THE NEVERENDING Intel vs AMD scuffle continues unabated with competition, legal tussles and so on. Right now AMD is knocked down in the CPU boxing match, and recuperating in the corner until their 45nm parts come out, hopefully bringing some hope. They'll need that across the board, from ultramobile to supercomputers.

Another conflict has seemingly been brewing slowly - first fairly silent, now plain overt. Yeah, Intel and Nvidia surely don't (any more) love each other. The reasons are quite a few, the newest one being - Larrabee.

Graphics

Up to now, despite the potential Intel-Nvidia acquisition made all but impossible these days, there was quite a bit of benefit for Nvidia from the recent rapid Intel processor performance advance. The combination of Core 2 Quad and SLI-enabled "speed demon" capability saw Nvidia sell quite a few overpriced high end Nforce chipsets to run one, two or more equally overpriced 8800-series graphics cards, thus sharing a big chunk of the "added value" in the Core 2-generation desktops, laptops and workstations.

In the meantime, no thanks to Nvidia's (real or perceived) development cycle slowdown, and continued high prices, DAAMIT picked itself up on the GPU side, and, combined with ever-improving Intel chipsets and their Crossfire capability, managed to get back into biz: a quad-core Intel QX9770 on Intel X48 chipset with two Crossfired 3870X2 cards is just as capable, but cheaper than, the same QX9770 on the newest Nforce 790i chipset and twin GeForce 9800GX2 cards in quad-GPU SLI setup.

The expected accelerated arrival of the RV770 and other next-generation DAAMIT GPUs late this year, will give further headaches to Nvidia which for now only has a 55nm G92 shrink coming up soon. Also, ATI's GPU newbies may somehow see their launch coinciding with the first Gainestown and Bloomfield Nehalem CPUs, whose Tylersburg chipset is, guess what, again CrossFire-ready.

The start of next year will, with the arrival of Larrabee, see both Intel and AMD - or whatever it morphs into by then - having the complete CPU, chipset platform and GPU stacks across all major market segments: mobile, desktop and GPGPU workstation. Nvidia will, of course, miss the first, by far the most important ingredient - the processors.

However, first Nvidia may be hit there where it hurts most - their flagship GPU business. The first Larrabees are expected to be far more flexible in the usage models, whether it is a DX10 gaming GPU, an OpenGL workstation engine, or a multifunctional GPGPU.

Couple it with a possible direct QPI connection option on top of the usual PCI-E, and Larrabee may end up as a tightly coupled graphics coprocessor, able to access all of the Nehalem's large system memory pool very fast, on top of its local graphics memory. Better still, having an X86 front end, Larrabee might be programmed inline as an X87-style coprocessor - remember those days?

Nvidia, of course, still has a stronghold - but not a stranglehold - among many game developers, with the long-perfected relationship over the years. Does it mean they can impose their model over the Intel one? Not really, those same game vendors have to increasingly support DAAMIT GPUs too these days, as the HD3870 seems to have the price-performance edge, especially with their new steppings.

If Intel sees the strategic value of the developer relationship on both gaming and workstation software front - besides of course getting both DirectX and OpenGL fully optimised - they will open up their fat checkbook as required and get that support to the hilt.

Couple this with the ability to - akin to the mainboard chipset business - encourage major vendors to design their truly own unique graphics cards on time for launch by working with them early in the GPU design stage, compared to NV and ATI making those same vendors stuck with their own board designs, and you g et the picture. That added value can translate into far higher margins for those Taiwanese vendors - which approach will they support, guess?

Chipsets

With the advent of X38 and now X48, Nvidia chipsets have mostly lost the performance advantage they enjoyed in the Nforce 680i vs Intel 975X days. The newest 790i doesn't justify being, say, twice the price of the X48: everything from memory, I/O and graphics capability or overclocking is pretty much equal, with the Intel offering running all that somewhat cooler due to the superior process.

Nvidia was given the Xeon license by Intel to enable creating dual-FSB high-end workstation and gaming chipsets which by right should have been the base of something like Skulltrail. Intel's existing Seaburg 5400 chipset, with those FB-DIMMs, isn't exactly ideal - a dual-FSB flavour of a Nforce 790i Ultra would be superb for the purpose. However, for reasons unknown, whether design, available process or a simple cancellation, this never materialised.

With Nehalem's on-die memory controllers, and high-performance Tylersburg chipset, there is far less space for Nvidia to cover - and make extra dosh from - once the new platform arrives this year. And that is, even assuming Nvidians get the QPI license.

License stand-off

That's where we touch the sensitive spot: the rumours circulate - bear in mind, just rumours - that Nvidia might not get the QPI license after all. It didn't deliver on the Xeon side, and its value add is far less in the Nehalem era: so, why give it to them at all?

Well, QPI block would cut Nvidia off ALL future Intel chipset business, and, quite possibly, future high-end tightly-coupled QPI-based GPU market where this connection may enable important performance wins.

If Nvidia, as some other rumours say, tried to actively lobby the abovementioned software developers not to support Larrabee, Intel would be unhappy.

What happens when, at the same time, that same Nvidia lobbies support around Taiwan for a hostile takeover of AMD due to quadruple market cap advantage - and all that despite US Govt's clear "no" as JenHsun may never pass their security clearance for AMD Uncle Sam military deals that Hector & Co might be living off right now?

And now, with that obviously a goner, Nvidia may be trying the same with VIA - not for their chipsets, for sure, or for the performance of their CPUs, but for the X86 license, pure and simple. If that deal was to go through, everyone aside of CPU team may as well look for new jobs.

Faced with this blatant attempt to forcibly build up direct competition across the board, Intel might end up sufficiently incensed to give them some serious spanking - a favourite Singapore pastime.

QPI license hold-off is just a start - if the VIA X86 CPU license is anything similar to the AMD one, Intel could attempt to withdraw that one as well.

Now, other newer rumours say that Nvidia is trying to threaten Intel with possible GPU-related lawsuits to block Larrabee take-off. That by itself may be problematic since Intel did get quite a hold of graphics patents too, not to mention some nice people from companies like 3DLabs who knew how to do high-end 3-D equal, if not better than, NV.

On top of possible Chipzilla IP counter-attack in this case, the Graphzilla could also face Intel's deeper - temporary at least - support for the ATI RV700 GPU family as a "complementary" solution to Larrabee. It's only a step further from the current in-built Crossfire preference, anyway. And, it helps Intel keep its main competitor "afloat" to stave off any monopoly investigations.

However, why not try the "middle path"? After all, if Nvidia doesn't make unnecessary trouble, Intel could extend them the QPI license - maybe for chipsets only, not for GPUs. It's not as if Nvidians may make full use of it: they just missed essentially dominating the high-end workstation and gaming market if they did the above mentioned dual-FSB Xeon Nforce chipset. That might even be positive for Nehalem platform-level diversity.

In summary, right now the two Santa Clara-based PC chip giants seem to be on a kind of collision course. In my mind, the collision is not necessary - but the smaller party has to be more nimble and flexible to avoid getting trampled on, in anger or otherwise. µ