With each new generation graphics are pushed further, delivering more power. We’ll typically we see a ~30% increase per generation as well as new features (such as tessellation) thrown into the mix. Increased processing power in turn requires building new graphics architectures with more of a chip’s basic constituent parts: transistors. Transistors form the building blocks of both Graphic Processing Units (GPUs) and CPUs alike; the long and short of it is that more transistors equals more ‘oomph’.
Nvidia and AMD (nee ATI) use the same manufacturer for their chips: Taiwan Semiconductor Manufacturing Corp (TSMC). The largest single chip ‘foundry’ in the world, TSMC traditionally performs a magical 'process shrink’ once every 18 months – taking each of the billions of transistors in a modern GPU and making them 20-30% smaller. This secret sauce keeps Nvidia and AMD (as well as Intel and Samsung etc, whose manufacturing arms do exactly the same) in the game of producing more powerful pieces of hardware with each generation, without requiring rolling blackouts across every city in the world. As the designers add more transistors, TSMC’s ‘process shrink’ keeps power draw, heat and production costs in check – without their ingenuity, adding more ‘oomph’ would be near impossible. However, in late 2009 something went desperately wrong at TSMC’s factories. In a ‘more wrong than In a ‘more wrong than SOPA and PIPA combined’ sort of way.
The graphics generation we saw in 2009, flagshipped by the HD 5870 and GTX 480 were released as planned, however due to these issues TSMC ended up producing a bucketload of faulty chips. The consumer never saw this directly; the broken GPUs were simply thrown away, however the failures pushed up prices and reduced availability. Even more disturbingly for AMD and Nvidia the next planned ‘process shrink’ to 32 nanometres didn’t seem to be working at all. TSMC responded by diverting their resources to fixing the myriad faults with their manufacturing – the scale of these issues still aren’t openly known, but it’s rumoured that as much as 90% of these chips were simply binned in the early days.
The chipmaker also completely cancelled the broken 32 nanometre (32nm) process in favour of a 28nm shrink that has ended up being delivered more than two years after the fact. At this point, things were a mess, and with their manufacturer’s stalling, so Nvidia and AMD’s plans followed suit – the follow on generation’s cards’ (GTX 580 and HD 6970 among them) transistors weren’t shrunk via the march of human ingenuity and so to avoid burning out power supplies they were redesigned with less transistors, running considerably more slowly than was originally planned.
It’s may not initially be obvious to every onlooker how significant the scale of the issue was; on the face of it (not) going from 40nm to 32nm may not seem like much of a problem, given 40 and 32 are reasonably proximate numbers. But the proof of the pudding is in the redesigning that occurred. Consider this: if the HD 6xx0 and GTX 5x0 had been made on a 32nm process, the silicon at their hearts would likely have been designed with 50% more transistors, each with a healthy MHz speed bump to boot.
All this would have occurred with reduced power draw. The HD 6870 would have been as fast as the 6970, the GTX 560 Ti as fast as the GTX 580, everything would have been scaled up, all without additional expense, heat or power draw. Shrinking transistors effectively single-handedly kept graphics chips from reaching insane proportions, for example the ‘Southern Islands’ chip in the modern HD 7970 is the exact same size (352mm^2) as the R580+ chip in the 2006 vintage X1950 XTX, inspite of containing over eleven times as many transistors:
In the world of microchips size really does matter, and the lack of movement hamstrung progress right up until the January 9 release of the AMD Radeon HD 7970, sitting in the Atomic labs right now. At this point, GPU manufacturing technology is getting on for three years old, approximately lemon-party age in computing terms. Where the rubber hits the road for us is right now – TSMC skipped but also in effect leapfrogged beyond its borked 32nm process, which has widened the gulf between previous and next gen GPUs. The HD 7970 and its yet to be released stable mates are jumping straight to 28nm; this literally halves the size of each transistor, in theory allowing manufacturers to simply double the amount of processing power in a card (which is exactly what one appears to have done).
There are other factors in play that will make this generation as big an issue for geeks as finding their downloading habits have cost society more than the entire planet is worth. For one, graphics chips are designed around common sets of features, predominantly Microsoft DirectX (DX) along with the much less frequently used OpenGL. As these features are added in DirectX, graphics card transistors are in a broad sense used-up without increasing performance, instead using the extra power running said features (like tessellation). This generation however sees no such sacrifices; while we’ve moved from DX11 up to DX11.1 the newest DirectX is concentrated on increasing speed rather than adding features. All up, then, with the potential to double transistors, increase clockspeeds dramatically, and nothing much holding the designers back it’s more than likely we’ll be seeing the largest graphics speed increase ever over 2012.