Comments on: Eviden Mainstreams BXI Interconnect Thanks To Ultra Ethernet And AI Boom https://www.nextplatform.com/2024/12/10/eviden-mainstreams-bxi-interconnect-thanks-to-ultra-ethernet-and-ai-boom/ In-depth coverage of high-end computing at large enterprises, supercomputing centers, hyperscale data centers, and public clouds. Thu, 19 Dec 2024 17:58:34 +0000 hourly 1 https://wordpress.org/?v=6.7.1 By: Thomas Hoberg https://www.nextplatform.com/2024/12/10/eviden-mainstreams-bxi-interconnect-thanks-to-ultra-ethernet-and-ai-boom/#comment-242123 Thu, 12 Dec 2024 14:34:30 +0000 https://www.nextplatform.com/?p=145117#comment-242123 I remember my time with my Bull colleagues very fondly, when my employer was still part of Atos.

And I also remember them being extremely proud of BXI, while my personal interest at the time was more in data plane processing (PIN?) on networks e.g. via P4 and PIM on RAM. BXI was all about better bang for the buck and my focus was on advancing architectures for value driven computing.

If I was a betting man, I’d say Marvell for both technical and cultural reasons.

When I read what Marvell is up to in trying to create a cheaper proprietary alternative to HBM for AI ASICs, that is very similar to what drove Bull to create BXI.

]]> By: Calamity Jim https://www.nextplatform.com/2024/12/10/eviden-mainstreams-bxi-interconnect-thanks-to-ultra-ethernet-and-ai-boom/#comment-242089 Thu, 12 Dec 2024 01:51:20 +0000 https://www.nextplatform.com/?p=145117#comment-242089 In reply to Paul Berry.

I guess, while the API (Portals) would be the same (up to possible version numbers), a differentiator would come in at the effectiveness of SmartNIC offload provided by the respective systems (including provisioning of data structure space for zero copy pipelining), and how much of the related gains in efficiency and performance eventually propagate over to UEC as well (possibly through evolved hardware, like BXI v5). I imagine that Cornelis/Libfabric faces similar opportunities and challenges for performance and flexibility in their HW/SW design optimizations to “minimize impedance mismatch” between apps and fabric — if I understand well …

]]> By: Timothy Prickett Morgan https://www.nextplatform.com/2024/12/10/eviden-mainstreams-bxi-interconnect-thanks-to-ultra-ethernet-and-ai-boom/#comment-242073 Wed, 11 Dec 2024 22:52:13 +0000 https://www.nextplatform.com/?p=145117#comment-242073 In reply to Paul Berry.

Interesting question.

]]> By: Paul Berry https://www.nextplatform.com/2024/12/10/eviden-mainstreams-bxi-interconnect-thanks-to-ultra-ethernet-and-ai-boom/#comment-242071 Wed, 11 Dec 2024 20:41:02 +0000 https://www.nextplatform.com/?p=145117#comment-242071 Both BXI and HPE’s slingshot use Sandia’s Portals messaging library as the basis for their MPI implementations.
Does UEC define portals layers, or is that vendor specific? One wonders if the portals layer will be the same between Eviden and HPE, with only the MPI library differing?

]]> By: Timothy Prickett Morgan https://www.nextplatform.com/2024/12/10/eviden-mainstreams-bxi-interconnect-thanks-to-ultra-ethernet-and-ai-boom/#comment-242065 Wed, 11 Dec 2024 18:22:00 +0000 https://www.nextplatform.com/?p=145117#comment-242065 In reply to Eric Olson.

Oh, I liked Token Ring. But IBM said it did not scale really well for large rings even back in the 14 Mb/sec days….

]]> By: Eric Olson https://www.nextplatform.com/2024/12/10/eviden-mainstreams-bxi-interconnect-thanks-to-ultra-ethernet-and-ai-boom/#comment-242060 Wed, 11 Dec 2024 17:06:05 +0000 https://www.nextplatform.com/?p=145117#comment-242060 Could more networking standards be what the world of AI and HPC needs? Omnipath is out at 400 Gbps. Where is that next generation token ring network that messages directly with the ring bus connecting CPU cores?

]]> By: Calamity Jim https://www.nextplatform.com/2024/12/10/eviden-mainstreams-bxi-interconnect-thanks-to-ultra-ethernet-and-ai-boom/#comment-242041 Wed, 11 Dec 2024 12:50:12 +0000 https://www.nextplatform.com/?p=145117#comment-242041 Cool to see some competition on the Internet side of HPC interconnects. The three exafloppers on the current Top500 are HPE Slingshot-11 machines, and, with JEDI and JETI, it seems that Jupiter will be a Quad-rail Infiniband job … BXI brings in much needed diversity and competition in this space IMHO! I hope Alice Recoque can benefit from the BXI v3 innovations, as a step up from CEA-HE/HF motors that run BXI v2.

Meanwhile, BXI v4 (2027), with UEC, PCIe6 and CXL (hopefully 3.x+) should be the real game changer!

]]>