From the Peripheral Component Interconnect (PCI) to the Compute Express Link (CXL), Jim Pappas has spent most of his 30+ year career driving standards initiatives and promoting their acceptance. Currently the director of technology initiatives at Intel, Pappas chairs the CXL Consortium, co-chairs the InfiniBand Trade Association (IBTA), and is a board advisor of the Storage Network Industry Association (SNIA), among many other leadership activities past and present.
We talked with Jim recently to ask about his history in high tech, his thoughts on the current landscape, and what makes for a successful initiative. Our conversation also yielded some surprises—including Jim’s view that the work of creating technology initiatives sometimes produces magic.
HPC Voices: Jim, you’ve been involved in establishing computer industry standards since the 1990’s, first at Digital Equipment Corporation (DEC) and later at Intel. How did you get started in this work?
My early years (1980’s) in the industry were at DEC, where I was a CPU engineering manager and, before that, a graphics architect. I earned some patents, but I didn’t talk with anyone outside the company. Then I attended an industry trade conference, and realized I really wanted to be involved in the industry.
Later, someone I respected at DEC introduced me to an Intel project called the Local Glueless Bus. I met with Intel, and we decided to make it into an industry standard. We formed a small team and renamed it Peripheral Component Interface, or PCI, and soon I was one of the most industry-facing engineers making PCI happen. The PC was just taking off and everyone wanted an easy way to connect to it.
HPC Voices: Which initiatives have had the greatest impact?
PCI is certainly one, although one of my later initiatives, PCI Express (PCIe), eventually killed it. Certainly USB—the Universal Serial Bus. PCI fixed the inside of the box as far as connectivity, but then we needed to attach devices externally. I was recruited into Intel in 1994, and we formed a consortium and invented USB.
USB is a great example of how, with ecosystem involvement, you can get much more than you asked for. Devices were much less power efficient in those days, and we came up with the idea to simply supply some power to the attached devices. Without us even envisioning the impact, USB became the universal power supply, and now USB supplies power to almost everything. It’s the most widely used interface in the history of computing. It strengthened the PC and allowed it to become the center of our digital lives.
InfiniBand is another great example. Half the HPC clusters out there are probably using InfiniBand. Also, we can’t overlook the Ethernet specification, which Intel, DEC, and Xerox co-developed and which became the basis for the IEEE 802.3 Ethernet standard.
HPC Voices: What goes into a successful initiative?
Many times when we start, there might be confusion and multiple efforts, so you have to work through those, but generally, when you put a good initiative together, the confusion evaporates. It drives clarity and allows progress to move forward.
Starting out, you need more focus than a group of 300 companies can give you. Most successful initiatives start with a small number of influential companies to set the direction. You need a small, knowledgeable team to get things started and put together the overall plan. Then you invite in the innovation of the industry and get others involved to make sure it will meet their needs.
If you’re going to get an entire industry to work together, you need a clear compliance and interoperability plan, and you must keep focused on the end result. You need to make the standard truly open and compatible, and use it to drive clarity and progress. The guiding principle is to not stifle innovation and allow the industry to succeed.
At a certain point, it’s a leap of faith.
HPC Voices: CXL is an important current initiative. What is its significance?
CXL is cool. It is a top-tier initiative that’s going to bring a dramatic change to computer architecture. In many ways, CXL is an extension to PCI. Part of the brilliance of PCI Express was to make every generation forward and backward compatible. This greatly reduces the risk of product development, so HPC users get more innovation and more devices to choose from as they’re configuring their HPC and AI clusters.
CXL extends PCI Express in very powerful ways. Where PCIe is a block level interface, CXL lets you use the same wires and infrastructure of PCIe, but also do memory transactions over the interface. Consider a dual-processor system where each processor can read/write the other’s memory and also locally cache the data using an interface such as Intel UPI, the Ultra Path Interconnect. Furthermore, the cache in each processor remains coherent with the other, so if data is changed in one processor, the adjacent processor’s cache is also updated. CXL lets processors attach to other devices with similar levels of memory and cache coherence, so it represents a big step forward for acceleration. You can use CXL with FPGAs and ASICs to offload AI functions from the CPU with same close interface, for example.
CXL works by means of three protocols. CXL.IO basically replaces the PCIe I/O bus and does I/O operations. CXL.cache keeps the caches coherent. CXL.mem allows the remote memory to be addressable by the CPU as part of the same coherent memory space. The three protocols can be interleaved together on a transaction-by-transaction basis, so each transaction over CXL can dynamically use any of those three types of cycles.
The consortium delivered the first version of the specification (CXL 1.1) in December of 2019. In Q4 of 2020, we released CXL 2.0, which adds the ability for multiple CPUs to access multiple accelerators through a switch. The first CXL systems will be delivered 2022, and we’re working on version 3.0 of the spec. Any company wanting to help shape the spec can join us, as CXL is a completely open consortium.
HPC Voices: Are there areas where you see a need for more industry collaboration?
Artificial intelligence and machine learning are huge areas, and I expect to see more standards innovation around their issues. The Internet of Things is an area with tremendous opportunities and a number of companies trying to be the control point, but it’s probably still too fragmented. That’s a tough one.
OneAPI is an important mechanism to get us to the level of choice that our customers want in terms of a truly open, unified programming model that crosses accelerator architectures. It’s built with the right priorities to advance the industry, and I expect it to be very important going forward.
HPC Voices: Why have these types of initiatives been such a strong focus for Intel?
Market leaders have a responsibility to advance the entire industry, so the platform can evolve and stay current. Intel has enough influence to help move the ball down the field. We often look at the industry and analyze the following: where we can add value, what needs to change, what’s holding it back, how do we move it forward, how do we make a difference, and then how do we get the industry to rally around the effort.
The technologies my team drives are ones that no single company can do on its own, at least not to the same level of impact. So the work we do is to try to align the entire ecosystem, including our competitors, to move forward and spark innovation across the industry, and then pull that innovation together to advance what the platform can do. Some of the ideas that come out of an initiative are spectacular, some fall off a cliff, but it’s all part of the process. The goal is always to have a useful result. Specs that no one uses are useless.
Intel is involved with hundreds of standards organizations, some with only a few people, and others which are very large. We join to offer leadership. We try to put the right engineers in place, hear the industry issues, adapt and change accordingly, and have the discipline to get things done.
HPC Voices: What are the benefits to the companies that participate in industry initiatives?
The bottom line is what I said earlier about USB. When you pull the industry together and work toward this innovation, you get much more than your original vision of this new type of technology. The pie gets larger because our machines can do more, and you build momentum in the marketplace.
You can attempt to tackle the problem on your own, but if it takes three or four years to develop a new technology, then in the end, all you have is what you envisioned when the product development started. In such a dynamic industry, that can be a disadvantage, because in the meantime, the industry will have changed. Having the ecosystem involved helps you keep up with where the industry is going rather than where it was when you started.
When you drive one of these initiatives, you not only get everything you wanted, but everything you didn’t know you were going to need. The ecosystem delivers things you never dreamed of, and it helps evolve the industry. It’s really a magical experience.
HPC Voices: Any plans to retire?
No. I’m having too much fun. I take inspiration from Ruth Bader Ginsburg, who worked till she was 87.
Intel technologies may require enabled hardware, software or service activation.
© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries. Other names and brands may be claimed as the property of others.