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 - Wednesday, September 30, 2009 - link
No they did not post earnings, other than in the sense IN THE RED LOSSES called sales.shotage - Wednesday, September 30, 2009 - link
I'm not sure what your argument is SiliconDuck..But maybe you should stop typing and go into hybernation to await the GT300's holy ascension from heaven! FYI: It's unhealthy to have shrines dedicated to silicon dude. Get off the GPU cr@ck!!!
On a more serious note: Nvidia are good, ATI has gotten a lot better though.
I just bought a GTX260 recently, so I'm in no hurry to buy at the moment. I'll be eagerly awaiting to see what happens when Nvidia actually have the product launch and not just some lame paper/promo launch.
SiliconDoc - Wednesday, September 30, 2009 - link
My aregument is I've heard the EXACT SAME geekfoot whine before, twice in fact. Once for G80, once for GT200, and NOW, again....Here is what the guy said I responded to:
" Nvidia is painting itself into a corner in terms of engineering and direction. As a graphical engine, ATI's architecture is both smaller, cheaper to manufacture and scales better simply by combining chips or expanding # of units as mfg tech improves.. As a compute engine, Intel's Larabee will have unmatched parallel thread processing horsepower. What is Nvidia thinking trying to pass on this huge, monolithic albatross? It will lose on both fronts. "
---
MY ARGUMENT IS : A red raging rooster who just got their last two nvidia destruction calls WRONG for G80 and GT200 (the giant brute force non-profit expensive blah blah blah), are likely to the tune of 100% - TO BE GETTING THIS CRYING SPASM WRONG AS WELL.
---
When there is clear evidence Nvidia has been a markleting genius (it's called REBRANDING by the bashing red rooster crybabies) and has a billion bucks to burn a year on R&D, the argument HAS ALREADY BEEN MADE FOR ME.
-----
The person you should be questioning is the opinionated raging nvidia disser, who by all standards jives out an arrogant WHACK JOB on nvidia, declaring DUAL defeat...
QUOTETH ! "What is Nvidia thinking trying to pass on this huge, monolithic albatross? It will lose on both fronts. "
---
LOL that huge monolithic albatross COMMANDS $475,000.00 for 4 of them in some TESLA server for the collegiate geeks and freaks all over the world- I don't suppose there is " loss on that front" do you ?
ROFLMAO
Who are you questioning and WHY ? Why aren't you seeing clearly ? Did the reds already brainwash you ? Have the last two gigantic expensive cores "destroyed nvidia" as they predicted?
--
In closing "GET A CLUE".
shotage - Wednesday, September 30, 2009 - link
Found my clue.. I hope you get help in time: http://www.physorg.com/news171819640.html">http://www.physorg.com/news171819640.htmlSiliconDoc - Thursday, October 1, 2009 - link
You are your clue, and here is your buddy, your duplicate:" What is Nvidia thinking trying to pass on this huge, monolithic albatross? It will lose on both fronts."
Now, I quite understand denial is a favorite pasttime of losers, and you've effectively joined the red club. Let me convert for you.
" What is Ati thinking trying to pass on this over length, heat soaked, barely better afterthought? It will lose on it's only front."
-there you are schmucko, a fine example of real misbehavior you pass-
AaronJD - Wednesday, September 30, 2009 - link
While I definitely prefer the $200-$300 space that ATI released 48xx at, It seems like $400 is the magic number for single GPUs. Anything much higher than that is in multi-GPU space where you can get away with a higher price to performance ratio.If Nvidia can hit the market with well engineered $400 or so card that is easily pared down, then they can hit a market ATI would have trouble scaling to while being able to easily re-badge gimped silicon to meet whatever market segment they can best compete in with whatever quality yield they get.
Regarding Larabee, I think Nvidia's strategy is to just get in the door first. To compete against Intel's first offering they don't need to do something special, they just need to get the right feature set out there. If they can get developers writing for their hardware asap Tesla will have done its job.
Zingam - Thursday, October 1, 2009 - link
Until that thing from NVIDIA comes out AMD has time to work on a response and if they are not lazy or stupid they'll have a match for it.So in any way I believe that things are going to get more interesting than ever in the next 3 years!!!
:D ?an't wait to hear what DirectX 12 will be like!!!
My guess is that in 5 years we will have a truly new CPUs - that would do what GPUs + CPUs are doing together today.
Perhaps will come to the point where we'll get blade like home PCs. If you want more power you just shove in another board. Perhaps PC architecture will change completely once software gets ready for SMP.
chizow - Wednesday, September 30, 2009 - link
Nvidia is also launching Nexus at their GDC this week, a plug-in for Visual Studio that will basically integrate all of these various API under an industry standard IDE. That's the launching point imo for cGPU, Tesla and everything else Nvidia is hoping to accompolish outside of the 3D Gaming space with Fermi.Making their hardware more accessible to create those next killer apps is what's been missing in the past with GPGPU and CUDA. Now it'll all be cGPU and transparent in your workflow within Visual Studio.
As for the news of Fermi as a gaming GPU, very excited on that front, but not all that surprised really. Nvidia was due for another home run and it looks like Fermi might just clear the ball park completely. Tough times ahead for AMD, but at least they'll be able to enjoy the 5850/5870 success for a few months.
ilkhan - Wednesday, September 30, 2009 - link
If it plays games faster/prettier at the same or better price, who cares what the architecture looks like?On a similar note, if the die looks like that first image (which is likely) chopping it to smaller price points looks incredibly easy.
papapapapapapapababy - Wednesday, September 30, 2009 - link
"Architecturally, there aren't huge lessons to be learned from RV770"SNIF SNIF BS!
"ATI's approach is much more cautious"
more like "ATI's approach is much more FOCUSED"
( eyes on the ball people)
"While Fermi will play games, it is designed to be a general purpose compute machine."
nvidia, is starting to sound like Sony " the ps3 is not a console its a supercomputer @ HD movie player, it only does everything" guess what? people wanted to play games, nintendo ( the focused company, did that > games, not movies, not hd graphics, games, motion control) Sony - like nvidia here- didn't have the eyes on the ball.