At the annual computer graphics mecca SIGGRAPH 2013, Nvidia has taken the wraps off the Logan (Tegra 5) SoC. Nvidia isn’t talking about the CPU yet (it’s probably Cortex-A15, like Tegra 4), but the GPU… oh the GPU is a mobile version of Kepler that blows every other mobile graphics solution out of the water. Mobile Kepler can match the performance of the iPad 4′s GPU while consuming one-third of the power (900mW vs. 2.6W); or, at the other end of the scale, mobile Kepler is about five times faster than the iPad 4 if it’s allowed to draw the same amount of power (2-3W). Perhaps more importantly, though, mobile Kepler finally brings feature/API parity to PCs, next-gen consoles, and smartphones. By bringing DirectX 11, OpenGL 4.4, OpenGL ES 3.0, hardware tessellation, and CUDA 5 support to smartphones, developers can write one game for a console, and then easily port it to every other platform.
Back when Nvidia first launched the Kepler architecture, the company claimed that the architecture would scale all the way up to supercomputers and down to smartphones — and, in the case of GTX Titan, it actually brought supercomputing power to the desktop. Given that the Tesla, GeForce, and Quadro variants of Kepler are all capable of drawing more than 250 watts, we understandably had our doubts that Kepler would scale down to the ultra-mobile envelope (2-3 watts). Seemingly, though, we needn’t have worried. In the video below, Nvidia’s Ira demo is running on early Logan SoC silicon — and the GPU is drawing just 2-3 watts (about the same power draw as the iPad 4′s PowerVR SGX 554MP4 GPU). This demo, which highlights the advances in rendering realistic faces, was originally shown running on a GTX Titan at GTC 2013 — a part that draws more than 100 times more power.
In the next video, you can see the Island demo, which was originally demonstrated on a desktop Fermi part. This demo shows OpenGL 4.3 and hardware tessellation running on Logan’s mobile Kepler GPU.
Yes, the demos don’t have the same polygon counts or high-res textures as their desktop-based counterparts, but they’re still technologically very impressive and way beyond the current state of the art for mobile graphics.
Internally, Logan’s GPU consists of a single Kepler SMX (the top-of-the-line GK110 in the GTX Titan uses 14 SMXes). There are probably some other changes to reduce power consumption, but it’s still essentially Kepler. Going by the graph below, which shows mobile Kepler to have five times the raw horsepower of the iPad 4′s SGX 554MP4 (77 gigaflops), we can infer that mobile Kepler will top out at 400 gigaflops – giving us peak theoretical performance that’s slightly better than the 8800 GTX (a seven-year-old chip) and significantly higher than the PS3. We have no idea of the clock speed or memory subsystem at this point, but we do know that a single Kepler SMX contains 192 CUDA cores — and as each core is good for two flops per cycle, that means mobile Kepler needs to be clocked at around 1GHz to hit 400 gigaflops.
We’re unlikely to see a smartphone GPU clocked at 1GHz — but even at a more reasonable 500MHz, mobile Kepler is still comparable to the PS3′s GPU, and better than any other mobile GPU on the market. Don’t forget Kepler’s best-in-class support for modern APIs, too.
The big question now, though, is when will Nvidia actually bring mobile Kepler to market? Tegra 4, which is behind schedule, will finally come to market next week with Nvidia’s own Shield portable. Given the delays, it’s uncertain how many Tegra 4 devices will actually arrive, and how many will have switched to another SoC, such as Qualcomm’s Snapdragon 800. The reference device shown in the videos above uses early silicon that has only just come back from the fabs — so, we’re probably looking at at least 12 months before Tegra 5 lands in consumer devices.
Still, when it does eventually arrive, we’ll not only have a mobile chip that allows for console-quality games, but it might also be the trigger for developers to seriously invest in high-quality mobile games. Exciting times indeed.