Hasn't nvidia secured a deal to produce a super computer using Fermi, that is 10 times faster than the current leading Jaguar (powered by AMD Opteron) and is being hailed as one of the most significant leaps in HPC progress?
Ten years is a long time. I don't see their competitors relaxed about it.
Universities and companies want a mid "cheap" supercomputer as the modern desktop can't cope with a size of dataset they want to process.
Mark Harris has been working with universities to mature the GPGPU market. The result is a myriad of university projects and now commercial research to use GPGPU to process medical scanner data, oil exploration data etc.
This hasn't been one sided - but first I want to break to look at the strategy in the toolsets.
GPGPU originally existed by loading your data into OpenGL textures (yes the same game textures), then programming the fragment shaders with your data processing program before rendering the texture to an off-screen buffer. Bingo we have computation.
Researchers started picking this up and coming up with ways to minimise the limitations when doing reductions using ping-pong etc and creating ways to maintain data efficiently. Brook, RapidMinds (acquired by Intel) were born to minimise the differences between nVidia and ATI.
Obviously it makes sense if you can control the software toolkit to adapt the platform to GPGPU - hence CUDA was born.
nVidia's approach is identical to Apple, a high cost closed platform with a "Boom it works" approach (we'll forget the early driver issues!).
CUDA became the language (much like Objective-C is to Apple). Nexus will become the defacto closed toolkit akin to Xcode on Apple with Objective-C/Cocoa (yes I know it runs GCC (thus apple get this for free) at the bottom but how many platforms actively use Objective-C instead of C/C++).
A faster, flexible but more costly toolchain to produce and maintain. It has to be free to start using, naturally they claw the cost back in the high price hardware (specifically the more expensive Tesla range is required).
ATI follow a more Linux open approach, so they produced Close to the Metal (CTM). Which basically allowed you to hand code GPU assembler. Open but so low level that nobody really wanted to touch it (switch GPU and prepare to recode to optimise!).
I was a registered CTM developer and looked at combining GCC and CTM however the GCC chain is so basic and olde-worlde it's not suited to SPMD programming as all the analysis and internals work for small SIMD optimisations.
Anyway I digress. The key here is that there is no GCC standard toolset currently to underpin OpenCL so there is a small lag as products appear..
Apple could see this being a useful technology for their media processing and started work on OpenCL.. eventually handing to Khronos.
AMD could see the value and dropped CTM.
A slower toolchain to produce as the standard becomes subject to design by committee. However it's cheaper and more likely to deliver a wider range of development environments for university and company budgets.
So now we have CUDA/Nexus vs OpenCL/<random toolkits> as the implementation platforms. Although nVidia support OpenCL I see it will become a second cousin to CUDA/Nexus.
So back to the part about being hailed as a major leap. It's a leap forward but not solely nVidia. SGI were doing this before nVidia existed (Mark Harris, IIRC was part of SGI).
Next question that needs to be asked - what's the market size (financially).
Universities by their nature are interested in low cost - is Fermi low cost when compared to the expected OpenCL compliant AMD/Intel platforms?
When AMD/Intel focus on the remaining discrete GPU market, then nVidia could find it's cash source squeezed. The same will occur in the commercial GPGPU market as there's money to be made in this blue ocean market.
10 years is a long time for nVidia's competitors to come up with competitive products in the GPGPU space. In this time it has to has to safeguard it's share - this is where the proprietary toolkit comes in, making it too costly for commercial deployments and applications to switch vendors.
This is the reason that nVidia don't like a level playing field such as DX and OpenCL.
For larger installations still - super computers - the entire platform is proprietary and as long as the development toolkit delivers then it's an easy life for upgrades and expansion.
The thorn is that the majority of code is C/C++ and x86 SIMD orientated. Again I can see the move to supporting C++/CUDA here. The requirement for logic operations requires nVidia's CPU program to bare fruit, although the majority of processing is multiply-and-add (the key GPU instruction). Their weakness is a lack of experience with interconnect performance between GPU nodes. They may opt, if sensible, to partner with CRAY in the mid-term.
AMD have some experience in this space and I would expect CRAY to partner in the open alliance to create supercomputers using AMD CPU, GPU technology as they have done in the past...
So although it's heralded as a "significant leap" it should be noted as a leap of market potential however it is a market which boarders the territory of some big companies, who may not be first to exploit the market but when they arrive will be quickly applying pressure by stretching their product portfolios down and leveraging their knowledge of the supercomputer market.
In this scenario I would be careful of an initial "success" blip for nVidia as they see sales rise in the fresh market, only to fall as the major competitors arrive. They do have a lock-in device but it's the market cashflow that will really see them succeed or fail, and I think 10 years is an awefully long time in a tank without partnering with some sharks.
