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
samspqr - Thursday, October 1, 2009 - link
ATI's availability will be sorted out soon, NVIDIA's weird design choices that are targeted at anything but graphics won'tin fact, I have just realized: NVIDIA IS DOING A MATROX!
(forget about graphics, concentrate in a proffessional niche, subsequently get run over by competitors in its former main market... eventually dissappear from the graphics market or become irrelevant? with some luck, RayTracing will be here sooner rather than later, ATI will switch to GPUcomputing at the right time -as opposed to very much too soon-, and we will have a 3 players market; until then, ATI domination all over)
andrihb - Thursday, October 1, 2009 - link
What a huge leap of the imagination :Psamspqr - Friday, October 2, 2009 - link
sorry, I was just trying to imagine how many weird things would have to happen so that we don't have a single GPU maker in the marketin any case, if you want some imaginative thinking, try here:
http://www.semiaccurate.com/2009/10/01/nvidia-fake...">http://www.semiaccurate.com/2009/10/01/nvidia-fake...
(I'm not sure yet who is the one making stuff up -charlie or nvidia-, but so far my bet would be on nvidia)
mindless1 - Saturday, October 3, 2009 - link
What they may have done is take an existing PCB design for something else, and tacked down the parts and air-wired them. It is a faster way to debug a prototype, as well as just drilling a few holes and putting makeshift screws in to test a cooling design before going to the effort of the rest of the support parts before you know if the cooling subsystem is adequate.IF that is the situation, I feel nVidia should have held off until they were further along with the prototypes, but when all is said and done if they can produce performance in line with the expectations, that would prove they had a working card.
IGoodwin - Friday, October 2, 2009 - link
First off, I don't know the truth about a fake or real Tesla being in existence; however, when an article shows a strong emotional bias, I do find it hard to accept the conclusions.Here is a link to the current Tesla product for sale online:
http://www.tigerdirect.com/applications/SearchTool...">http://www.tigerdirect.com/applications...tails.as...
This clearly shows the existing Tesla card with screws on the end plate. Also, if memory serves, having partial venting on a single slot for the new Tesla card would equal the cooling available on the ATI card. Also, six-pin connector is in roughly the same place.
As for the PCB, it is hidden on the older Tesla screen shots, so nothing can be derived.
The card may be fake, or not, but Charlie is not exactly unbiased either.
jonGhast - Saturday, October 3, 2009 - link
"but Charlie is not exactly unbiased either."What's the deal with that, I keep trying to read Semi's articles, though his 'tude towards MS and Intel is pretty juvenile, but I've got to ask; did somebody at Nvidia gang rape his mom?
mindless1 - Saturday, October 3, 2009 - link
I simply assume he is either directly or indirectly on ATI's payroll.Fudzilla wrote "The real engineering sample card is full of dangling wires." To display such a card to others they could simply epoxy down some connectors and solder the wires to them.
monomer - Friday, October 2, 2009 - link
Here's an article from Fudo saying that the card was a mock-up. Nvidia claimed it was real at the conference, and are now saying its a fake, but that they really, truly, had a real one running the demos. Really! I completely believe them.http://www.fudzilla.com/content/view/15798/1/">http://www.fudzilla.com/content/view/15798/1/
Yojimbo - Thursday, October 1, 2009 - link
What makes you think it isn't the right time? You can only really tell in hindsight, but you give in your post any reason that you think now is not the right time and later, when amd is gonna do it, is the right time. I think the right time is whenever the architecture is available and the interest is there. Nvidia has, over the past 5 years, been steadily building the architecture for it. Whether the tools are all in place yet and whether the interest is really there remains to be seen.It has nothing to do with matrox or any shift to a "professional niche." Nvidia believes that it has the ability to evolve and leverage its products from the niche sector of 3d graphics into a broader and more ubiquitous computing engine.
wumpus - Thursday, October 1, 2009 - link
Do you see any sign of commercial software support? Anybody Nvidia can point to and say "they are porting $important_app to openCL"? I haven't heard a mention. That pretty much puts Nvidia's GPU computing schemes solely in the realm of academia (where you can use grad students a cheap highly-skilled labor). If they could sell something like a FEA package for pro-engineer or solidworks, the things would fly off the shelves (at least I know companies who would buy them, but it might be more a location bias). If you have to code it yourself, that leaves either academia (which mostly just needs to look at hardware costs) and existing supercomputer users. The existing commercial users have both hardware and software (otherwise they would be "potential users"), and are unlikely to want to rewrite the software unless it is really, really, cheaper. Try to imagine all the salaries involved in running the big, big, jobs Nvidia is going after and tell me that the hardware is a good place to save money (at the cost of changing *everything*).I'd say Nvidia is not only killing the graphics (with all sorts of extra transistors that are in the way and are only for double point), but they aren't giving anyone (outside academia) any reason to use openCL. Maybe they have enough customers who want systems much bigger than $400k, but they will need enough of them to justify designing a >400mm chip (plus the academics, who are buying these because they don't have a lot of money).