Will CXL spell the end for boutique composable infrastructure vendors?

Compute express link (CXL) is often touted as the holy grail of composable infrastructure for its ability to disaggregate system memory from the processor. With the first compatible processors just months away, one has to wonder: what will happen to the likes of GigaIO, Fungible, or Liqid when CXL finally arrives? Will these …

[TeeCee]

...ability to disaggregate system memory from the processor.

Oooo. Local Bus is back!

To paraphrase Pete Townshend: Meet the new bus, same as the old bus...

[Anonymous Coward]

Way too soon

While the article is interesting, it's background, not actionable, information. As the article points out, this is years away in an industry that thinks in fiscal quarters. As such, I see no incentive for PCIe switch vendors to move, or even think very much about ditching their proprietary protocols.

If I had to speculate (and it's Sunday so, why not), Infiniband's strategy seems much more likely for PCIe switch vendors, given the realistic target audience for CXL. Small datacenters do not seem likely to invest in the paradigm shift until this becomes a proven technology. Its early implementations are likely to be in HPC and HPC systems are more than ready to adopt proprietary kit like Infiniband.

I would predict that the vendors, to the extent that they are doing anything, are polishing their balance sheets and hoping for a buyout.

But, as I said, it's way too soon. It won't even show up on Gartner Trends for a couple of years.

[DS999]

This is really more similar to fiber channel switches

Both in terms of how specialized the market will be and in terms of how it'll be used in the datacenter.

[CapeCarl]

big/LITTLE (server not core)

Hmmm, is this a niche solution looking for an industry-wide problem?

Is the overhead cost of connecting PCIe 5/6/7 at rack level (retimers? optical?) worth more than the overhead of a few stranded vCPUs or GBs of RAM within a pizzabox-ish server in the Cloud or a large cluster?

---

I think back to my previous (almost) decade with a 100K+ core financial research cluster, yes some server nodes had 2x the RAM of others (a modest bow to the fact that not all jobs were the same size)...But it could seem that over the past couple decades that the world has gotten better at sharding large jobs over multiple small servers (obviously not all problems can be decomposed this way, but based on that fact that we have been living with the pizzabox hardware economic model for so long, many/most jobs must be able to fit into this hardware layout).

So as in the case of my previous IT life, with the proposed rack-level decomposed hardware architecture (nRU box of RAM, nRU box of compute cores, ... and a CXL TOR switch), the cluster job scheduler now adds the step of looking for a rack with the required available resources in order to compose a server to run a given job (that has been defined well enough for all this effort of perfectly sizing the server to be worth the cost).

Hmmm I don't know, maybe this Brave New World quickly degenerates into a pile of greater-than-pizzabox sized servers (database, JVMs...) and pile of smaller-than-pizzabox sized servers (web, ...)....Which isn't a whole lot different than what I was doing circa 2000 when specing $5K Sun boxes as web servers up to $800K Sun boxes for large Oracle databases.

Maybe this all works IF the hardware costs for composability head towards zero...Yes one can swap out CPUs, RAM etc independently over time, but hardware gets written off in 3 - 5 years anyway and the cost of hardware swapping in a data center is not free (I have done a lot of it)...So yeah, that rack-level composability cost has to be cheaper / overcome the years of Cloud vendor research into packing VMs into smallish (1 or 2 socket) servers and sharding (decomposing) large jobs over lots of said "standard" servers.

....Or maybe this topic is just too large for my 1 vCPU / fixed-sized RAM brain to faithfully comprehend // Ah, maybe in The Matrix that brain connector could let me use "human CXL" to borrow unused brain cells from other nearby pods...