Comments on: This AI Network Has No Spine – And That’s A Good Thing https://www.nextplatform.com/2024/08/23/this-ai-network-has-no-spine-and-thats-a-good-thing/ In-depth coverage of high-end computing at large enterprises, supercomputing centers, hyperscale data centers, and public clouds. Thu, 29 Aug 2024 03:14:55 +0000 hourly 1 https://wordpress.org/?v=6.7.1 By: tom https://www.nextplatform.com/2024/08/23/this-ai-network-has-no-spine-and-thats-a-good-thing/#comment-232908 Mon, 26 Aug 2024 17:03:05 +0000 https://www.nextplatform.com/?p=144567#comment-232908 Grammar police here: it’s “comprised from”, “composed of”. Comprised basically means “squished together”, so “it was squished together from parts”. I thought this was the first thing they teach you when you get a job in writing. 😉

]]> By: Slim Albert https://www.nextplatform.com/2024/08/23/this-ai-network-has-no-spine-and-thats-a-good-thing/#comment-232865 Sun, 25 Aug 2024 22:22:39 +0000 https://www.nextplatform.com/?p=144567#comment-232865 Cool idea to grant this the TPM “obvious once you think about it” (or OOY-TAI) award! As long as the HBDs have fast all-to-all interconnects, severing (or inverting) the spine does look like a winning proposition. The “outer” network then goes from a rather hierarchical leaf-spine, spine-rib, or hub-spoke Clos configuration, with ingress/afferent, central/middle, and egress/efferent parts, into more of a synaptic exoskeleton, a Fujitsu-style n-D TOFU torus, a near-mesh, or a dragonfly hairnet it seems, but with crisscrossing tracks (or bias ply).

I’d invest the money saved by going spineless on getting the fastest possible switches ever for that rail network though, not putting it in a piggy bank.

]]> By: Tony dixon https://www.nextplatform.com/2024/08/23/this-ai-network-has-no-spine-and-thats-a-good-thing/#comment-232829 Sun, 25 Aug 2024 17:10:48 +0000 https://www.nextplatform.com/?p=144567#comment-232829 This is a interesting concept and is similar to the host network concept described in my book (found here https://www.amazon.co.uk/introduction-compartmentalisation-management-migration-micro-segmentation/dp/B09L3PNXZL)

Where this article talks about the use of shortest routes to remove switches from the network fabric, my concept is discuss in the context of improving speed and security within business domains and private clouds. I love to explore if these two concepts could support each other in a wider range of workloads.

]]> By: Timothy Prickett Morgan https://www.nextplatform.com/2024/08/23/this-ai-network-has-no-spine-and-thats-a-good-thing/#comment-232813 Sun, 25 Aug 2024 14:04:19 +0000 https://www.nextplatform.com/?p=144567#comment-232813 In reply to Rys.

Ah yes, but this eliminates the spine and its costs and all of those transceivers! And makes use of networking already in the box.

]]> By: itellu3times https://www.nextplatform.com/2024/08/23/this-ai-network-has-no-spine-and-thats-a-good-thing/#comment-232776 Sun, 25 Aug 2024 06:30:17 +0000 https://www.nextplatform.com/?p=144567#comment-232776 Seems a very complex way to describe a well-localized geometry of processing, so you can use a low bandwidth global link.
This only duplicates the processor-cache1-cache2-cache3-etc locality of processing pattern known for a thousand years.

]]> By: Rys https://www.nextplatform.com/2024/08/23/this-ai-network-has-no-spine-and-thats-a-good-thing/#comment-232601 Sat, 24 Aug 2024 06:45:00 +0000 https://www.nextplatform.com/?p=144567#comment-232601 Cell based fabric in Arista made sure to evenly use the network links in a rail optimised design and avoids the idle condition of links.

]]>