When we ran some benchmarks on Nvidia’s new Maxwell GTX 750 Ti, we were impressed, and noted that the chip’s power efficiency could make it a serious headache for AMD in the near future.
Leaked benchmarks from the upcoming GTX 860M point to the terrifying truth of that statement: While most high-end laptop GPUs have a TDP of 75W, the 860M displaces just 45W. Perhaps even more interestingly, if the chip is as capable as these tests indicate, it could make a potent upgrade for smartphones, tablets, and other ultra-mobile devices in the 20nm 2015/Tegra K1 timeframe as well.
First we’ll discuss laptops and desktops, since that’s what we have performance figures for. According to leaked information, the GeForce GTX 860M has 640 CUDA cores, 40 texture mapping units (TMUs), and 16 ROPs. At first glance, that would put it well behind the current GTX 760M, which has 768 cores, 64 ROPs, and 16 TMUs. It’s further dwarfed by the GTX 770M, which offers a 960:80:24 configuration.
Appearances, however, can be deceiving. In 3DMark 11′s Performance preset, the GTX 860M reportedly puts up a score of P5339, as compared to the GTX 770M’s P5023. (P stands for Performance).
The tables turn a bit in the Extreme preset, where the GTX 770M’s score of X1747 narrowly wins past the GTX 860M’s X1662. This mirrors what we saw in desktop tests, where the GTX 650 Ti Boost was occasionally significantly faster than the GTX 750 Ti in tests that were memory bandwidth-bound or scaled extremely well with Kepler’s 192-core SMX architecture.
The real kicker, however, is the claimed 45W TDP envelope for the new Maxwell. This would make sense, given that the desktop variant has a 60W TDP and Nvidia can bin the mobile parts for superior power consumption. If it’s true, this puts AMD in a very tight spot with respect to the laptop market. AMD’s 28nm GCN parts are well-aligned against Kepler in terms of rated TDPs – the GTX 770M is reportedly a 75W part. At 45W, Mobile Maxwell is in a class of its own.
The road ahead
Nvidia is holding back for 20nm before it launches a full spread of Maxwell products on the desktop, and that’s likely true in mobile as well. The 860M is probably positioned to catch the sweet spot of the midrange space – and since it offers nearly top-end performance, it’s going to do that extremely well. There’s no word on whether or not the 860M supports SLI configurations (the desktop variant doesn’t), or when we’ll see products actually shipping with the new GPU inside them.
One of the most exciting facets of this architecture, however, is the performance it should offer in a handheld. Tegra 4 may not have lit up the sales charts, but the Tegra K1's mobile GPU architecture really ought to shine, especially when it comes to highlighting the performance benefits that CUDA and GPU physics can bring to the table in mobile games.
Next year, we’ll presumably see a 20nm version of Tegra K1 with more Project Denver-derived CPU cores and an integrated Maxwell GPU. Given that improvements to TDP and smaller dies directly translate to better experiences in handhelds (since every modern environment is TDP constrained), Nvidia should be able to deliver a killer upgrade 10-12 months from now.
This kind of debut in laptops could spur additional interest in licensing the company’s graphics technology. We haven’t heard of anyone signing up for a Kepler license to date, but power efficiency turns heads these days. If Maxwell keeps delivering like this, it’s going to attract attention – possibly from companies like MediaTek or RockChip, who are hungry to differentiate themselves and gain market share against giants like Qualcomm or Samsung.