Today Intel is launching CXL or Compute Express Link along with the consortium that now controls it. At a high level CXL is CCIX with proprietary Intel sprinkles on top, something SemiAccurate feels is the point of the exercise.
To start out looking at the tech, the details are pretty sparse but as of this writing the entire spec should be up on the CXL web site here. (Note: You need to be a member to see the spec but membership is said to be free at the lowest level) So what is CXL? It is a protocol that rides over the physical layer of PCIe5.0 and offers a low latency memory interface/window and cache coherency. Sound familiar? If you said, “Gee that sounds exactly like CCIX“, no points because for the moment that is exactly what it is. (Note 2: We are more than happy to concede there may be substantial differences between CCIX and CXL but until _BOTH_ specs are public, we can’t say for sure)
CXL was developed by Intel and as of this writing the 1.0 spec has been given to the CXL Consortium. This consortium has nine members, Alibaba, Cisco, Dell EMC, Facebook, Google, Hewlett Packard Enterprise, Huawei, Intel, and Microsoft. The CXL Consortium (CXLC from now on) is said to be aiming for a yearly cadence on spec updates so expect the 2.0 next year and 3.0 to be announced in 2021 or about the same time as the first hardware bearing CXL 1.0 is likely to hit the market. For those keeping track of hardware roadmaps, think Sapphire Rapids. You may also notice that this means Sapphire will have PCIe5 not PCIe4.
There are three interesting technical-ish bits that are part of CXL that may or may not be the case with CCIX, backwards compatibility, asymmetry, and IP protection. The backwards compatibility part seems obvious enough and given that the engineer spearheading CXL for Intel, Jim Pappas, was responsible for USB, PCI, and PCIe, it seems likely they will pull it off. Asymmetry is another one that will be familiar to those who care about low level details, putting the complexity on the host side and making targets cheap and simple is usually a good thing for consumers.
Last up is that CXLC offers IP protection, a good thing, no a necessary thing in the space if you want to make silicon. How deep and how useful this protection is depends on the details which are not as of yet released, but given the track records of those involved it will probably be just fine.
That brings up the elephant in the room, the track record of those involved. CXL is nothing more, the note above notwithstanding, than an Intel branded non-compatible version of CCIX. At the public level of detail both do exactly the same thing other than CCIX being based on PCIe4 and has devices based on the spec available now. Intel may be able to learn from CCIX and do a better job but there will be at least three generations of CCIX based hardware available before the first CXL device hits the market. Better is a question you can’t answer yet, available is.
Then you get to motivations. Intel has long been about proprietary hardware with some semblance of open connectivity. USB 1/1.1 shenanigans aside the company has been pretty good about open interfaces until fairly recently. Then we came to Thunderbolt and ‘open’ being redefined as, “we are the only ones making silicon for the host” to intentionally shut out competition. When Thunderbolt was crushed by USB3 in the market, Intel played the lowest forms of IP and compliance games to stifle it’s adoption.
Those games worked well enough to delay mass market adoption of USB3 for 12-18 months. Like the USB 1/1.1 debacle this deadlock was only broken when the rest of the industry decided to tell Intel to get stuffed and threatened, credibly, to go their own way and make a new spec without Intel. Suddenly the compliance problems ceased and cheap USB3 silicon flooded the market. Thunderbolt eventually fixed most of the crippling problems it faced but silicon still costs well over an order of magnitude more than USB3 per port and is available from only one vendor, Intel. It is said to be an ‘open’ standard though.
While SemiAccurate doesn’t think CXL or the CXLC will play this kind of game, the incentive for any non-Intel hardware player to jump on board is roughly zero. If you look at the list of members for the CXLC, there is one thing that stands out, the lack of hardware makers. Of the eight non-Intel members, four are cloud providers, three are device makers/OEMs that are effectively in Intel’s pocket, and Huawei which could be a number of categories depending on which part of the company is pushing membership. Compare and contrast that with the CCIX membership list, specifically the hardware makers in each.
If you look at CXL from a high level view, it is just an Intel flavored CCIX that they can control. There are between zero and one hardware maker on board so that you are likely to get Intel CPUs, NICs, FPGAs, and storage supporting CXL in the 2021 time frame. Everything else on the market will support CCIX although some vendors may support both specs. Is CXL anything more than a late attempt to claw back control of a key piece of enterprise tech by Intel? Ask us in three years. Could CXL have a killer feature or two that CCIX lacks? Possibly but doubtful. Should hardware makers trust Intel to do the right thing here and not pull a USB3 or Thunderbolt? Trust is earned and Intel is in a very deep self-dug hole, but if you are making hardware that is your call.S|A
Latest posts by Charlie Demerjian (see all)
- Where do Intel and TSMC go from here? - Aug 11, 2020
- Intel should not launch Ice Lake-SP - Aug 3, 2020
- How fast is Intel’s Ice Lake-SP CPU? - Jul 30, 2020
- What is Intel making at TSMC? - Jul 28, 2020
- Intel’s 7nm meltdown takes it’s first high level head - Jul 27, 2020