Catching up with Cray
Cray is a legendary name in supercomputing hardware. Cray CTO Bill Blake (Netezza’s early-rise VP Development) seem to be there in part because of Cray’s name and history. I’m now consulting to Cray largely because of Bill Blake, specifically to Cray subsidiary Yarcdata. Along the way, I’ve picked up enough about Cray in general — largely from Bill and from Cray president Pete Ungaro — to perhaps be worth splitting out as a separate post.
Cray business highlights include:
- After a meandering and financially disappointing journey, Cray is again a stand-alone public company.
- Cray is a computer systems company.
- Cray makes a large fraction of its revenue from selling and supporting a small number of supercomputers, largely to scientific, technical, and government customers.
- Even so, Cray sells systems at a broad range of price points. Storage products are in the mix as well.
I haven’t sorted through all the details in Cray’s SEC filings, but huge government contracts play a big role, as do the associated revenue recognition delays.
At the highest level, Cray’s technical story looks like:
- The paradigm of a Cray system is:
- Lots of processing power
- Lots of memory
- Lots of threads to use all that silicon
- Very fast context-switching among the threads
- Everything looking like a near-seamless whole
- Cray is embarking on its last multi-year generation of computers that rely on proprietary parts, specifically in the area of interconnect; more precisely, Cray is embarking on it last multi-year generation of interconnect chip technology. In fact, …
- … Cray has already sold its interconnect hardware business to Intel, for integration into its chips or packages, and is going to be licensing it back (exclusively) over the next years.
The core Cray use case is to model the world with the best possible fidelity. Canonical Cray applications include:
- Weather forecasting.
- Nuclear explosion simulation.
- Aircraft modeling. (I came up with the bolded formulation above after Pete told of an aircraft engine maker who could simulate a whole turbine on Cray, as opposed to just modeling a single fan.)
Graph-oriented systems such as terrorist networks are also now in the mix, which is how Yarcdata got started.
Do an image search on “Cray”, and you’ll find multiple generations of circular computer. Why circular? To minimize the diameter, of course. Why minimize the diameter? So that signals would take the fewest possible clock cycles to get from one part of the system to another, allowing the maximum number of computer parts to work together in a semi-seamless whole.
The Cray-1 ran at 80 megahertz. The distance light can travel in 12.5 nanoseconds is on the same order as the diameter of that machine.
I mention that because, while the geometry has now changed, similar priorities guide Cray’s system design to this day. Cray’s fundamental design philosophy could be called scaling with minimal compromise; Cray wants to stick together huge numbers of computer parts and have them work with as close to linear scalability as ever possible.
Truth be told, I don’t know how close to linear scalability Cray achieves, and I haven’t even heard them frame the issue that way. Even so, I’m pretty confident my formulation is accurate.
As Pete puts it, Cray’s development focus is computer architecture, and especially three specific areas:
- Interconnect.
- Operating systems and programming tools.
- System packaging (e.g. energy efficiency).
The specifics in that that keep coming up when I talk with Yarcdata are:
- Many simultaneous threads.
- Many servers’ worth of RAM addressable in a single pool. (Especially this one.)
Thus, Cray’s custom interconnect technology is a big part of the story. So when I read something that sounds like Cray sold its interconnect technology to Intel, confusion set in.
The story turns out to go like this:
- Cray’s current interconnect ASIC, Gemini, is approaching the end of its life span.
- Cray’s next-generation interconnect chip, Aries, is slated for “serious use” in 2013.
- Cray believes that for both power and performance reasons, processor and networking chips need to be(come) closely tied together, in the same package and perhaps even on the same chip. Intel believes the same thing, and/or generally want to conquer the silicon world.
- So Cray sold its interconnect hardware/chip technology to Intel, including engineers, patents, and so on.
- Nonetheless, Cray will have exclusive right to Aries.
- Post-Aries, Cray will use standard Intel interconnect technology (built with the help of what Cray bought from Intel), and compete in the interconnect area on software.
Cray notes that the software interconnect engineers it is keeping are more numerous than the hardware interconnect engineers being transitioned to Intel.
I guess if Cray hadn’t done the Intel deal, the next generation would have been called Pisces … but that name sounds somewhat fishy.
And with that, I’ll turn to a discussion of Cray’s data-oriented subsidiary Yarcdata, wherein are mentioned figures such as “1 million threads” and “half a petabyte of RAM”.
Comments
One Response to “Catching up with Cray”
Leave a Reply
[…] Cray’s strategy these days seems to be: […]