Note: This article originally appeared on theregister.com, click here to read the original piece.
Regardless, we’re looking forward to a shake-up in server architecture
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 composable infrastructure outfits adapt and thrive, or be relegated to the scrap pile of tech vendors that had the right idea before the market was ready for it? And by composable infrastructure, we mean rethinking your IT resources as pools of compute, storage, networking, and accelerators, and allocating these to workloads in the most efficient way on the fly as needed.
“For some of them,” Gartner analyst Tony Harvey told The Register, “it’s not going to make a lot of practical difference. If you’re a GigaIO, for example, you’re doing a PCIe switch, and you’re going to have to upgrade to PCIe 5.0, and you’re going to support CXL.”
While CXL may force these vendors to change and adapt, their investments in PCIe switching puts them in a unique position to lead a new market segment for CXL switches, he added.
The idea being that, eventually, CXL switches will replace Ethernet switches for in-rack communications. This will enable customers to interface any number of disaggregated compute, memory, or storage appliances within the rack. In other words, the dream of multivendor composable infrastructure fully realized.
However, CXL-based composable infrastructure remains years away. While CXL version 1.1 will make its debut with Intel’s Sapphire Rapids and AMD’s Genoa later this year, the switching and fabric capabilities required to realize these composable infrastructure use cases won’t arrive until the launch of CXL 2.0 and CXL 3.0-based silicon.
The imminent birth of a CXL switching ecosystem?
In many respects, CXL represents more of an opportunity for vendors like Liqid, GigaIO, and Fungible than a threat. However, the technology will undoubtedly drive competition in the niche composable infrastructure market, as major OEMs, such as Dell, HPE, and Lenovo, begin introducing their own CXL-based infrastructure.
As a result, Harvey predicts that many composable infrastructure vendors could find themselves the target of acquisitions by OEMs looking to bolster their PCIe switching capabilities.
Alternatively, he said, PCIe switching could follow the path taken by Ethernet, with these composable infrastructure vendors eventually becoming the CXL equivalents to Arista, Juniper, or Cisco.
- Marvell CXL roadmap goes all-in on composable infrastructure
- PCIe 7.0 pegged to arrive in 2025 with speeds of 512 GBps
- How CXL may change the datacenter as we know it
- Will optics ever replace copper interconnects? We asked this silicon photonics startup
This route isn’t without challenges of its own, and will depend heavily on whether the vendors are willing to open their platforms up to the kind of Ethernet-style standardization required for interoperability.
“To create that open market for PCIe switches, you’re going to have to be vendor independent and vendor neutral, and also create this Ethernet-style world where people expect everything to work with everything else,” Harvey said. “I don’t care who made my network card or my switch. I plug them in and they work.”
On the flip side, some vendors may double down on their proprietary architectures and create the CXL equivalent of Infiniband.
A disaggregated future
CXL switching is likely to have a much wider appeal than composable infrastructure, according to a recent Gartner report, which predicted the segment would account for less than 10 percent of enterprise server spend through 2025.
“There are bits of the market where it fits and bits of market where it doesn’t,” Harvey said. “In the high-performance compute space and managed service providers or cloud providers, it’s a beautiful solution.”
They compose once and then use it for five years and then shut it down
Most customers that are buying composable infrastructure today aren’t actually taking advantage of that composability in their day-to-day operation, he explained. “You talk to anybody, and they’ll tell you what they actually do is compose once and then use it for five years and then shut it down.”
Part of this is because many enterprise applications simply aren’t designed to take advantage of composability.
“Is anyone really going to take their Oracle ERP system, shut it down, resize the service it’s running on, then restart that ERP system in order to take advantage of composability,” he said. “The answer to that question is no.”
But that’s not to say there isn’t value in the modularity unlocked by these technologies.
“CXL has the opportunity to completely change the way we architect servers. That’s going to be really interesting in the market because we’ve not seen a server re-architecture for years. We’re all still building 1U or 2U rack servers that look identical,” Harvey said. “CXL can change all of that.”
So while enterprises might not recompose their infrastructure on the fly, the ability to deploy large pools of memory or other accelerators in a disaggregated fashion still has its appeal and presents an opportunity for the likes of GigaIO, Liqid, Fungible, and any number of other vendors to tie them all together. ®
Tobias Mann Sun 10 Jul 2022 // 12:59 UTC // 16:27 UTC
View source version on theregister.com: https://www.theregister.com/2022/01/20/compose_memory_across_a_hpc/