Cache-Coherent Interfaces for High-end GPUs

By: Jim Wilson (jim.wilson.delete@this.noemail.com), May 12, 2022 5:43 pm
Room: Moderated Discussions
In March 2022, NVIDIA announced that their H100 GPU has a 900 GBytes/sec cache-coherent interface to their upcoming ARM CPU. Compute Express Link (CXL), formerly called Intel Accelerator Link (IAL), provides 128 GBytes/sec over x16 PCIe gen5.

Questions for discussion:
1. Will NVIDIA’s NVLink and AMD’s Infinity Fabric replace PCIe as the CPU interface to most high-end non-Intel accelerators?

2. Should AMD, Intel and Apple make the PCIe pins on their server processors configurable as NVLink?

3. Should the current x16 PCIe interface for GPUs be replaced with a x32 or x64 interface in servers and workstations?

Eight channels of DDR5-4800 provides 307 GBytes/sec in one direction, which is 4.8x more than x16 PCIe gen5 provides in one direction (64 GBytes/sec). The price of a wider connector is completely insignificant compared to the price of a high-end GPU.

There could be motherboards with one x64 connection per CPU socket and other motherboards with two x32 connections per CPU socket. When there are two x32 connections between a CPU socket and two GPUs, x64 connectors could still be used and the motherboard could use half of each connector to connect one GPU directly to the other GPU. This could be indicated by using two different colors for each half of a x64 connector.
 Next Post in Thread >
TopicPosted ByDate
Cache-Coherent Interfaces for High-end GPUsJim Wilson2022/05/12 05:43 PM
  Cache-Coherent Interfaces for High-end GPUsWes Felter2022/05/13 01:11 PM
    Cache-Coherent Interfaces for High-end GPUsJim Wilson2022/05/13 03:41 PM
      Cache-Coherent Interfaces for High-end GPUsWes Felter2022/05/13 04:33 PM
        Cache-Coherent Interfaces for High-end GPUsJim Wilson2022/05/13 05:45 PM
          Cache-Coherent Interfaces for High-end GPUsanon22022/05/13 07:28 PM
            Cache-Coherent Interfaces for High-end GPUsJim Wilson2022/05/14 02:35 AM
          Cache-Coherent Interfaces for High-end GPUsWes Felter2022/05/14 09:22 AM
            Cache-Coherent Interfaces for High-end GPUsclue22022/05/14 09:54 AM
            Cache-Coherent Interfaces for High-end GPUsKevin G2022/05/17 02:40 PM
  Cache-Coherent Interfaces for High-end GPUsFreddie2022/05/14 07:43 AM
    Cache-Coherent Interfaces for High-end GPUsJim Wilson2022/05/14 02:11 PM
      Cache-Coherent Interfaces for High-end GPUsFreddie2022/05/14 03:13 PM
        Cache-Coherent Interfaces for High-end GPUsJim Wilson2022/05/14 05:28 PM
          Cache-Coherent Interfaces for High-end GPUsFreddie2022/05/15 02:12 AM
            Cache-Coherent Interfaces for High-end GPUsDoug S2022/05/15 10:25 AM
            Cache-Coherent Interfaces for High-end GPUsJim Wilson2022/05/15 07:37 PM
  Cache-Coherent Interfaces for High-end GPUsDoug S2022/05/15 10:45 AM
    Cache-Coherent Interfaces for High-end GPUsJim Wilson2022/05/15 05:15 PM
      Cache-Coherent Interfaces for High-end GPUsDoug S2022/05/16 12:58 AM
        Third Party GPUs in Future Mac ProJim Wilson2022/05/16 06:04 AM
          Third Party GPUs in Future Mac ProDoug S2022/05/16 09:05 AM
            Third Party GPUs in Future Mac ProJim Wilson2022/05/16 02:15 PM
              Third Party GPUs in Future Mac ProUngo2022/05/16 04:52 PM
                Third Party GPUs in Future Mac ProJim Wilson2022/05/16 06:10 PM
              Third Party GPUs in Future Mac Pro---2022/05/16 10:15 PM
                Third Party GPUs in Future Mac ProJim Wilson2022/05/17 06:44 AM
            Third Party GPUs in Future Mac Proanon12022/05/17 04:24 AM
          Third Party GPUs in Future Mac Pro---2022/05/16 10:04 PM
            2022 Mac ProJim Wilson2022/05/17 06:55 AM
          Third Party GPUs in Future Mac Proanon12022/05/17 01:27 AM
            Third Party GPUs in Future Mac ProJim Wilson2022/05/17 07:14 AM
              Third Party GPUs in Future Mac Proanon12022/05/18 12:51 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell avocado?