With Nvidia and AMD both releasing a new generation of graphics cards over the next few months an increasing amount of information is flowing, and much of it pointing in the same direction. Then of course, we have our sources...
There are two things we have been completely certain of for a good while now. Firstly the increasing difficulty in shrinking silicon experienced by all chipmakers means TSMC’s 20 nanometer (nm) ‘process shrink’ will not debut for almost another year; both AMD and Nvidia’s next-gen will be made on the existing 28nm process like their predecessors. In addition both the Red and Green team are slated to release tweaked and enhanced versions of their 2012 architectures rather than all-out new designs. These facts do limit what they can do – both will be building physically larger pieces of silicon and just how large they can get depends on the efficiency of their designs – they have to keep within acceptable power envelopes.
Given this, it is worth a quick summary of the current architectures. With its current ‘Southern Islands’ cards AMD developed a largely new design: it was technically innovative and a first foray into a general computing design (e.g. video decoding and supercomputing applications). However the Red Team somewhat unwisely released it with a gargantuan amount of overclocking headroom - as we know, that unnecessarily dampened performance until faster ‘stock’ speeds were delivered with driver and firmware changes.
Nvidia took a different route with Kepler, its radically redesigned of the flawed 2010 vintage ‘Fermi’, for one reducing its general computing capbilities. Critically it also moved to a single clock speed for the entire piece of silicon with huge results for efficiency. The effect of the redesign has been impressive – Kepler uses less silicon, much less power and is incredibly fast - generally winning out against bigger, hotter AMD chips this generation.
As a result 2012 saw the Green Team taking the holy GPU trifecta of outright speed, performance-per-watt and low temperatures. However Nvidia only won each of these by a small margin; never truly outclassed, Southern Islands was kept very much in the game by the aforementioned speed bumps as well as continual price cuts (the HD 7950 almost halving in price since debut for example). So the situation as of early 2013 is that we have been blessed with two good designs this generation, Kepler is faster and more efficient, however AMD’s current baby may have more scope for improvement.
As we noted above, in the previous generation (HD 7xx0 and GTX 7x0) Taiwan Semiconductor Manufacturing Corp (TSMC) the foundry that manufactures both Nvidia and AMD’s graphics chips, made a long overdue move from 40nm to 28nm. In doing so, they literally halved the size of each of the billions of transistors in that generation. This secret sauce allowed both AMD and Nvidia to fit much more firepower into the same amount of silicon, without increasing power draw and allowed for epic performance increases. While that kind of gargantuan leap forward will not occur this generation (Nvidia may release 20nm GPUs at the end of the year) there is still a big difference between today’s 28nm process and the one that debuted twelve months ago. Back then there were a lot of faulty pieces that came of the production lines (more so for big chips) and limits on the number of chips that could be produced - this pushed up prices.
Nowadays 28nm chips are much easier to make, they also ‘leak’ less, making them less power hungry and thus cooler. Finally technical maturity and experience gained over of the past year could allow Nvidia and AMD to tweak their designs on a technical level. A gargantuan leap forward it may not be, but both manufacturers can field larger chips, do so cheaply and without increasing power draw or heat too much. A promising combination...