NVIDIA's Fermi: Architected for Tesla, 3 Billion Transistors in 2010
by Anand Lal Shimpi on September 30, 2009 12:00 AM EST- Posted in
- GPUs
A More Efficient Architecture
GPUs, like CPUs, work on streams of instructions called threads. While high end CPUs work on as many as 8 complicated threads at a time, GPUs handle many more threads in parallel.
The table below shows just how many threads each generation of NVIDIA GPU can have in flight at the same time:
| Fermi | GT200 | G80 | |
| Max Threads in Flight | 24576 | 30720 | 12288 |
Fermi can't actually support as many threads in parallel as GT200. NVIDIA found that the majority of compute cases were bound by shared memory size, not thread count in GT200. Thus thread count went down, and shared memory size went up in Fermi.
NVIDIA groups 32 threads into a unit called a warp (taken from the looming term warp, referring to a group of parallel threads). In GT200 and G80, half of a warp was issued to an SM every clock cycle. In other words, it takes two clocks to issue a full 32 threads to a single SM.
In previous architectures, the SM dispatch logic was closely coupled to the execution hardware. If you sent threads to the SFU, the entire SM couldn't issue new instructions until those instructions were done executing. If the only execution units in use were in your SFUs, the vast majority of your SM in GT200/G80 went unused. That's terrible for efficiency.
Fermi fixes this. There are two independent dispatch units at the front end of each SM in Fermi. These units are completely decoupled from the rest of the SM. Each dispatch unit can select and issue half of a warp every clock cycle. The threads can be from different warps in order to optimize the chance of finding independent operations.
There's a full crossbar between the dispatch units and the execution hardware in the SM. Each unit can dispatch threads to any group of units within the SM (with some limitations).
The inflexibility of NVIDIA's threading architecture is that every thread in the warp must be executing the same instruction at the same time. If they are, then you get full utilization of your resources. If they aren't, then some units go idle.
A single SM can execute:
| Fermi | FP32 | FP64 | INT | SFU | LD/ST |
| Ops per clock | 32 | 16 | 32 | 4 | 16 |
If you're executing FP64 instructions the entire SM can only run at 16 ops per clock. You can't dual issue FP64 and SFU operations.
The good news is that the SFU doesn't tie up the entire SM anymore. One dispatch unit can send 16 threads to the array of cores, while another can send 16 threads to the SFU. After two clocks, the dispatchers are free to send another pair of half-warps out again. As I mentioned before, in GT200/G80 the entire SM was tied up for a full 8 cycles after an SFU issue.
The flexibility is nice, or rather, the inflexibility of GT200/G80 was horrible for efficiency and Fermi fixes that.
Facebook
Twitter
RSS
415 Comments
View All Comments
Zingam - Thursday, October 1, 2009 - link
No no! This is just on paper! When will see it for real!! Oh... Q2-3-4 next year! :)So you cannot claim they have the better thing because they don't have it yet! And don't forget next year we might have the head-smashing Larrabee!
:)
Who knows!!! I think you are way to biased and not objective when you type!
chizow - Thursday, October 1, 2009 - link
Heheh if Q2 is what you want to believe when you cry yourself to sleep every night, so be it. ;)Seriously though, its looking like late Q4 or early Q1 and its undoubtedly meant for one single purpose: to destroy the world of ATI GPUs.
As for Larrabee lol...check out some of the IDF news about it. Even Anand hints at Laughabee's failure in his article here. It may compete as a GPGPU extension of x86, but not as a traditional 3D raster, not even close.
SiliconDoc - Thursday, October 1, 2009 - link
Gosh you'd be correct except here is the FERMIhttp://www.fudzilla.com/content/view/15762/1/">http://www.fudzilla.com/content/view/15762/1/
There it is bubba. you blew your yap wide open in ignorance and LOST.
Good job, you've got plenty of company.
ClownPuncher - Thursday, October 1, 2009 - link
Wow, a video card! On top of that pcb could be a cat shit for all we know. The card does not exist, because I can't touch it, I can't buy it, and I can't play games on it.Also, the fact that you seem to get all of your info from Fudzilla speaks volumes. All of your syphillus induced mad ramblings are tiresome.
Lifted - Thursday, October 1, 2009 - link
I see what appears to be a PCB with some plastic attached, and possibly a fan in there as well. Yawn.ksherman - Wednesday, September 30, 2009 - link
Really like these kind of leaps in computing power, I find it fascinating. A shame that it seems nVidia is pulling a bit away from the mainstream graphics segment, but I suppose that means that the new cards from ATI/AMD are the undisputed choice for a graphics card in the next few months. 5850 it is!fri2219 - Wednesday, September 30, 2009 - link
For the love of Strunk and White, stop murdering English in that manner- it detracts from the text buried between banner ads.Sunday Ironfoot - Wednesday, September 30, 2009 - link
nVidia have invented a new way to fry eggs, just crack one open on top of their GPU and play some Crysis. :-)SiliconDoc - Wednesday, September 30, 2009 - link
Let's crack it on page 4. A mjore efficient architecture max threads in flight. Although the DOWNSIDE is sure to be mentioned FIRST as in "not as many as GT200", and the differences mentioned later, the hidden conclusion with the dissing included is apparent.Let's draw it OUT.
---
What should have been said 1st:
Nvidia's new core is 4 times more efficient with threads in flight, so it reduces the number of those from 30,720 to 24,576, maintaining an impressive INCREASE.
---
Yes, now the simple calculation:
GT200 30720x2 = 61,440 GT300 24576x4 = 98,304
at the bottom we find second to last line the TRUTH, before the SLAM on the gt200 ends the page:
" After two clocks, the dispatchers are free to send another pair of half-warps out again. As I mentioned before, in GT200/G80 the entire SM was tied up for a full 8 cycles after an SFU issue."
4 to 1, 4 times better, 1/4th the clock cycles needed
" The flexibility is nice, or rather, the inflexibility of GT200/G80 was horrible for efficiency and Fermi fixes that. "
LOL
With a 4x increase in this core design area, first we're told GT200 "had more" then were told Fermi is faster in terms that allow > the final tale, GT200 sucks.
--
I just LOVE IT, I bet nvidia does as well.
tamalero - Thursday, October 1, 2009 - link
on paper everything looks amazing, just like the R600 did in its time, and the Nvidia FX series as well. so please, just shut up and start spreading your FUD until theres real information, real benches, real useful stuff.