No, mobile GPUs (still) aren't going to catch dedicated consoles by the end of 2017

Every now and then, one of the mobile vendors releases a report claiming that mobile graphics either have hit console-level quality or are about to do so. This time, ARM is the company arguing the point. At the Casual Connect conference in Amsterdam, ecosystem director Nizar Romdan told reporters that the company will build chips that can catch, then outperform the GPUs inside the PlayStation 4 and Xbox One by the end of 2017.

"Mobile hardware is already powerful," said Romdan at Casual Connect Europe. "If you take today’s high-end smartphone or tablet, the performance is already better than Xbox 360 and PlayStation 3. It’s catching up quickly with Xbox One and PlayStation 4."

This graph implies that desktop GPU compute performance skyrocketed from 2009 to 2012 -- which is true, as far as it goes -- and then stopped thereafter, just below the 6TFLOP/s mark. In reality, single-precision floating point performance has continued to rise since then, albeit not at the same rate. Nvidia's M40 GPU accelerator offers up to 6.8TFLOPS/s, while AMD's Fury X is up to 8.6TFLOPS/s. This graph implies that desktop performance stopped. It didn't.

Since ARM wants to compare against consoles, however, we'll focus our discussion there. It's absolutely true that mobile GPUs have made great strides in recent years, both in terms of raw performance and power efficiency. They also support more advanced APIs than the older consoles, which further improves their ability to render complex scenes using less power, or to use more advanced effects. None of this, however, rewrites the laws of physics.

Here are power consumption figures for the Xbox One and PS4, as measured by the NDRC not long after both platforms launched.

The only data we care about in this case are the game play figures. A next-generation version of either console would use significantly less power if it was built on a 14nm process node, and even less if it used the 10nm technology that might be available by the end of 2017. For argument's sake, let's assume that the 14nm node could cut power consumption by 40%, with the 10nm node cutting another 40% off that. By 2017, our hypothetical PS4 draws 70W under full load, while the Xbox One is down to just 56W. If we assume another 25% improvement thanks to architectural enhancements, the PS4 is just 56W and the Xbox One is a positively svelte 45W.

Typical smartphone TDP is in the 4-5W range. Tablets get more headroom, but 10-12W is the usual max. Micro-consoles like the Shield might have 20W envelopes for gaming, but that's still less than half the TDP of our hypothetical Advanced PlayStation 4. Our estimates on potential power savings for current-gen consoles, meanwhile, are extremely optimistic, essentially assuming that the entire platform can be scaled to reduce power at a constant rate.

No mobile device in 2017 is going to be capable of delivering the PS4's graphics in a 4-5W TDP. And while we're on that topic, GFLOPS/s is a lousy metric for comparing performance between vastly different architectures and form factors. If the SoC can't keep those GPU cores fed with information, all the theoretical floating-point performance in the world won't mean squat. There's a reason why modern GPUs are backed by multi-channel sophisticated memory interfaces clocked at high speeds and with large integrated caches. All of these technologies boost sustained performance at the cost of power consumption. Even ARM ultimately acknowledged that mobile can't really compete on these terms. Here's The Verge again:

The problem with "console quality"

The real problem here, of course, is that "console quality" is being treated as synonymous with "High quality gaming." I enjoy an amazing visual experience as much as any gamer, but games don't need console quality graphics to offer deep, engaging experiences. Heck, depending on the kind of game you want to build, that kind of focus on visuals can actually harm the final product.