'New clock, not just tick or tock'
"The industry needs a new clock, not just another tick or tock," Freund said with a laugh in an interview with El Reg at SC10. Whatever Calxeda is working on specifically, Freund will not say. But he was willing to confirm that Calxeda is working on ARM processors specifically for servers, and that the company will not be launching servers itself but rather working with OEM partners to put its chips inside of their boxes.
Freund was also willing to talk about the design goals that Calxeda has set for itself, making it clear that these goals did not constitute a specific processor roadmap as we are accustomed to seeing from Intel, AMD, IBM, and Oracle. The "server on a chip" design that Calxeda is cooking up will offer the same performance at the system level (with more chips per system) as an equivalent x64-based machine, with the resulting ARM server being half the cost, using one-tenth the energy, and occupying one-tenth the space.
"The baggage of compatibility takes up too much real estate on the chip," says Freund in talking about x64 chips from both Intel and AMD. Which is why chip startup Tilera is making system-on-a-chip designs based (everyone conjectures) on the MIPS architecture that can pack 64 cores and everything it needs onto a single chip and is working on future variants that will do 100 and then 200 cores. (See here for details on the latest Tilera chips, which offer very good integer performance but which do not have floating point math.) Tilera's future "Stratton" chips will allow server makers to cram up to 40,000 cores into a server rack by 2013, if all goes well; half that will be possible by the end of 2011 using the 100-core TileGx-100 SoCs.
So why come into this dense, low-power server racket when Tilera has such a big lead and not very much business as yet? (Not enough to dent Intel, at least.)
"We think the ARM chips have the ecosystem of developers and tools that the MIPS chips do not," says Freund. And more importantly, ARM chips have the force of history on their side.
Setting aside the issues of memory addressing for the moment, which are an issue for both Tilera and for anyone making ARM processors for servers, owing to the 32-bitness of the designs, Calxeda and Marvell agree that there is a place for 32-bit servers in the data centers (and maybe the living rooms) of the world. (The future "Eagle" Cortex-A15 reference chips from ARM Holdings, the chips have a funky virtualized 40-bit mode and do not go all the way up to 64 bits.) Calxeda knows that it is going to take some time for ARM to go all the way to 64 bits and is nonetheless convinced that it can make creditable server chips using the ARM architecture.
For a lot of hyperscale data center customers, power draw is as important an issue as processor throughput, and for workloads that can be parallelized on tiny machines, a swarm of low-powered ARM chips with a clean, lean, and mean implementation of Linux might be a better answer than running Linux on x64 chips. "We believe there are plenty of workloads that can work fine with 32-bit processors," Freund says.
This will not be the first time that hoards of cheap machines have assaulted the data center and knocked an incumbent off its tiles excepting the most pesky legacy applications. Mainframes owned the data center for a decade and a half until proprietary minicomputers surrounded them, attacking from the bottom and eventually pulling applications off the mainframe and, more importantly, getting new applications that would have otherwise gone onto the mainframe. Unix-based RISC servers came in, again growing up from workstations, to attack the bottom of the mainframe and minicomputer markets and growing to respresent half of the server racket at the height of the dot-com boom. And then the x86 desktop chip grew up in the 1990s to become the dominant server platform, chewing into the bottom of the RISC/Unix and proprietary mini base and expanding to essentially take over.
Now, it will be the turn of the ARM-based cell phone chip, which Calxeda is going to help grow into a bottom dwelling shark that eats its way into the heart of the server space.
"It has always been a lot easier to scale up a new technology up to take on an established one than it has been to scale that established technology down to meet a new threat," explains Freund, and considering that he was the chief cheerleader for IBM RISC/Unix and mainframe platforms for the past decade, that's a stunning admission. ®
Tilera's architecture isn't MIPS-based
Just a minor correction. Tilera's 64-core system-on-chip design isn't based on the MIPS architecture, it's their own. It has a VLIW 3-pipeline instruction set with DSP-like extensions. It comes with a gcc-based toolchain, and a port of SMP Linux supported by kernel 2.6.36.
Perhaps RISC will save us all. We will just have to wait and see if they can produce something better than the Cell proc.
So who's right then?
All those other articles saying that the Really Big Deal in servers these days is the amount of RAM you can cram in or some startup who reckons that they can blow the doors off the industry while limited to 32-bit addressing?
I'm not convinced that having enough in the whalesong budget to seriously piss off the silly name god makes 'em any more likely to be right. The fact that the rebranding jossstickery would appear to be the only item of substance in a two-page piece about their latest "annoucements" just adds to the aura of FAIL.