Reply
  • KananX
    dalek1234 said:
    "single threaded performance". I call total BS.

    A CPU with 54 (or whatever the number is) cores, and you're planning on using just one core? Jensen must think we are all stupid. The truth is most likely AMD. Nvidia doesn't want help AMD grow by giving them money.

    The problem I see with this reasoning is "power-consumption". The biggest cost for a lifetime of a supercomputer is electricity. I wonder how many customer's will walk away from DGX H100 AI becasuse it has "Intel, power hungry, inside.

    I have another theory about Nvidia choosing Intel: AMD not willing to sell CPUs to Nvidia, because AMD needs them for their own CPU + Accelerator Cards offerings for upcoming supercomputers, and doesn't have any, or many, to spare.
    No it makes sense. Nvidia uses a software scheduler probably even still with newer GPUs and those are heavily single thread dependent unlike the hardware scheduler that AMD uses since a long time now, at least R9 290 days, which works way better with CPUs that have less single thread power but a lot of cores.
    Reply
  • dalek1234
    KananX said:
    No it makes sense. Nvidia uses a software scheduler probably even still with newer GPUs and those are heavily single thread dependent ...

    Interesting. I didn't know that.
    Reply