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
SiliconDoc - Thursday, October 1, 2009 - link
Nice consolation speech.I guess you expected " you're right ", but somehow lying to make you feel good is not in my playbook.
Now, next time you don't take it so seriously as to reply, and then still, be pathetic enough to get it wrong. Hows that for a fun deal ?
Maian - Wednesday, September 30, 2009 - link
Where's snakeoil when you need him... I don't give a shit about vendor, but the flame wars here are hilarious :DLifted - Wednesday, September 30, 2009 - link
What flame war? It's just a single nut barking at everyone for no reason. If he has a problem with the article he's sure making it difficult to figure out what it is with all his carrying on and red rooster nonsense.Does anyone (besides the nut) actually care what is said in this article? It's simply something to pass the time with, and certainly not worth getting upset over. Is the nut part of the nvidia marketing machine or merely a troll? It almost seems as if he's writing in a manner as to cover up his true identity. Yes silicondoc, it IS that obvious.
SiliconDoc - Thursday, October 1, 2009 - link
Wow, a conspiracist.Well, for your edification, you didn't score any points, since the readers here get all uppity about what's in the articles, so they have shown a propensity to care, even if you're just here to pass the time, or lie your yapper off for the convenient line it provides you for this momment.
Usually, the last stab of the sinking pirate goes something it like: " It doesn't matter !"
Then Davey Jones proves to 'em it does.
-
Nice try, but the worst problem for you is, it matters so much to you, you think I'm not me. NOW THAT's FUNNY !
ahhahahahaaha
ClownPuncher - Wednesday, September 30, 2009 - link
Aspbergers.Kaleid - Wednesday, September 30, 2009 - link
No, most people with Asperger's are highly functional. This is something else.Finally - Thursday, October 1, 2009 - link
It's Rain Man?tamalero - Friday, October 2, 2009 - link
its assburgers, google it.redpriest_ - Wednesday, September 30, 2009 - link
I can't help but note for the record that um, the card isn't out yet, so how can they win when no one knows when you can actually buy one yet? And for the record, I have a 5870 in my system, right now, that can play games....right now. I went to a retail store and bought it. That's how simple it was. I know you've been posting tons of FUD in the other review forums about how it's unavailable etc etc but the fact is, it IS available, and multiple people can own one.Also, let me state for the record that I have owned nvidia GPUs in the past so that I'm vendor agnostic. I buy whatever solution is available and better. KEY POINTS: AVAILABLE. BETTER.
SiliconDoc - Thursday, October 1, 2009 - link
Here's how they can win, here the NVidia master holds FERMI up for all to see !http://www.fudzilla.com/content/view/15762/1/">http://www.fudzilla.com/content/view/15762/1/
Aww, dat too bad for the wittle wed woosters. It really is real, little red lying hoods.