Cheap as chips: There's no such thing as a free lunch any Moore
Industry faces down Spectre of performance optimisation
Column A year ago this column mourned the death of Moore's Law, the 1965 paper so beloved by both engineers and computer scientists because of ongoing performance benefits seemingly so effortlessly achieved.
We suggested in our death notice that in lieu of flowers, donations should be lavished on Intel shares.
Researchers now believe the data-leaking Spectre chip flaw – as well as other speculative execution vulns – to be a fundamental drawback, introduced by certain classes of optimisations that kept performance-per-Watt inching forward even as process sizes crashed to a halt somewhere between 14 and 10 nanometers.
SPOILER alert, literally: Intel CPUs afflicted with simple data-spewing spec-exec vulnerabilityREAD MORE
Dead, buried, mourned, computing now gingerly walks through a post-Moore minefield – where every step forward represents a risk, and nothing comes for free anymore.
Three devices introduced over the course of the last year – two of which I purchased for my own use – frame this story perfectly.
When Microsoft announced Surface Go, I shouted "Shut up and take my money!", only to be disappointed by a device that has the right form factor and weight but feels so underpowered it sometimes makes me want to weep tears of frustration. Desiring a mid-class device, something between a desktop and a tablet, it feels too slow and too battery-hungry for a decent day of work.
If Moore's Law had ground along, Surface Go would sport an Intel CPU fabricated on a 7nm process – with more than twice the grunt of the model Microsoft shipped. That could have been a nearly perfect laptop/tablet – fast, cheap and energy efficient. Microsoft managed one of those three – it's cheap enough – but missed completely on the other two because you can no longer simply spec latest-generation components and expect them to provide great performance.
Crypto crash leads to inventory pile-up at Nvidia, sales slaughteredREAD MORE
Next – and perhaps most surprisingly – Nvidia hit the wall with its latest RTX generation. Amdahl's Law (PDF) has triumphed over Moore: instead of fabricating more transistors for even greater levels of parallelism, Nvidia designed an entirely new architecture to do "real-time ray tracing" – or, at least, something much closer to ray tracing than traditional real-time 3D.
This unexpected turn away from parallelism into dedicated hardware design means every AAA games developer now needed to rewrite the forest of highly hardware-specific rendering code to exploit the strengths of a new chip architecture that, truth be told, hasn't sold as well as hoped. Nvidia has been caught out in the classic chicken-and-egg confronting all hardware houses – without a significant installed base, it won't attract software developers, and without those developers, no one will buy its thousand-dollar cards.
Finally, there's Apple, whose purchase of PA Semi back in 2008 has proven its worth: the current A12X generation of chips – powering my new iPad Pro – prove to have the best performance-per-watt of any CPU in history. iPad Pro is everything Surface Go should have been: snappy, responsive and grunty enough to handle computationally intensive tasks while maintaining long battery life, making it without doubt the best laptop I've ever owned. (If only it ran macOS!)
Microsoft's cheapo Surface: Like a netbook you can't upgradeREAD MORE
Apple has managed to find a middle ground between Microsoft, which is stuck using bog-standard components, and Nvidia, which is flailing as it tries to drag everyone into its new architecture. As the glue that holds Apple's entire mobile product line together, iOS engages hardware capacity where it's available, making the most of whatever hardware you've got – because Apple designed that hardware from transistors on up.
According to 2017 Turing Award winners John Hennessy and David Patterson, that fusion of custom hardware and operating systems points to our future. We can't separate them easily anymore, not if we want to see steady performance gains. We're already living in the post-Moore's Law future, which is both more diverse and more specific. It's a future where we need to think carefully before we act – a future where we'll have to work hard if we want, er, Moore. ®