Cloud Hosting

When system architects sit down to design their next platforms, they start by looking at a bunch of roadmaps from suppliers of CPUs, accelerators, memory, flash, network interface cards and PCI-Express controllers and switches. And the switches are increasingly important in system designs that have a mix of compute and memory types and for clusters that will be sharing components like accelerators and memory.

The trouble is this: The roadmaps are not really aligned well. Most CPU and GPU makers are trying to do major compute engine upgrades every two years, with architectural and process tweaks in the year in between the major launches so they have something new to sell every year. Makers of chips for networking switches and interface cards in the Ethernet and InfiniBand markets tend to be on a two-year cadence as well, and they used to tie their launches very tightly to the Intel Xeon CPU launch cadence back when that was the dominant CPU in the datacenter, but that rhythm has been broken by the constantly redrawn roadmaps from Intel, the re-emergence of AMD as a CPU supplier, and a bunch of other Arm CPU makers, including at least three hyperscalers and cloud builders.

And then there is the PCI-Express bus, which has been all over the place in the past two decades. And while PCI-Express specifications have been released in a more predictable fashion in recent years, PCI-Express controllers have been faithful to the PCI-Express roadmaps but PCI-Express switches are well behind when it comes to product launches from MicroChip and Broadcom.

Sitting here on a quiet July morning, thinking about stuff, we think all of these roadmaps need to be better aligned. And specifically, we think that the PCI-SIG organization that controls the PCI-Express specification and does so through a broad and deep collaboration with the IT industry, needs to pick up the pace and get on a two-year cadence instead of the average of three it has shown in the past two decades. And while we are thinking about it, we think the industry would be better served with a short-cadence jump to PCI-Express 7.0, which needs to be launched as soon as possible to get I/O bandwidth and lane counts in better alignment with high throughput compute engines and what we expect will be an increasing use of the PCI-Express bus to handle CXL-based tiered and shared main memory.

Dont get us wrong. We are grateful that the PCI-SIG organization, a collaboration between all kinds of companies in the datacenter and now at the edge, has been able to get the PCI-Express bus on a predictable roadmap since the very late PCI-Express 4.0 spec was delivered in 2017. There were some tough signaling and materials challenges that kept the datacenter stuck at PCI-Express 3.0 for seven years, and we think Intel, which dominated CPUs at the time and dragged its feet a little bit on boosting I/O because it got burned with SATA ports in the chipsets used with the Sandy Bridge Xeon E5s that came out later than expected in March 2012. Rumors abounded about the difficulties of integrating PCI-Express 4.0 and PCI-Express 5.0 controllers into processors since then.

Generally, a PCI-Express spec is released and then within about a year or so we see controllers embedded in compute engines and network interface chips. So when PCI-Express 4.0 came out in 2017, we saw the first systems using it coming out in 2018 specifically, IBMs Power9-based Power Systems machines, followed by its use in AMD Rome Epyc 7002s launched in August 2019. Intel didnt get PCI-Express 4.0 controllers into its Xeon SP processors until the Ice Lake generation in April 2021.

And even with the short two-year jump to the PCI-Express 5.0 spec in 2019, it wasnt until IBM launched the Power10 processor in its high-end Power E1080 machines in 2021 that it became available in a product. AMD didnt get PCI-Express 5.0 into a server chip until the Genoa Epyc 9004s launched in November 2022 and Intel didnt get PCI-Express 5.0 into a server chip until the Sapphire Rapids Xeon SPs launched in January 2023.

So it was really a three-year cadence between PCI-Express 4.0 and 5.0 products, as expressed in the controllers on the CPUs, even if the spec did a two-year short step.

We think that the specs and the products need to get on a shorter two-year cadence so the compute engines and the interconnects can all be lined up together. And that includes PCI-Express switch ASICs as well, which have traditionally lagged pretty far behind the PCI-Express specs for the 3.0, 4.0, and 5.0 generations that they were widely available.

The lag between PCI-Express ports and PCI-Express switches at any given generation are a problem. That delay forces system architects to choose between composability (which ideally uses PCI-Express switches at the pod level) or bandwidth (which is provided through a direct server slot). Systems and clusters need to be designed with both composability and bandwidth and we would add high radix to the mix as well.

At the moment, there are only two makers of PCI-Express switches, Broadcom (through its PLX Technologies acquisition a number of years ago) and MicroChip. We profiled the MicroChip Switchtec ASICs at the PCI-Express 5.0 level way back in February 2021, which scale from 28 to 100 lanes and from 16 to 52 ports, but as far as we know, they are not shipping in volume. Broadcom unveiled its PCI-Express 5.0 chip portfolio back in February 2022, including the ExpressFabric PEX 89100 switch, which has from 24 to 144 lanes and from 24 to 72 ports. We are confirming if these are shipping as we go to press and have not heard back yet from Broadcom.

Our point is that PCI-Express switches have to be available at the same time that the compute servers, memory servers, and storage servers are all going to be created using chips that support any given level of PCI-Express. On Day One, in fact. You have to be able to embed switches in the servers and not lose bandwidth or ports or sacrifice radix to get bandwidth. We therefore need lots of suppliers in case one of them slips. This is one of the reasons why we were trying to encourage Rambus to get into the PCI-Express switch ASIC racket recently.

All of this is top of mind just as the PCI-SIG has put out the 0.3 release of the PCI-Express 7.0 spec.

Lets take a look at the projections we did for the PCI-Express roadmap a year ago when the PCI-Express 6.0 spec was wrapped up and PCI-Express 7.0 appeared on the horizon:

The PCI-Express 7.0 spec is not expected to be ratified until 2025, and that means we wont see it appearing in systems until late 2026 or early 2027. We think this wait is far too long. We need PCI-Express 7.0 to provide the kind of bandwidth accelerators need to chew on an enormous amount of data that is required to run a simulation or train an AI model. We need it matched up with a fully complex CXL 4.0 specification for shared and pooled memory.

We understand that it would be hard to accelerate PCI-Express 7.0 controllers and switches to market, and that all manner of products would also have to be accelerated. Compute engine and peripheral makers alike would be hesitant to not try to squeeze as much investment as possible out of their PCI-Express 6.0 product cycles.

Still, as PCI-Express 6.0 is put into products and goes through its rigorous testing which will be needed because of the new PAM-4 signaling and FLIT low-latency encoding that it makes use of we think the industry should start accelerating and match up to the CPU and GPU roadmaps as best as possible and to get onto a two-year cadence alongside of them.

Get the components in balance and then move ahead all at once, together.

Go here to see the original:
PCI-Express Must Match The Cadence Of Compute Engines And Networks - The Next Platform