Are we at the end of Moore's law already?

Recently NVidia and AMD (what was ATI) released their new graphics cards. For the past few years they have been releasing the new cards in the winter ready for Christmas. Each year Moore's law has brought big performance improvements. Perfornance was not quite doubling every 18 months but it was close enough. A graphics card 2 years old would not be able to play the most demanding modern games. That has changed this season with AMD and NVidia bringing new cards to market that get about 15% extra performance. Its not a good improvement and unfortunately the draught will last more than a year.

Both NVidia and AMD use TSMC for their manufacturing. The improvements in performance come mostly from an increase in the number of transistors that are available for the cards. The more transistors they have the more shaders they can have. 3D graphics is thankfully a highly parallel process and GPUs are already up to 512 or more actual processors working on parts of the problem. Improvements will likely be seen up to well above a million processors so future increases in transistors should bring good improvements in performance without sacrificing efficiency.

Unfortunately TSMC didn't deliver 32nm transistors in time for Christmas. Instead it scrapped the entire 32nm project. TSMC is now focussing on 28nm process and skipping 32nm altogether after having a raft of problems. So this seasons graphics cards are based on 40nm technology just like the year before. Its not surprising therefore that the optimisations that the two companies have applied have brought minor improvements in performance. All they could do was use what they had learnt in the real world about which bits of the card were the bottlenecks in real games and computations and feed that back into the design process. Its worthwhile taking some time to optimise but 15% isn't a good result.

Global foundries are another company available to AMD and NVidia for making their GPUs but unfortunately they too have cancelled their 32nm project and gone straight to 28nm. Intel has however managed to get 32nm working and they use it in their latest i7 CPUs. However its not used in a lot of their devices, most of the volume that Intel sells is at 45nm, which makes me suspect the yields are quite low. Intel does not sell their manufacturing capability to their competitors however so we wont be seeing graphics cards based on Intel's process.

More worrying is that Intel's move from 45nm to 32 nm brought little gain. There was a small reduction in power for the same number of transistors and a slight reduction in voltage. But overall it didn't allow Intel to produce CPUs that ran at a faster clock speed (about 5%), nor pack more cores onto a CPU. It didn't even save much in the way of power consumption.

Power consumption is the big problem in CPU and Graphics cards today. As transistors have gotten smaller the amount of power consumed per transistor has dropped, but it hasn't dropped at the same rate as the size. This has led to both the CPU and GPU hitting cooling limits. For the CPU cooling seems to be limited to around 120W and for GPUs the maximum power draw of the entire card is 300W as defined by the PCI Express specification. Both are at those limits and have been for several years already.

Since Intel's 32nm brought little reduction in power consumption there is no reason to believe that TSMC and Global Foundries 28nm process will bring substantial gains either. There will be a small reduction in power allowing NVidia and AMD to improve the performance of their GPUs a bit, but its not going to be close to doubling performance as the reduction wont address the heat problems much.

It looks like we are approaching the end of Moore's law. Heat is making it impossible to add additional transistors into CPUs and GPUs and the end result is that today’s computer technology is about as fast as silicon transistors can go. Future gains are going to have to be based on architecture changes in the hardware to be more efficient and software algorithm improvements. That's not going to bring gains at the same pace we've seen in the past. What this all probably means is that you should buy one of the latest high end computers today, with a top graphics cards.

I bought an Intel i7 920 (2.8Ghz 4 core CPU) 2 years ago and Intel's current plans for desktop PCs don't show an improvement on the Horizon yet, its still pretty much the fastest CPU you can reasonably buy. Intel's next CPUs packages a rubbish graphics chip onto the CPU. This makes PCs smaller but brings no actual performance improvements to the CPU. If you bought a top end Intel i5/i7 today the actual CPU looks like it might be near the top performance available for many years. Odds are it'll be fast enough until it fails in 10 years time. The same is true of a top end graphics card, one bought two years ago is only 25% slower than today’s greatest and the horizon shows no actual improvements coming.

The end of Moore's law and the performance gains it brought has finally arrived and the news didn't even bother to pick up the story.