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
Voo - Saturday, October 3, 2009 - link
You may overseen it, but there was a edit by an administrator to one of his posts which did exactly what you want ;)james jwb - Sunday, October 4, 2009 - link
that's good to hear :)Hxx - Friday, October 2, 2009 - link
By the looks of it, Nvidia doesn't have much going on for this year. If they loose the DX11 boat against ATI then I will pity their stockholders. About the only thing that makes those green cards attractive is their Physics spiel. Now if ATI would hurry up and do somethin with that Havoc, then dark days will await Nvidia. One way or the other, its a win-win for the consumer. I just wish their AMD division would fare just as well against intel.Zool - Friday, October 2, 2009 - link
I dont wont to be too pesimistic but availability in Q1 2010 is lame late. Windows 7 will come out soon so people will surely want to upgrade to dx11 till christmas. Also OEM market which is actualy the most profitable. Dell, HP and others will hawe windows 7 systems and they will of course need dx11 cards till christmas.(amd will hawe hopefully all models out till that time)Than of course dx11 games that will come out in future can be optimized for radeon 5K now while for gt300 we dont even know the graphic specs and the only working silicon dont even resemble to a card.
Very bad timing for nvidia this time that will give amd a huge advantage.
Zool - Friday, October 2, 2009 - link
Actualy this could hapen if u merge a super gpgpu tesla card and a GPU and want to sell it as one("because designing GPUs this big is "fucking hard"). Average people (maybe 95% of all) dont even know what Megabyte or bit is not even GPGPU. They will want to buy a graphic card not cuda card.If amd and microsoft will make heawy DX11 pr than even the rest of nvidias gpus wont sell.
PorscheRacer - Friday, October 2, 2009 - link
As with anything hardware, you need the killer software to have consumers want it. DX11 is out now, so we have Windows 7 (which most people are taking a liking to, even gamers) and you have a few upcoming games that people look to be interested in. For GPGPU and all that, well... What do we have as a seriously awesome application that consumers want and feel they need to go out and buy a GPU for? Some do that for F@H and the like, and a few for transcoding video, but what else is there? Until we see that, it's going to be ahrd to convince consumers to buy that GPU. As it is, most feel IGP is good enough for them...PorscheRacer - Friday, October 2, 2009 - link
Actually, thinking about this... Maybe if they were able to put a small portion of this into IGP, and include some good software with it, maybe the average consumer could see the benefits easier and quicker and be inclined to go for that step up to a dedicated GPU?RXR - Friday, October 2, 2009 - link
DocSilicon, you are one funny as hell mental patient to be!. I really hope you dont get banned. You just made reading the comments a whole lot more fun. Plus, it's win win. You get to satisfy your need to go completely postal at everyone, and we get a funny sideshow.- Friday, October 2, 2009 - link
Great words but nothing behind! Fermis is Nvidias Prescott or should I say much like the last Voodoo chip that never really appeared on the market? Too many transistors are not good ...ioannis - Friday, October 2, 2009 - link
Although the Star Trek TNG reference is ok, 'Nexus' should have been accompanied by a Blade Runner reference instead, Nexus-6 :)