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
LawRecordings - Thursday, October 1, 2009 - link
Buwahahaha!!!What a sad, lonely little life this Silicon Doc must lead. I struggle to see how this guy can have any friends, not to mention a significant other. Or even people that can stand being in a room with him for long. Prolly the stereotypical fat boy in his mom's basement.
Careful SD, the "red roosters" are out to get you! Its all a conspiracy to overthrow the universe, and you're the only one that knows!
Great article Anand, as always.
Regards,
Law
Vendor agnostic buyer of the best price / performance GPU at the time
SiliconDoc - Thursday, October 1, 2009 - link
They can't get me, they've gotten themselves, and I've just mashed their face in it.And you're stupid enough to only be able to repeat the more than thousandth time repeated internet insult cleche's, and by your ignorant post, it appears you are an audiophile who sits bouncing around like a retard with headphones on, before after and during getting cooked on some weird dope, a HouseHead, right ? And that of course does not mean a family, doper.
So you giggle like a little girl and repeat what you read since that's all the stoned gourd can muster, then you kiss that rear nice and tight, brown nose.
Don't forget your personal claim to utter innocence, either, mr unbiased.
LOL
Yep there we have it, a househead music doused doped up butt kisser with a lame cleche'd brain and a giggly girl tude.
Golly, what were you saying about wifeless and friendless ?
ClownPuncher - Thursday, October 1, 2009 - link
What exactly is a cleche?Is it anything like a cliche?
Your spelling, grammar, and general lack of communication skill lead me to think that you are actually a double agent, it's an act if you will...an ATI guy posing as a really socially stunted Nvidia fan in an attempt to turn people off of Nvidia products solely by the ineptitude of your rhetoric.
UNCjigga - Wednesday, September 30, 2009 - link
I'd hate to have a political conversation with SiliconDoc, but I digress...Some very interesting information came out in today's previews. Will Fermi be a bigger chip than Cypress? Certainly. Will it be more *powerful* than Cypress? Possibly. Will it be more expensive than Cypress? Probably. Will it have more memory bandwidth than Cypress? Yes.
Will it *play games* better than Cypress? Remains to be seen. Too many factors at play here. We don't know clock speeds. We have no idea if "midrange" Fermi cards will retain the 384-bit memory interface. We have
For all we know, all of Fermi's optimizations will mean great things for OpenCL and DirectCompute, but how many *games* make use of these APIs today? How can we compare DirectX 11 performance with few games and no Fermi silicon available for testing? Most of the people here will care about game performance, not Tesla or GPGPU. Hell, its been years since CUDA and Stream arrived and I'm still waiting for a decent video encoding/transcoding solution.
Calin - Thursday, October 1, 2009 - link
Even between current cards (NVIDIA and AMD/ATI) the performance crown moves from one game to another - one card could do very well in one game and much worse in another (compared to the competition). As for not yet released cards, performance numbers in games can only be divined, not predictedBull Dog - Wednesday, September 30, 2009 - link
So how much in NVIDIA's focus group partner paying you to post this stuff?dzoni2k2 - Wednesday, September 30, 2009 - link
You seriously need to take your medicine. And call your shrink.dragonsqrrl - Thursday, October 1, 2009 - link
I know it seems like SiliconDoc is going on a ranting rage, because he kinda is, but the fact remains that this was a fairly biased article on the part of Anandtech. I've been reading reviews and articles here for a long time, and recently there has been a certain level of prejudice against Nvidia and its products that I haven't noticed on other legitimate review cites. This seems to have been the result of Anandtech getting left out of the loop last year. Throughout the article there is a pretty obvious sarcastic undertone towards what the Nvidia representatives say, and their newly announced GPU. I can only hope that this stops, so that anandtech can return to its former days of relatively balanced and fair reporting, which is all anyone can ask of any legitimate review cite. Articles of this manner and tone serve no purpose but to enrage people like SiliconDoc, and hurt Anandtech's image and reputation as a balanced a legitimate tech cite.Keeir - Thursday, October 1, 2009 - link
Curious in where you see the Bias.I see a little bit of the tone, but it seems warranted for a company that has for the last few years over-promised and under delivered. Very similar to how AMD/ATI was treated upto the release of the 4 series. Nvidia needs to prove (again) that it can deliever a real innovative product priced at an affordable level for the core audience of graphics cards.
Here we are, 7 days after 5870 launch and Egg has 5870s for ~375 to GTX 295s at 500. Yet again, ATI/AMD has made it a puzzling choice to buy any Nvida product more than 200 dollars.... for months at a time.
SiliconDoc - Friday, October 2, 2009 - link
What's puzzling is you are so out of touch, you don't realize the GTX295's were $406 before ati launched it's epic failure, then the gtx295 rose to $469 and the 5870 author edsxplained in text the pre launch price, and now you say the GTX295 is at $500.Clearly, the market has decided the 5870 is epic failure, and instead of bringing down the GTX295, it has increased it's value !
ROFLMAO
Awwww, the poor ati failure card drove up the price of the GTX295.
Awww, poor little red roosters, sorry I had to explain it to you, it's better if you tell yourself some imaginary delusion and spew it everywhere.