Nvidia CEO Jen-Hsun Huang took to the stage this morning at the company's GPU Technology Conference to deliver the keynote, and over the next two-and-a-half hours he proceeded to expound on most of Nvidia's current product lineup. Among the tech demos and examples of how Nvidia's technology is being used by various companies, Huang gave us some details on the company's product roadmaps for both its GPUs and its Tegra mobile processors.
The next two years will see Nvidia GPUs and systems-on-chips (SoCs) borrowing plenty of ideas from one another, with the GPUs becoming ever more integrated and the SoCs becoming more powerful and even sharing GPU architectures with their desktop cousins. Let's walk through the roadmap and talk about exactly how Nvidia's lineup will evolve in 2014 and 2015.
Nvidia's current GPU architecture is called Kepler, and it has come to market in the company's GeForce 600-series GPUs, as well as in various Quadro cards, the Nvidia VGX virtualized graphics cards, and the Nvidia Grid server, among others. Nvidia's roadmap has the Kepler architecture staying on through 2013, but it will be replaced next year by a new architecture called Maxwell.
With each subsequent GPU generation, it's a given that performance will increase, and that will also be true of Maxwell. The truly interesting thing the architecture brings to the table, though, is something called "unified virtual memory." In short, this technology would make the GPU's memory accessible to the CPU, and the main system memory accessible to the GPU—all of the memory in your computer will be usable by both of the major processors. AMD's version of this idea, which it calls Heterogeneous Systems Architecture (HSA), is also due at some point in 2014.
This idea has a number of implications for overall system performance: for some GPU operations that require a large amount of memory but don't necessarily need that memory to be particularly fast, it can now easily access the 8GB or 16GB of system RAM that's becoming increasingly common in today's systems. If the CPU is performing tasks where memory access time is of paramount importance, it can use the GPU's fast GDDR memory rather than the system's standard DDR. Nvidia didn't go into any detail on just how the system would juggle all of that memory and decide which operations use which memory, but if done well it's an interesting way to make sure you're getting the most out of your hardware.
Looking further into the murky future, Hsun-Huang then talked about Volta, Maxwell's successor. Volta concerns itself with reducing the amount of space that data must move to make it from the GPU to the memory and then back again. By stacking the RAM directly on top of the GPU (something Nvidia calls "stacked DRAM") rather than placing it around the GPU on the graphics card or system board, Nvidia can drastically increase the card's theoretical memory bandwidth: Huang promised bandwidth to the tune of one terabyte a second, a huge upgrade to even the 288.4GB/s bandwidth offered by Nvidia's massive just-launched Titan graphics card.
Nvidia provided no timeline for when Volta-based products would make it to market, but based on the two-year life cycles for the company's other architectures, it would make sense for it to hit the market sometime in 2016.
To date, Nvidia's Tegra GPUs have used two different kinds of shaders—vertex shaders and pixel shaders. This is also true of both Tegra 4 and Tegra 4i. The number of these shaders (and, thus, the GPUs' capabilities) have increased with every generation, but they still lag behind Nvidia's desktop and laptop GeForce parts.
This will change with Nvidia's next-generation Tegra part, codenamed "Logan," which will finally bring "unified" shaders to the company's mobile processors. These shaders, based on the same Kepler architecture as the GeForce 600-series PC GPUs, will support all of the same APIs as Nvidia's current GeForce cards, including CUDA 5 and OpenGL 4.3. They'll also support things like PhysX, bringing more of the features of desktop graphics cards to phones and tablets.
Nvidia didn't talk much about the specific CPU architecture that Logan will use, but the company suggested that it would continue to use the same general configuration as Tegra 4 and Tegra 3 before it—four CPU cores paired with a single low-power "companion core" designed to reduce the chip's power draw during light use or when idle. We should see more of Logan later this year, and the chip should reportedly begin shipping at some point in early 2014.
Logan's successor, codenamed "Parker," will advance things on three fronts. First, it will continue importing its GPU from the GeForce line, working in 2014's "Maxwell" architecture and its attendant performance and feature additions. Next, it will use a FinFET manufacturing process not dissimilar to the one Intel is currently using for its 22nm Ivy Bridge chips. Assuming that Nvidia continues to use its longtime manufacturing partner Taiwan Semiconductor (TSMC) for this chip, the planned 2015 availability for the chip likely means that Parker will be made on a 16nm process, down from 28nm for Tegra 4 and 4i.
Finally, Parker will be the first Tegra chip to move away from ARM's Cortex CPU architecture. The chips will still use the ARM instruction set, but like Qualcomm and Apple before it, Nvidia is going to begin building its own CPU cores tailored to its own needs. We've known about this CPU architecture, codenamed Project Denver, for a little over two years now, but the Parker chip will be the first shipping product to actually use it.
At this point, the main thing we know about Project Denver's architecture is that it will be 64-bit, which also opens the door to the server room—Nvidia's efforts with its VGX server graphics cards and the Grid gaming server, among other things, suggests that this wouldn't be much of a stretch. AMD will be introducing its own ARM-based Opterons in 2014, and in the event that these do well, Nvidia would be well-positioned to create its own product based on the Denver architecture.