For well over a decade, Nintendo has pursued a console strategy that put an emphasis on new and emerging gameplay as opposed to pushing graphics and CPU horsepower. This strategy has had decidedly mixed results, and with the Nintendo Switch, there was speculation that the company might have adopted a fundamentally different approach. Unlike the Wii and Wii U, Switch is supposed to be a hybrid mobile and living room console, suggesting Nintendo might put much more emphasis on adopting a leading-edge process node and cutting-edge graphics technology to reduce power consumption and improve performance-per-watt.
A new report from GamesBeat suggests this is not the case — at least not to the extent that fans may have hoped. Sources have stated that the SoC inside the Nintendo Switch is built on Maxwell, not Pascal, and uses 20nm technology, presumably from TSMC. This would be ironic if true, considering Nvidia was the first major firm to raise concerns about the suitability of TSMC’s 20nm node for GPUs back in 2012, but the mobile SoC that Nvidia built for Nintendo would be a far cry from the 150-250W TDP graphics cards Nvidia had in mind back then. Maxwell was a power-efficient architecture and the die shrink from 28nm to 20nm would definitely give the chip a little extra headroom, but the fact that it’s a tablet part is going to put significant restrictions on how much horsepower Nintendo can deliver.
To some extent, this may not matter. The Switch only has a 720p screen and Nintendo could stretch its power budget further by cutting a game’s native resolution to 540p and upscaling for 720p or even 1080p when plugged into a television. The PlayStation 4 Pro has demonstrated that this kind of approach can produce very good results, even if they aren’t quite as good as native content. GamesBeat expects the Switch to offer ~1 TFLOP of processing power, which would be ~2x better than the Tegra X1 Nvidia launched back in January 2015. A die-shrink to 20nm could’ve helped Nvidia get partway there, but we’ll have to wait and see if the chip offers this much processing power. At the same time, however, we have to note that TFLOPS are the wrong way to compare potential horsepower against graphics processors.
The TFLOPS metric that Nvidia, AMD, and Intel all collectively like to quote is nothing but a measurement of peak throughput given an idealized set of instructions and an infinite power budget. It can be used to approximate performance within the same GPU family, but it doesn’t tell you much about performance between different families. TFLOPS only captures the number of floating point operations a GPU can complete per second; it doesn’t tell you how long a tablet can sustain that rate, how much fill rate the GPU offers, or how many texture mapping units it has. That’s why we typically refer to these three measurements in sequence. The Tegra X1, for example, is a 512:16:16 design, with 512 cores, 16 texture mapping units (TMUs) and 16 render outputs (ROPS). Power consumption, not TFLOPS, is almost certainly the limiting factor here.
Given these new revelations, we can start to put some upper and lower boundaries on the Switch’s potential performance. The worst-case scenario — and I find this unlikely — is that the Switch would offer Wii U-equivalent graphics in a mobile form factor. What seems more likely is that Nintendo will offer some iterative improvements over the Wii U, but that developers will still have to work hard to balance frame rate, power consumption, and system battery life.
As for why Nintendo didn’t use Pascal or 14/16nm technology, the answer is simple: When Nintendo started working the Switch, those technologies weren’t mature enough for adoption. Nintendo appears to have recognized that one of the things that crippled the Wii U was the lack of games immediately after launch, and wanted to get hardware into developer hands quickly to avoid a repeat of this disaster. That meant working with technology that was readily available, and that meant rolling a custom version of Nvidia’s X1 rather than waiting until a Pascal-derived SoC was available.