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
The R600 was great, you idiot.Of course, when hating nvidia is your real gig, I don't expect you to do anything but be parrot off someone else's text and get the idea wrong, get the repeating incorrect.
-
The R600 was and is great, and has held up a long time, like the G80. Of course if you actually had a clue, you'd know that, and be aware that you refuted your own attempt at a counterpoint, since the R600 was "great on paper" and also "in gaming machines".
It's a lot of fun when so many fools self-proof it trying to do anything other than scream lunatic.
Great job, you put down a really good ATI card, and slapped yourself and your point, doing it. It's pathetic, but I can;t claim it's not SOP, so you have plenty of company.
papapapapapapapababy - Wednesday, September 30, 2009 - link
because both ms and sony are copying nintendo...that means, next consoles > minuscule speed bump, low price and (lame) motion control attached. All this tech is useless with no real killer ap EXCLUSIVE FOR THE PC! But hey who cares, lets play PONG at 900 fps !
Lonyo - Wednesday, September 30, 2009 - link
Did you even read the article?The point of this tech is to move away from games, so the killer app for it won't be games, but HPC programs.
SiliconDoc - Thursday, October 1, 2009 - link
I think the point is - the last GT200 was ALSO TESLA -- and so of course...It's the SECOND TIME the red roosters can cluck and cluck and cluck "it won't be any good" , and "it's not for gaming".
LOL
Wrong before, wrong again, but never able to learn from their mistakes, the barnyard animals.
Zingam - Thursday, October 1, 2009 - link
Last time I bought the most expensive GPU available was Riva TNT!Sorry but even if they offer this for gamers I won't be able to buy it. It is high above my budget.
I'd buy based on quality/price/features! And not based on who has the better card on paper in year 20xx.
SiliconDoc - Thursday, October 1, 2009 - link
Well, for that, I am sorry in a sense, but on the other hand find it hard to believe, depending upon your location in the world.Better luck if you're stuck in a bad place, and good luck on keeping your internet connection in that case.
ClownPuncher - Thursday, October 1, 2009 - link
Or maybe he has other priorities besides being an asshole.SiliconDoc - Thursday, October 1, 2009 - link
Being unable, and choosing not to, are two different things.And generally speaking ati users are unable, and therefore cannot choose to, because they sit on that thing you talk about being.
Now that's how you knockout a clown.
Lord 666 - Wednesday, September 30, 2009 - link
That actually just made my day; seeing a VP of Marketing speak their mind.Cybersciver - Friday, October 2, 2009 - link
Yeah, that was cool.Don't know about you guys, but my interest in GPU's is gaming @ 1920X1200. From that pov it looks like Nvidia's about to crack a coconut with a ten-ton press.
My 280 runs just about everything flat-out (except Crysis naturally)and the 5850 beats it. So why spend more? Most everything's a consul port these days and they aren't slated for an upgrade till 2012, least last I heard.
Boo hoo.
Guess that's why multiple-screen gaming strating to be pushed.
No way Jose.