Nvidia GPU Technology Conference, Final Day 3
Meeting with Nvidia Officials
ABT was fortunate to have over an hour to sit down with Nvidia’s Drew Henry, the general manager of the GeForce business unit and Jason Paul, GeForce product manager. These men are sincerely enthusiastic and attentive – and you have to consider that they have been doing these same exact kinds of interviews with the press basically non-stop for 2 days! This is very rare to find in most other companies and yet we encountered it over-and-over. Clearly these people do what they love and love what they do and it is noted as being one of the keys to personal success, never mind for success as a company.
The second meeting was with Andy Keane, general manager of the GPU computing business unit and with Andrew Humber, Tesla & CUDA Enterprise Products. Before the meeting we chatted about many things including moving to a new house – which one of them had to do over the weekend, something I hate doing. Each of these meetings went over whitepapers and answered questions that we had. Mostly it was clarifying what was said in the conference and emphasizing what Nvidia wanted us to take away from the conference; that they were committed to this new direction in parallel computing, while not forgetting the rest of their ecosystem and customers.
At this meeting was clarified the naming of the new architecture and we got to look at the new Fermi GPU. They said they were announcing Fermi early to prepare the industry for it. We could not help but think to ourselves that the timeliness of their announcement would also prepare their fans to ignore the competition’s latest DX11 offerings and to wait for the Fermi GeForce GTX.
They insisted that Fermi would also be a formidable gaming GPU and it would be the single fastest GPU when it is shipped this year . They said that Fermi’s 384-bit bus is very well-balanced with GDDR5’s increased bandwidth, although lesser models are set up to also use GDDR3. They pointed out their philosophical differences with ATi and said that within 18 months everyone would see the clear divergence. ATi is mostly committed to graphics chips while Nvidia wants to elevate the GPU to a true co-processor with the CPU. They also pointed out that PhysX was making the biggest difference in the gaming experience along with 3D Vision.
They pointed out that Ion was not an Atom exclusive and that Nvidia would build it for any CPU. However, they were not going to build DMI for Intel nor were they going to build new motherboards for AMD as their high-end CPU business had mostly collapsed; so, no new core logic for AMD in the near future is the prognosis.
They also talked about their licensing of SLi to motherboard manufacturers and how it “just works” because they have one of the largest compatibility testing labs in the world where they test all conceivable hardware combinations. They are proud of their “team testing” also, where every game setting is tested individually . They also said to look forward to a much improved hybrid SLi for extreme power savings in notebooks.
We also discussed the improvements in the Fermi architecture over Tesla’s besides the obvious doubling of the cores to 512. They pointed out the new GigaThread technology that allows for each solver to be a kernel to execute instructions is much faster than waiting in a queue for the previous instructions to complete in the old serial fashion. Thus there is greater efficiency when all 512 cores are used instead of having idle waiting time.
They were very proud of their new debugging tools which would allow a much higher quality of application to be created more quickly and we met CUDA’s creator, Ian Buck, in our second session. We asked if the philosophical differences between their chief scientists have made any changes yet. Bill Dally brings extraordinary parallel programming experience to Nvidia but Fermi was started about four years ago; long before he came to Nvidia.
