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
AtwaterFS - Wednesday, September 30, 2009 - link
4 reals - this dude is clearly an Nvidia shill.Question is, do you really want to support a company that routinely supports this propaganda blitz on the comments of every Fn GPU article?
It just feels dirty doesn't it?
strikeback03 - Thursday, October 1, 2009 - link
I doubt SiliconDoc is actually paid by nvidia, I've met people like this in real life who just for some reason feel a need to support one company fanatically.Or he just enjoys ticking others off. One of my friends while playing Call of Duty sometimes just runs around trying to tick teammates off and get them to shoot back at him.
SiliconDoc - Thursday, October 1, 2009 - link
If facing the truth and the facts makes you mad, it's your problem, and your fault.I certainly know of people like you describe, and let's face it, it is one of YOUR TEAMMATES---
--
Now, when you collective liars and deniars counter one of my pointed examples, you can claim something. Until then, you've got nothing.
And those last 3 posts, yours included, have nothing, except in your case, it shows what you hang with, and that pretty much describes the lies told by the ati fans, and how they work.
I have no doubt pointing them out "ticks them off".
The simple fix is, stop lying.
Yangorang - Wednesday, September 30, 2009 - link
Honestly all I want to know is:When will it launch? (as in be available for actual purchase)
How much will it cost?
Will this beast even fit into my case...and how much power will it use?
How will it perform? (particularly I'm wondering about DX11 games...as it seems to be very much a big deal for ATI)
but heh none of these questions will be answered for a while I guess....
I'm also kinda wondering about:
How does the GT300 handle tessellation?
Does it feature Angle-Independent Anisotropic Filtering?
I could really couldn't give a crap less about using my GPU for general computing purposes....I just want to play some good looking games without breaking the bank...
haukionkannel - Thursday, October 1, 2009 - link
Well it's going to be DX11 card, so it can handle tessalation. How well? That remains to be seen, but there is enough computing power to do it guite nicely.But the big guestion is not, if the GT300 is faster than 5870 or not, It most propably is, but how much and how much it does cost...
If you can buy two 5870 for the prize of GT300, it has to be really fast!
Interesting release and good article to reveal the architecture behind this chip. I am sure, that we will see more new around the release of Win7, even if the card is not released until 2010. Just to make sure, that not too many "potential" customers does not buy ATI made card by that time.
Allso as someone said before this seams to be guite modular, so it's possible to see some cheaper cut down versions allso. We need competition to low and middle range allso. Can G300 design do it reamains to be seeing.
SiliconDoc - Thursday, October 1, 2009 - link
Well, that brings to mind another anandtech LIE.--
In the 5870 article text post area, the article writer and tester, responded to a query by one of the fans, and claimed the 5870 is "the standard 10.5 " .
Well, it is NOT. It is OVER 11", and it is longer than the 285, by a bit.
So, I just have to shake my head, and no one should have wonder why. Even lying about the length of the ati card. It is nothing short of amazing.
silverblue - Thursday, October 1, 2009 - link
http://vr-zone.com/articles/sapphire-ati-radeon-hd...They say 10.5".
SiliconDoc - Thursday, October 1, 2009 - link
I'm sorry, I realize I left with you in the air, since you're so convinced I don't know what I'm talking about." The card that we will be showing you today is the reference Radeon HD 5870, which is a dual-slot graphics card that measures in at 11.1" in length. "
http://www.legitreviews.com/article/1080/2/">http://www.legitreviews.com/article/1080/2/
I mean really, you should have given up a long time ago.
silverblue - Friday, October 2, 2009 - link
Anand, could you or Ryan come back to us with the exact length of the reference 5870, please? I know Ryan put 10.5" in the review but I'd like to be sure, please.It's best to check with someone who actually has a card to measure.
silverblue - Friday, October 2, 2009 - link
You know something? I'm just going to back down and say you're right. You might just be, but I couldn't give a damn anymore.