Gearbox President, CEO, and all-around forgetful person Randy Pitchford got into a fight with the head of Microsoft’s Xbox division over the meaning, nature, and current applicability of Moore’s Law, including whether or not it still exists in the first place.
Merry Christmas. Let’s get to it.
First off, Xbox lead Phil Spencer tweeted about how the slowing down of Moore’s Law meant that software innovations like variable rate shading (VRS) are going to be important in the next generation.
DF does good work. As Moore’s law slows and our perf ambitions increase it leads to design innovation. Software innovations like VRS will also be critical – Xbox Series X rewrites the rules of console design – and the power level should be extraordinary https://t.co/1Kddsolxti
— Phil Spencer (@XboxP3) December 19, 2019
Pretty standard stuff, really. But then Randy Pitchford decided to give a lecture on how Moore’s Law works:
Is Moore’s Law slowing down? How many transistors in the Series X? What if Moore’s Law is like the 4-minute mile? Your ambitious message for the Xbox One X was inspiring, but for Series X, well, this feels more like an excuse. https://t.co/7ihDmwWzO3
— Randy Pitchford (@DuvalMagic) December 19, 2019
The problem with trying to talk about Moore’s Law, inevitably, is that Moore’s Law is not just one thing. There’s the original law as formulated by Gordon Moore in 1965, which predicted a doubling of transistor density every year. Later, he revised this to every two years. In 2015, Intel’s then-CEO Brian Krzanich acknowledged that the cadence had slowed down to a true density doubling more like every 2.5 years. For decades, Moore’s Law was often conflated with a different, related principle called Dennard scaling, which predicted that smaller transistors would use proportionally less voltage and could therefore be clocked higher using less energy. Because Moore’s Law functionally enabled Dennard scaling, people tended to drop Dennard and just say “Moore’s Law” as a way of referencing the “smaller, faster, cheaper” treadmill.
Unfortunately, Pitchford decided to double-down with a graph that doesn’t really prove his point.
First of all, the rate of transistor density has slowed — you can see it flattening out somewhat over time, compared to the speed of improvement earlier in history. But remember, this is only a graph of transistor count alone. What happens when we include other data points that are often discussed as part of “Moore’s Law?”
The data says that Moore’s Law is unquestionably slowing down, and that the factors that used to drive higher performance aren’t scaling any more. Furthermore, while density is still improving, clocks aren’t. Single-thread performance is moving upwards only very slowly. Intel has managed to nudge its 14nm all the way up to 5GHz in some high-end SKUs, but its current crop of 10nm chips have much lower clocks. AMD’s 7nm CPUs are far faster than previous parts, but they don’t hit super-high clock speeds, either. The clock speed component of Moore’s Law, which was always borrowed from Dennard anyway, is finished. Density improvements continue, but getting useful gains out of them in terms of overall performance has become vastly more difficult. The entire reason AMD adopted chiplets was because continuing to shrink certain aspects of the CPU now produces more negative effects than positive ones.
Companies are never going to stop talking about Moore’s Law. It’s too convenient a method for summarizing the idea that computers get better over time, and the public is familiar enough with it to have a vague grasp of that concept. AMD and Intel are turning to technologies and ideas they haven’t previously used, like chiplets for AMD and new interconnect technologies for Intel, but we’re pulling in knowledge from domains that didn’t used to be part of the lithographic shrink process. Variable rate shading, the feature Spencer mentions, likely is important for the future of Xbox efficiency, but part of why we need new approaches is because the old ones are running out of gas.
I expect the Xbox Series X to offer a significant performance improvement over the Xbox One X, but I think the biggest gains are going to be on the CPU and storage side of the equation. Using an eight-core Ryzen CPU instead of a Jaguar part will deliver an enormous uplift in CPU performance, while the shift from HDDs to SSDs — and not just any drives, but ultra-high-speed drives, by all reports — should allow for storage improvements so dramatic, it could lead to changes in game design. The gains on the GPU side will also be significant, but I think the CPU and storage uplift will be the bigger drivers.