facebook rss twitter

NVIDIA reveals upcoming GPU architectures

by Pete Mason on 22 September 2010, 10:53

Tags: NVIDIA (NASDAQ:NVDA)

Quick Link: HEXUS.net/qaz5y

Add to My Vault: x

NVIDIA CEO Jen-Hsun Huang kicked off this year's GPU Technology Conference (GTC) in San Jose and used his keynote to shed a little light on the company's future GPUs - including revealing the architectures' codenames for the first time.

Entering into production sometime at the end of 2011 will be the Kepler architecture (continuing the theme of great scientists started with Tesla and Fermi).  This will be based on a 28nm process and progress on the design is apparently progressing rapidly.

The focus for NVIDIA is on efficiency, with Huang telling the audience that "transistors are free, power is not".  Accordingly the company is promising to increase performance per watt by three to four times with the Kepler GPUs.

Following on from this will be the Maxwell architecture, which is expected to arrive in "a few years".  While there were few details on what these chips would bring, the CEO anticipated a 16-fold increase in performance per watt over Tesla, showing almost exponential growth from generation to generation.

In the same way that Fermi brought ECC, Maxwell will introduce a series of new features specifically for high-performance computing.  This includes pre-emption, virtual-memory support and a reduced reliance on the CPU, meaning that less processing time will be wasted while the GPU waits for data from the main processor.

For now, this is all that we know about NVIDIA's next-generation architectures.  While the focus of GTC is on GPGPU and high-performance computing, there will undoubtedly be more news over the coming days on the company's plans for the future.



HEXUS Forums :: 5 Comments

Login with Forum Account

Don't have an account? Register today!
Performance per watt.. who'd have thought? :P
Seems kinda wishful/aspirational rather than informed projections to me, but then what do I know?
Well now AMD are dropping the ATI maybe nvidia are trying to become the new ATI?
I think it's an excellent idea for NVIDIA to be focusing on PPW and efficiency after what happened with the original GF100 architecture, but it's interesting to note that Fermi actually has better PPW than Tesla did, just because you get a lot of performance for every watt the card uses doesn't mean it's not going to use a hell of a lot of them and still run extremely hot.

Still, it should be an interesting battle between the Kepler architecture and AMD's 7000 series (presuming Northern Islands is released around the same time). Although I guess it remains to be seen if NVIDIA can even match the timeframe given within that roadmap, considering how long it took them to finalize Fermi.
Pandora92
but it's interesting to note that Fermi actually has better PPW than Tesla did, just because you get a lot of performance for every watt the card uses doesn't mean it's not going to use a hell of a lot of them and still run extremely hot.

You'd have to be a right muppet not to get better performance per watt whilst transitioning to a lower process size. If the process size halves, then you get 4 transistors where previously you had just one. There'll be differences in leakage current for your smaller transistors, but you should certainly end up with at least twice the performance per watt afterwards.

What I'd be interested to see is the performance per watt per transistor. Then you factor out the gains Moore's Law give you.

Jen-Hsun Huang
“transistors are free, power is not”

Sweet. Can I have 3 billion of your transistors in the form of a GTX 480 then please Jen-Hsun? :clapping: