- Wednesday, 10 February 2010
- INQ Mobile
- RSS
There's really no way to fix breaking news - Chicago Savant
Cell is no longer HPC material
AS THE NEWS BROKE at Heise.de, you could almost feel the Cell collective - that is, the Sony PS3 developer community and gamers - reel in shock at the sharp, jagged bits of an interview with IBM's Deep Computing VP, David Turek, saying Cell was to be no more. Of course, since he is the VP of HPC at IBM he was just talking about HPC rather than everything else.
According to the IBM executive's crystal ball, Cell is now no longer the right platform on which to develop HPC computing and so IBM will be shifting its focus from Cell-based co-processing to OpenCL-based co-processing - AMD's GPU stuff, in not so many words. This means that while Cell served its purpose in proving parallel processing was the way to go, development costs of further Cell based products become pointless as GPGPU computing becomes more widespread. Considering AMD is one of IBM's closest research partners this hardly comes as a surprise.
We rang up IBM PR but, rather unsurprisingly, its message queue was full. So we'll give you our view on the matter.
So far, IBM's latest supercomputing solutions, such as Roadrunner, revolved around AMD's Opteron processors with Cell as a co-processor unit. The latter, a cell unit dubbed PowerXCell 8i, is capable of delivering massive FPU performance, greater than that of AMD's x86 CPUs by a factor of 100. What IBM wants now is to move this model to CPU plus GPGPU by signing on to the OpenCL express as another brand of parallel computing to accomplish such tasks as nuclear weapons simulations.
There are two wide-ranging implications in IBM's new stance. First, there will be no further development on the Cell architecture on IBM's HPC side. There will be further solutions and support from IBM using Cell, and other products in other lines with the current Cell chips. It's good, it works, so why not? What Turek has said is that the future is GPGPU either with APU solutions, as AMD calls them, or with CPUs closely coupled to discrete GPGPU coprocessing units - essentially what AMD's been going on about for the past year.
Second, IBM's Deep Computing division is precisely the real estate that Nvidia is looking to conquer. Turek had previously stated that the future of supercomputing will be a 'hybrid' solution - simple enough, as x86 plus GPGPU computing is hybrid enough. However, siding with OpenCL effectively puts Nvidia in a pinch. This is a first class IBM and AMD tag team against Nvidia.
Notwithstanding all of this, the future of the Cell processor is uncertain. It hasn't made a hit with consumer electronics devices as Toshiba and Sony had promised, and the encroachment of GPGPU processing definitely throws a spanner in the works. PS3 will live out its usefulness, but in all likelihood, and with the path taken by CPU and GPU developers, the PS4 will be a far cry from its predecessor. µ
Share this:
ShareBusiness Intelligence Software
Customer Relationship Management (CRM)
Enterprise Accounting Software
Enterprise Resource Planning (ERP)
Enterprise Systems Management
The potential for this is massive if the proper thinking and execution is allowed to play out. Read that no egos just scientists doing what they do best.
Oh and by the way, I am so very, very glad someone ( IBM) teams with AMD against NVidia for competition.
¨Gentlemen,Start your engines.¨
they charged too flipping much!
I love to have the Cell in my TV. I love the cell on my smartphone. Who wouldn't?! But they charge the damn thin too much!
I can understand that it will be expensive for a system that have hundreds or thousands of Cell. Not for any consumer system that only have 1 or 2 Cell inside!
This is another marketing failure from IBM. They did it once for OS/2, now they did it again for Cell...
OpenCL is supported by Nvidia
You can program the cell with OpenCL.
"OpenCL-based co-processing - AMD's GPU stuff"
What? Givew credit where credit is due. When it came to development of openCL with Apple AMD/ATI had very little to do with it. openCL was modeled around CUDA and it's chief developers were Apple and Nvidia. The author of this article CLEARLY has no sniff about openCL and it's development. IBM and AMD are simply now saying what nvidia has been saying for years, HPC computings future is on GPU solutions where you want the most bang for the buck. Also openCL is not tied to GPU only solutions but can also be used with processors and DSP's. At least give the proper companies credit, Apple and Nvidia for bringing us openCL.
@Deanjo
OpenCL was developed by Apple alright, but with colaboration of AMD, IBM, Intel AND Nvidia. They are not so big partner of the project, less than the others, they only want to make sure that they have a compatible product with the open standard, as it'd be bad to nVidia to have openCL AMD only.
Their focus is on CUDA only, as if anyone in HPC business with sane mind would strap himself with a nVidia only product.
Check your facts fanboi. Sure Nvidia was involved with developing OpenCL, but it was pushing its own agenda on CUDA. AMD has been equally behind OpenCL from the get-go. Why single out Apple and Nvidia? AMD and Intel were both there, as too were many other companies who will never get recognition.
Apple was the first to integrate Open CL into something on your desktop: OS X. I don't think that actually comes close to the level of computing OpenCL is supposed to bring about in HPC.
Nvidia wanted CUDA to be the next big thing. They made it proprietary, royalty-free sure, but run only on Nv hardware. So nice of them... eh?
IBM also said they were planning on GPGPU solutions based on integrated integrated CPU+GPU computing. Unless Nvidia has pulled a CPU from out of thin air, where the heck did you conclude Nvidia is somehow a part of IBM's plan?
Maybe the phrasing isn't the best, granted.
@ Jean:
Check your facts there.
"OpenCL was initially developed by Apple Inc., which holds trademark rights, and refined into an initial proposal in collaboration with technical teams at AMD, IBM, Intel, and Nvidia. Apple submitted this initial proposal to the Khronos Group. On June 16, 2008 the Khronos Compute Working Group was formed[1] with representatives from CPU, GPU, embedded-processor, and software companies. This group worked for five months to finish the technical details of the specification for OpenCL 1.0 by November 18, 2008.[2] This technical specification was reviewed by the Khronos members and approved for public release on December 8, 2008.[3]"
http://en.wikipedia.org/wiki/OpenCL
Originally designed by Apple, who then created aconsortium of interested parties to refine it into a neutral, agreed standard that Khronos could endorse.
Using the fake penis experts from Wackypedia for your supporting evidence is one sure sign of ignorance.
Be that as it may be, however, I don't see where JHH and the rest of the lying, scabulous criminal organization known as nVidia should get any more credit for OpenCL than the other "neutralizing" partners.
Oh, and Roger in Spain, IBM doesn't knee-jerk. I've been around their operations long enough (as a contractor) to understand the deliberation processes. A Senior VP doesn't talk on the record unless the topic has been well-vetted internally.
nVidiots should just give up and go away. With IBM partnering with their, well, partners at AMD (pretty obvious, given their close working relationship), JHH and his fellow criminals stand about as much chance in the HPC market as a monkey with an abacus.
Organizations looking for HPC solutions aren't interested in talking with a mere chip designer. They talk with solution providers.
As has been pointed out, the author's association of OpenCL with AMD is off base. But to me, the real misdirection of this article is the un-qualified assumption that OpenCL is the next big thing in GPGPU. In OpenCL and CUDA (both very low-level APIs), a lot of code must be written to move and manipulate data structures between the host and device (and back). The real breakthrough will be when compilers start supporting GPUs and do this for you. PGI already has added this to their C and Fortran compilers using OpenMP-style directives and it won't be long until Intel, GNU and the rest follow suit. Why spend your time writing in a language like OpenCL (which may be portable at a library-level, but never at a performance-level) when you can just recompile your existing code?