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
I'm sure Anand brought it out of him with his bias.Already on page one, we see the UNFAIR comparison to RV870, and after wailing Fermi "not double the bandwidth" - we get ZERO comparison, because of course, ATI loses BADLY.
Let me help:
NVIDIA : 240 G bandwidth
ati : 153 G bandwidth
------------------------nvidia
---------------ati
There's the bandwidth comparison, that the biased author couldn't bring himself to state. When ati LOSES, the red fans ALWAYS make NO CROSS COMPANY comparison.
Instead it's "nvidia relates to it's former core as ati relates to it's former core - so then "amount of improvement" "within in each company" can be said to "be similar" while the ACTUAL STAT is "OMITTED !
---
Congratulations once again for the immediate massive bias. Just wonderful.
omitted bandwith chart below, the secret knowledge the article cannot state ! LOL a review and it cannot state the BANDWITH of NVIDIA's new card! roflmao !
------------------------nvidia
---------------ati
NVIDIA WINS BY A VERY LARGE PERCENTAGE.
konjiki7 - Friday, October 2, 2009 - link
http://www.hardocp.com/news/2009/10/02/nvidia_fake...">http://www.hardocp.com/news/2009/10/02/..._fakes_f...
Samus - Thursday, October 1, 2009 - link
Thats great and all nVidia has more available bandwidth but....they're not anywhere close to using it (much like ATi) so exactly what is your point?SiliconDoc - Friday, October 2, 2009 - link
Wow, another doofus. Overclock the 5870's memory only, and watch your framerates rise. Overclocking the memory increases the bandwith, hence the use of it. If frames don't rise, it's not using it, doesn't need it, and extra is present.THAT DOESN'T HAPPEN for 5870.
-
Now, since FERMI has 40% more T in core, and an enourmous amount of astounding optimizations, you declare it won't use the bandwith, but your excuse was your falsehood about ati not using it's bandwith, which is 100% incorrect.
Let's pretend you meant GT200, same deal there, higher mem oc= more band and frames rise well.
Better luck next time, since you were 100% wrong.
mm2587 - Thursday, October 1, 2009 - link
you do realize the entire point of mentioning bandwidth was to show that both Nvidia and AMD feel that they are not currently bandwidth limited. They have each doubled their number of cores but only increased bandwidth by ~%50. Theres no mention of overall bandwidth because thats not the point that was being made. Just an off hand observation that says "hey looks like everyone feels memory bandwidth wasn't the limitation last time around"Zingam - Thursday, October 1, 2009 - link
ATI has it here and has it now! NVIDIA does not win because on paper I have 50 billion transistors GPU on 1 nm process! I win! ;)You are a retarded fanboy! And I am not. I'd buy what's best for my money.
SiliconDoc - Thursday, October 1, 2009 - link
Behold the FERMI GPU unbeliever !http://www.fudzilla.com/content/view/15762/1/">http://www.fudzilla.com/content/view/15762/1/
That's called, COMPLETED CARD, RUNNING SILICON.
Better luck next time incorrect ignorant whining looner.
siyabongazulu - Friday, October 2, 2009 - link
Do you see any captions on that site? I don't think so. Nowhere does it mention that it's a complete card. So please stop lying because that goes to show how ignorant you are. Any person with a sound mind can and will tell you that it's not a finished product. So come up with something more valid to show and rant about. Sorry that your big daddy Heung hasn't given you your green slime if you like it that way. Just wait on the corner and when he says, GT300 is a go and tests confirm that it trumps 5870 then you can stop crying and suck on that.silverblue - Thursday, October 1, 2009 - link
When's it coming out?I mean, you have all the answers.
SiliconDoc - Thursday, October 1, 2009 - link
Well thanks for the vote of confidence, but yesterday on the launch, according to the author, right ?LOL
Ha, golly, what a pile.