Comments on: Aurora In A Socket: What Intel’s “Falcon Shores” XPU Might Do https://www.nextplatform.com/2022/02/28/aurora-in-a-socket-what-intels-falcon-shores-xpu-might-do/ In-depth coverage of high-end computing at large enterprises, supercomputing centers, hyperscale data centers, and public clouds. Thu, 08 Jun 2023 03:27:22 +0000 hourly 1 https://wordpress.org/?v=6.7.1 By: Timothy Prickett Morgan https://www.nextplatform.com/2022/02/28/aurora-in-a-socket-what-intels-falcon-shores-xpu-might-do/#comment-182407 Thu, 10 Mar 2022 23:58:10 +0000 https://www.nextplatform.com/?p=140104#comment-182407 In reply to JayN.

I suppose it could be. But I was thinking higher up in the memory hierarchy. We are always limited by our initial assumptions.

]]>
By: JayN https://www.nextplatform.com/2022/02/28/aurora-in-a-socket-what-intels-falcon-shores-xpu-might-do/#comment-182405 Thu, 10 Mar 2022 23:56:08 +0000 https://www.nextplatform.com/?p=140104#comment-182405 “Koduri added that the Falcon Shores XPU would have what he called “extreme bandwidth shared memory,” ”

Why wouldn’t this just be another implementation of the Rambo L2 SRAM ache used on Ponte Vecchio?

One of the recent Intel patent applications shows an entire base layer with embedded sram chiplets. See Fig 32 of US20210150663.

]]>
By: Timothy Prickett Morgan https://www.nextplatform.com/2022/02/28/aurora-in-a-socket-what-intels-falcon-shores-xpu-might-do/#comment-182404 Thu, 10 Mar 2022 23:48:14 +0000 https://www.nextplatform.com/?p=140104#comment-182404 In reply to JayN.

Well, there are three different ways to skin that potential cat, and I just suggested one of them. I guess we will see.

]]>
By: JayN https://www.nextplatform.com/2022/02/28/aurora-in-a-socket-what-intels-falcon-shores-xpu-might-do/#comment-182403 Thu, 10 Mar 2022 23:36:29 +0000 https://www.nextplatform.com/?p=140104#comment-182403 “There could be an native oneAPI offload across what we presume will be UltraPath Interconnect links between the CPU and GPU parts of the chips, although they could be Xe links.”

I thought a primary feature of cxl is to provide the biased cache coherency, which effectively pools up all the cache coherency transfers while the slave has ownership. Why would they abandon that and return to UPI? Wouldn’t that also require the GPU to have the complicated and proprietary home agent?

]]>
By: Timothy Prickett Morgan https://www.nextplatform.com/2022/02/28/aurora-in-a-socket-what-intels-falcon-shores-xpu-might-do/#comment-179979 Wed, 02 Mar 2022 00:49:55 +0000 https://www.nextplatform.com/?p=140104#comment-179979 In reply to Venkat S.

HA! Yes, Indeed. I also presume not!

]]>
By: Venkat S https://www.nextplatform.com/2022/02/28/aurora-in-a-socket-what-intels-falcon-shores-xpu-might-do/#comment-179977 Wed, 02 Mar 2022 00:44:02 +0000 https://www.nextplatform.com/?p=140104#comment-179977 Hope it isn’t going to consume Aurora (wall) power as well, including cooling power

]]>