Tuesday, 21 September 2010 18:49

Nvidia details performance improvements of new GPU architectures

Written by Jon Worrel

nvidia

GTC 2010: Kepler in 2011, Maxwell in 2013

During the GPU Technology Conference, Nvidia CEO Jen-Hsun Huang surprised hardware journalists with a very surprising announcement regarding the lineup of next-generation Nvidia GPU architectures. 

"For the very first time in the history of our company, we are going to tell you the codenames and the direction of our next several generations of processors," said Jen-Hsun Huang. We’re shipping Fermi today, which brought to the world very high-performance double precision. Fermi is rated at 768GFLOPS peak."

Jen-Hsun presented a keynote slide to the audience revealing two upcoming GPU architecture codenames, Kepler 28nm (2011) and Maxwell (2013).

"Our next-generation GPU, Kepler, also named after a scientist, is expected to deliver 3-4x the performance-per-watt of Fermi. Kepler is based on 28nm and we expect to go into production next year. By the time that we’re done with the Kepler family, we wil have probably invested a couple billion dollars in R&D for it. Kepler will achieve a big step up in comparison to Fermi in the area of performance-per-watt”

Jen-Hsun went on to explain the major design problem that faces parallel computing architectures. In perspective - "transistors are free, but power is not. If we are conscious about the use of performance-per-watt architectural ideas, then we'll continue to expand performance along with the number of transistors in new architectures."

Nvidia's CEO continued by announcing Maxwell, an even more futuristic architecture expected to be delivered in 2013. "Maxwell is going to be 16x performance improvement relative to our last GTC (2009). In just a few more years, we’re going to see a 16x improvement in performance for parallel computing applications. Between now and Maxwell, we’re going to introduce features like virtual memory. We’re going to enhance the GPU’s ability to autonomously process, so it’s less dependent on the CPU, along with a very large improvement in performance.”

Last modified on Wednesday, 22 September 2010 18:03
blog comments powered by Disqus

Comments  

 
+21 #1 Jigar 2010-09-21 19:04
This will really boost certain video encoding application that takes advantage of GPU let alone be the games. I just hope ATI also jumps into this race (Application support i.e.). We really need competition here or else this market will be monopolized by Nvidia.
 
 
+37 #2 FnuGk 2010-09-21 20:48
FINALLY they started thinking about performance-per-watt
 
 
+7 #3 spineless 2010-09-21 20:55
AMD just won
 
 
+34 #4 yourma2000 2010-09-21 21:06
Everytime I hear Nvidia it's always about GPGPU and CUDA, But what about gaming and gamers?
 
 
+20 #5 spede 2010-09-21 21:56
"is expected to deliver 3-4x the performance-per-watt of Fermi" That should happen pretty much automatically switching from 40nm to 28nm. AMD will get to 28nm first as usual.
 
 
+14 #6 Alexko 2010-09-21 22:53
"Fermi is rated at 768GFLOPS peak."

No, it isn't.

"Double Precision floating point performance (peak): 515 Gflops"

And I'm not the one saying that, NVIDIA is: http://www.nvidia.com/object/product_tesla_C2050_C2070_us.html > specifications.
 
 
-13 #7 Super XP 2010-09-21 23:38
I wonder how AMD's future GPU code name Cyclops architecture is going to compete with NVIDIA's offerings.
 
 
+2 #8 Squall_Leonhart 2010-09-21 23:54
Quoting Alexko:
"Fermi is rated at 768GFLOPS peak."

No, it isn't.

"Double Precision floating point performance (peak): 515 Gflops"

And I'm not the one saying that, NVIDIA is: http://www.nvidia.com/object/product_tesla_C2050_C2070_us.html > specifications.









its closer to 670Gflops
 
 
+5 #9 Alexko 2010-09-22 08:05
Quoting Squall_Leonhart :
Quoting Alexko:
"Fermi is rated at 768GFLOPS peak."

No, it isn't.

"Double Precision floating point performance (peak): 515 Gflops"

And I'm not the one saying that, NVIDIA is: http://www.nvidia.com/object/product_tesla_C2050_C2070_us.html > specifications.









It's closer to 670Gflops















Nope. It would be on the GTX 480, except DP is crippled on GeForce products, so 515 DP GFLOPS is the best Fermi can do in practice. Besides, even if if were 672 GFLOPS, that's still not 768!
 
 
+1 #10 Nerdmaster 2010-09-22 11:41
Quoting Jen-Hsun Huang:
In perspective - "transistors are free, but power is not. If we are conscious about the use of performance-per-watt architectural ideas, then we'll continue to expand performance along with the number of transistors in new architectures."

Are you kidding me? Adding more transistors to a single die and you get an exponential increase to the cost.
 

To be able to post comments please log-in with Disqus

 

Facebook activity

Latest Commented Articles

Recent Comments