By now, it's clear that the very high end Nvidia Kepler parts aren't going to see the light until much later this year, while the GK104 definitely can't fight AMD GPUs for the performance crown. On top of that, Intel is eating away at the low end, where Sandy Bridge Xeons chip at profitable entry level Quadro business, and Ivy Bridge is to follow…
Nvidia, as a corporate entity, still suffers from certain heavy structural weaknesses – the lack of an X86 CPU cannot be compensated by ARM Tegra, as there are a dozen other ARM vendors, plus the hungry upcoming Chinese mobile CPU companies, which fight for the same market, some of them with far higher financial resources – and, in a few cases, own fabs – than Nvidia itself. Add to that a speculative story which I heard in China, that only a very large order of many thousands of Nvidia Tesla GPGPU cards for their Tianhe multi-petaflop supercomputer few years ago actually saved Nvidia from a rumoured Intel takeover attempt – not that I would recommend such thing to Intel for their own corporate health anyway – by artificially boosting their stock price at the time, and the situation becomes truly intriguing.
Having exploited their own past performance leads to the fullest, Nvidia knows well the benefits of the 'waterfall effect' where many users, including our own readers, watch who holds the current speed lead at a given time and, based on that, affect the buying decisions for the lower-cost parts towards the same vendors. Right now, AMD holds that lead with the HD7970 and, soon, the HD7990.
The GK100 'Kepler' and its GK110 follow on were supposed to wrest that lead back to Nvidia, just a month or two from now. However, right now it looks a gone case, at least for the next half year. For whatever reasons, the GK104 upper mid-range part, a follow on to the GeForce 560 family, is now the top-end part to launch, and it can't – I mean can't – win over the HD7970 in any meaningful use. That role was reserved for the GK100 and follow ons.
Either way, the lack of the premium 'crown holder' part is bound to affect Nvidia's image. Unlike AMD and its – problematic at times, too – CPU business, and with the chipset presence long gone, Nvidia has no other major PC foothold aside the GPUs. AMD can, at least, alternate between attractive CPU/APU and GPU launches to keep the market buzzing, even if some things go wrong. On the other hand, the 'fashion fad' tablets and smartphones, with zero CPU vendor loyalty to boot, will never fully replace the long term stability of the PC market.
On the other hand, as covered here before, Intel had – at truly minimal cost to themselves – suddenly started eroding Nvidia's key profit center and 'leadership refuge' of sorts, where they invested top dollar, and gained top profits in return: the Quadro OpenGL professional GPU family. Even though the Sandy Bridge Xeon E3 integrated GPUs can only target the entry level of Quadro line, the 'good enough yet certified performance' approach simply works well for polygon-based, zero-effects but large and profitable 3-D CAD market. The Ivy Bridge will, with further performance doubling and more app certifications, deepen the attack as well.
So, Nvidia has to address both the high end and low end GPU market problems this year. The performance leadership is lacking, and that will affect things down the product line too. On the high end, besides obviously pushing forward the true high end part if the circumstances allow, it should also embrace one good idea that AMD has, if they want more acceptance in mainstream compute GPGPU in the future – no deliberate crippling of double precision FP math functionality in high end consumer parts. AMD never curtailed the DP FP performance in their high end HD6900 and HD7900 lines, and gained more traction even in HPC because of that. After all, the HD7990 is a Teraflop-class double precision FP monster.
On the low end, cutting Quadro profits to keep the line survive might be a good long term idea, to justify the users to still bother about the products. Simply, even if the Intel OpenGL graphics in their processors is somewhat slower (and how much speed besides vertex processing you need anyway for typical polygon-based 3-D building or machine models), it comes in fully certified, supported, and – free. The company has invested so much over the years in this flagship product brand and the market segments it covers, it would be a pity to see it squeezed hard by Intel from the low end and AMD at the high end by the increasingly assertive FirePro OpenGL line too, with improved drivers and apps certifications now…