GTC 2012: Nvidia CEO Teases Future Tesla, Demos GeForce GRID, But Keynote Lacked Follow-Through
The event started off positively enough with a discussion of the GTX 690 and some performance highlights from the GK104 (Kepler's) debut. Jen-Hsun followed with a discussion of how Kepler moves the bar forward compared to Fermi, and is capable of handling scientific workloads that dwarf those of its predecessor. He also discussed a pair of new technologies -- HyperQ and Dynamic Parallism.
Both of these functions are designed to grant Kepler more independent decision-making capabilities. Hyper-Q refers to how workloads are scheduled across the chip's streaming processors. Fermi and previous chips had a single work queue, while Kepler can handle up to 32 queues.
Dynamic parallelism refers to Kepler's ability to independently launch new threads or schedule tasks to run without referring back to the CPU. With Fermi and Tesla-class GPUs, the CPU handled setting up a task and dispatching it to the GPU; the GPU ran the task and then returned the results. Kepler is capable of handling such tasks itself, which should reduce overall latency and may improve execution efficiency.
There is, however, a significant and confusing caveat to both of the above. Jen-Hsun never differentiated between the twin GK104 chips that power the GTX 690 and the capabilities we've just discussed; both GPUs were referred to as Kepler. What the next slide shows, however, is that there's a distinct difference between the current capabilities of Kepler products and what's coming in the future.
Tesla K10 is based on the current GK104. Tesla K20 is presumably based on GK100, and won't launch until the end of the year. This confusion colored the rest of the CEO's presentations; which presumably all revolve around the capabilities of hardware we won't see in market for 8-9 months. At this point, things slowed down considerably. Jen-Hsun said at one point that he'd be discussing a GPU "many times more powerful than Kepler," as well as a GPU much smaller than anything Nvidia had currently built, but never returned to either point.
Nvidia has set its sights on two new markets -- workstation GPU virtualization and the Cloud. For the former, the company is banging the drum once again about virtualized GPU technology built into Citrix software and what it can do. The demo for these products ran ~40 minutes, despite the fact that, to the viewer, this mostly consisted of watching how quickly one could navigate a 2D desktop or watch a product render through a window. It's the type of technology that could make lives' easier for a handful of people. Go Nvidia.
He followed this with a brief discussion of a product Nvidia calls GeForce GRID. GeForce GRID is a rendering technology that purportedly allows for fast, seamless rendering off-site with the game code being piped in and run on local devices, be they a tablet or a PC. There are some significant latency challenges associated with doing this quickly, but Nvidia's demo of the technology ran quite smoothly. It would've made a much, much bigger impression if the company had set up an external site and given out a URL gamers could visit to test the sample program from various locations.
Cloud gaming is something Nvidia continues to return to, and it's true that there are significant potential benefits, but we suspect that this is one technology that needs to bake another few years before it's done. Scaling and building out a cloud gaming center is a challenge all its own, and that's before you tackle the question of keeping a game responsive during peak hours, issues surrounding data privacy, and the overarching need for cheap, high-bandwidth, low-latency connections in order to keep such games running smoothly. There are fundamental questions as to whether fast-action games will ever be particularly well suited to the wireless devices Nvidia and other companies like to demonstrate running their software. GeForce GRID is an interesting idea, but there are strategic challenges to consider.
That was it. No more news on GK100 yet, no future GPU data, no discussion of whether or not Tegra 4 will include the first CUDA-compatible GPU based on Kepler technology. Hopefully there will be more news to come out of GTC in the hours and days ahead.