Nvidia is "just getting started" in the mobile device market but this week the graphics chip manufacturer unveiled big plans for its line of low-power Tegra processors at the GPU Technology Conferentece in San Jose, California.
During his opening keynote, Nvidia CEO Jen-Hsun Huang was characteristically candid in his assessment of what his company has done with Tegra thus far. The first iteration of the mobile chip, launched back in 2008 and finding its limited way into devices like Microsoft's Zune, is "something we don't really like to talk about." Successive generations of Tegra Systems-on-a-Chip (SoCs) have achieved far more success - Nvidia has seen its Tegra business grow rapidly in recent years - but the company is anything but satisfied.
Nor should it be. According to analysts, Nvidia has less than five per cent of the mobile SoC market, while Qualcomm, Samsung, and Apple combine for around 70 per cent. There's plenty of room to grow.
And growth, according to Nvidia, starts with making Tegra better - much better.
What you may not know is that the graphics processors Nvidia uses in its current, third-generation Tegra 3 chips aren't the company's latest, greatest GPUs. Far from it, in fact. Instead, Tegra uses a GPU architecture based on the GPUs Nvidia was selling back in the mid-2000s. In other words, Tegra's architecture pre-dates not just the current Kepler GPUs powering Nvidia's GeForce products for PCs, but also Tesla and Fermi.
So how does Tegra compete at all? Nvidia threw "a ton of GPU cores" and "a lot of software" at its SoCs to get them to perform as well as they do, according to Pat Moorhead, principal analyst for Moor Insights & Strategy.
Starting in 2014, however, Nvidia will take a different tack. That's when the company plans to release a fifth-gen Tegra SoC code named Logan that will utilise the Kepler architecture. The following year, Nvidia will move Maxwell, the successor to Kepler, into its Tegra product line with a mobile chip code named Parker.
"When we do that, everybody else is going to be on notice," said Nick Stam, director of technical marketing for Nvidia. Mobile graphics from Qualcomm and ARM's Mali GPUs will have a tough time keeping up with Tegra products boasting Nvidia's most advanced GPUs, he said.
Meanwhile, Nvidia will soon roll out its fourth-generation of Tegra, Tegra 4, in the second quarter of this year. That product will be available in a two-chip solution with 72 GPU cores and a separate LTE modem chip. By the fourth quarter, the company will bring out Tegra 4i, which has 60 GPU cores but has an i500 modem integrated on the same chip.
This gets Nvidia into the territory Qualcomm currently owns in terms of a fully integrated mobile SoC - and in fact, the Tegra 4i package is just half the size of Qualcomm's Snapdragon 800.
Bring on Logan and Parker
But it gets better. The following year, with Logan, Nvidia will not just use its Kepler architecture for the first time in a Tegra SoC, but it will support its CUDA programming language on a mobile chip, another first. And then in 2015, Nvidia rolls its next-gen Maxwell GPU architecture into its sixth-generation Tegra SoCs, while also utilising CPUs based on ARM's Denver instruction set (and keep in mind Nvidia has its first unlimited CPU license with this generation and can design the CPU however it likes) to produce its first-ever 64-bit mobile device processors.
Introducing CUDA and 64-bit to Tegra will open up the platform and "bring the computing power and knowledge of thousands of programmers around the world" to it, Stam said.
The only question, according to Moorhead, is can Nvidia do it? The challenge will be to add the horsepower of Nvidia's most advanced GPU technologies to a mobile SoC while keeping the power draw in the 1 to 4-watt envelope, the analyst said.
One way to get there is through process shrinks which Nvidia's foundry partners have planned for the coming years. Tegra should be manufactured on the 20 to 22nm process node, according to Moorhead. Stam also pointed out that with Parker, Nvidia will move to FinFET 3D transistors, similar to Intel's tri-gate technology, which reduce power leakage considerably.
There's another technology Nvidia has planned for its Maxwell-generation GPU architecture which could make things very interesting, Moorhead said. That's virtual shared memory, a way to eliminate complicated methods needed for programming how memory shared between the CPU and GPU is allocated. Nvidia looks to be developing a proprietary means for doing this, but the Heterogeneous System Architecture (HSA) partnership formed by Advanced Micro Devices, Qualcomm, ARM, and others last year is doing something similar but in an open standards framework.
"Maxwell gets you to something very similar to what the HSA guys are getting to, making the GPU and CPU equal partners in terms of how memory is used. Today, in phones, tablets, and PCs, the CPU has the primary access to the memory," Moorhead elaborated.
"You have to tell the GPU and the program when to move the GPU's memory on the card to real memory, and programming that is very complex. Unified virtual memory re-architects the memory subsystem and makes this far simpler, meaning you don't have to be a 'ninja programmer' to work with these GPUs."