Last week on Bit-Tech, editor Tim Smalley wrote an insightful piece discussing whether or not AMD might see fit to move all ATI GPUs to SOI at 32nm. Whilst Tim brought up some great points, some seemed to clash with things we'd heard from the industry, so we thought we'd follow up.
As Tim correctly writes in his piece, integrating the GPU onto the CPU die is of paramount priority to AMD, and has been ever since the firm bought ATI. Also, as any chipmaker worth its ketchup knows, process tech is king when it comes to the semiconductor industry, and AMD will want to ensure it is using cutting edge stuff to stay ahead of Nvidia and not too far behind Intel.
Tim accurately predicts that when AMD does indeed take the plunge and build GPUs onto chips, it will most likely do so with the same 'Silicon on Insulator' (SOI) process tech CPUs are churned out on.
But following on from this statement, Tim goes on to say that it would be unlikely for AMD to migrate away from SOI "any time soon- at least, not without a completely new CPU architecture."
While this may be true for the current architecture, the future of APUs (accelerated processing units that integrate CPUs and GPUs) could realistically be both bulk and SOI. For example, high-performance SOI might make sense for products like AMD's upcoming Llano, but when the firm eventually brings low-power cores to market - *cough* see the here again-gone again "Bobcat" - core bulk looks to be a much more viable and cost effective option.
There is no doubt that, as Tim says, moving to SOI "requires a completely different mindset" and the decision does need to be taken early(ish) in the design phase, as one can't design a piece of silicon for a bulk process only to move it to SOI at the last minute. But just how early on that decision needs to be made is the issue.
From what we've heard in our discussions with GlobalFoundries, the decision to go bulk or SOI is made approximately six months before product design begins (a 12-18 month process in it's own right).
Performance, power consumption, die size and cost all weigh heavily in this decision, so there are several factors to take into account.
Bit-Tech is correct in its assertion that, in the case of high end GPUs filtering down to more mainstream ones, it makes little sense to completely re-architect the chip from scratch when the blood, sweat and tears has already gone into it. And yes, it would also be very expensive, so it would definitely be surprising if AMD moved its whole family of GPUs to an SOI process upon Fusion's release. So this is most probably true for discrete GPUs.
But Tim goes on to say that "there are no longer separate GPU and CPU design teams" at AMD, something which is (in our humble opinion) incorrect, as we have been told there are indeed separate design teams for discrete graphics and integrated design teams for APUs.
Tim also predicts that if AMD adopts SOI for its GPUs, punters will likely see higher clock speeds than the ones ticking away on AMD's current GPUs offerings, but from what we've been told, SOI doesn't necessarily mean higher frequencies.
Historically, Intel CPUs often had a frequency advantage over AMD's but suffered greatly in performance per watt prior to the Core 2 architecture.
SOI is marginally more expensive than bulk, but ultimately the real promise for GPUs would be in big potential gains in the performance per watt arena, so the price, in this case, might be right.
That said, outside of integrated CPU/GPUs, its doubtful we'll be seeing much change. Globalfoundries and TSMC are currently slugging it out to win the battle for GPUs at 28nm nodes using a bulk CMOS process.
Technologies like high-k metal gate and immersion lithography - both of which TSMC will only BEGIN introducing at 28nm, while GloFo will be miles ahead on third gen immersion and second gen High-k - will likely be the deciding factors in the war for the best tech recipe for graphics supremacy.
Tim concludes that AMD moving its GPUs to a different process technology could leapfrog NVIDIA, if it manages to find the right balance between performance, yield, power consumption and profits.
This is ultimately where Globalfoundries may prove to be the big differentiator for ATI GPUs. If GloFo manages to bring massive dies like Istanbul to market with high yields on an advanced process, it will probably bode very well for GPUs in the DX 11 generation and beyond.