Architecting Fermi: More Than 2x GT200

NVIDIA keeps referring to Fermi as a brand new architecture, while calling GT200 (and RV870) bigger versions of their predecessors with a few added features. Marginalizing the efforts required to build any multi-billion transistor chip is just silly, to an extent all of these GPUs have been significantly redesigned.

At a high level, Fermi doesn't look much different than a bigger GT200. NVIDIA is committed to its scalar architecture for the foreseeable future. In fact, its one op per clock per core philosophy comes from a basic desire to execute single threaded programs as quickly as possible. Remember, these are compute and graphics chips. NVIDIA sees no benefit in building a 16-wide or 5-wide core as the basis of its architectures, although we may see a bit more flexibility at the core level in the future.

Despite the similarities, large parts of the architecture have evolved. The redesign happened at low as the core level. NVIDIA used to call these SPs (Streaming Processors), now they call them CUDA Cores, I’m going to call them cores.

All of the processing done at the core level is now to IEEE spec. That’s IEEE-754 2008 for floating point math (same as RV870/5870) and full 32-bit for integers. In the past 32-bit integer multiplies had to be emulated, the hardware could only do 24-bit integer muls. That silliness is now gone. Fused Multiply Add is also included. The goal was to avoid doing any cheesy tricks to implement math. Everything should be industry standards compliant and give you the results that you’d expect.

Double precision floating point (FP64) performance is improved tremendously. Peak 64-bit FP execution rate is now 1/2 of 32-bit FP, it used to be 1/8 (AMD's is 1/5). Wow.

NVIDIA isn’t disclosing clock speeds yet, so we don’t know exactly what that rate is yet.

In G80 and GT200 NVIDIA grouped eight cores into what it called an SM. With Fermi, you get 32 cores per SM.

The high end single-GPU Fermi configuration will have 16 SMs. That’s fewer SMs than GT200, but more cores. 512 to be exact. Fermi has more than twice the core count of the GeForce GTX 285.

  Fermi GT200 G80
Cores 512 240 128
Memory Interface 384-bit GDDR5 512-bit GDDR3 384-bit GDDR3

 

In addition to the cores, each SM has a Special Function Unit (SFU) used for transcendental math and interpolation. In GT200 this SFU had two pipelines, in Fermi it has four. While NVIDIA increased general math horsepower by 4x per SM, SFU resources only doubled.

The infamous missing MUL has been pulled out of the SFU, we shouldn’t have to quote peak single and dual-issue arithmetic rates any longer for NVIDIA GPUs.

NVIDIA organizes these SMs into TPCs, but the exact hierarchy isn’t being disclosed today. With the launch's Tesla focus we also don't know specific on ROPs, texture filtering or anything else related to 3D graphics. Boo.

A Real Cache Hierarchy

Each SM in GT200 had 16KB of shared memory that could be used by all of the cores. This wasn’t a cache, but rather software managed memory. The application would have to knowingly move data in and out of it. The benefit here is predictability, you always know if something is in shared memory because you put it there. The downside is it doesn’t work so well if the application isn’t very predictable.

Branch heavy applications and many of the general purpose compute applications that NVIDIA is going after need a real cache. So with Fermi at 40nm, NVIDIA gave them a real cache.

Attached to each SM is 64KB of configurable memory. It can be partitioned as 16KB/48KB or 48KB/16KB; one partition is shared memory, the other partition is an L1 cache. The 16KB minimum partition means that applications written for GT200 that require 16KB of shared memory will still work just fine on Fermi. If your app prefers shared memory, it gets 3x the space in Fermi. If your application could really benefit from a cache, Fermi now delivers that as well. GT200 did have an L1 texture cache (one per TPC), but the cache was mostly useless when the GPU ran in compute mode.

The entire chip shares a 768KB L2 cache. The result is a reduced penalty for doing an atomic memory op, Fermi is 5 - 20x faster here than GT200.

A Different Sort of Launch A More Efficient Architecture
Comments Locked

415 Comments

View All Comments

  • SiliconDoc - Thursday, October 1, 2009 - link

    Plenty hard, but they GOT HER DONE, and here is the pic of her

    http://www.fudzilla.com/content/view/15762/1/">http://www.fudzilla.com/content/view/15762/1/

    Yes, now about that fantasy paper anand was spewing on - yes he won't get one for two months, but AS I SAID, WE ALREADY KNOW IT BEATS the ati epic failure.
  • rennya - Thursday, October 1, 2009 - link

    Where can I get that GPU? At least at my place I can get a 5870 GPU if I want to, but not so for this GPU.
  • SiliconDoc - Thursday, October 1, 2009 - link

    Well go get one.

    Now were down to the launch and paper lies in this article, were lies, as I've said. Bigger lies by the red texters. If I were Anand I'd be giggling at you fools.
  • rennya - Thursday, October 1, 2009 - link

    Newegg link please. Or any other online retailer websites for the matter.

    Did I already said to stop it with the paper launch already? That only exists in your dreams you know. Or maybe America. But such thing is not true here. Just because America doesn't have enough unit it doesn't mean it is true everywhere else.
  • SiliconDoc - Thursday, October 1, 2009 - link

    Oh, so sorry mi' lady, here is your newegg link, you'll see 2 greyted out 5850's and THAT'S IT.
    http://www.newegg.com/Product/ProductList.aspx?Sub...">http://www.newegg.com/Product/ProductLi...10679497...
    Can't buy em. Paper, e-paper in this case, digital nothing.
    --
    Now if we only could send you some money, and you could trot over to your imaginary shops.... gee if only someone gave you some money, you could HAVE A PICTURE, because veryone knows handy little cameras are BANNED there, huh.
    Gee, all those walks to work.. and not ten seconds to take a pic, and now it's really, too late LOL
    ahhahahahahaaa
    --
    More of that wonderful "red rooster evidence".
  • ClownPuncher - Thursday, October 1, 2009 - link

    I bought one, 5850 that is. They are popular, so they sell out. The same thing happened to me when the 8800GT's launched, bought 2 and they were sold out 15 minutes later.

    If people can buy them, it isn't a paper launch. Give it up. There were cards for sale from many etailers on launch day for the 5870 as well.

    You're saying the 8800GT and 8800GTS g92 were paper launches also? They were selling out in minutes.
  • SiliconDoc - Friday, October 2, 2009 - link

    Funny how you wait for the perfect post to claim your lie is true.
    You're a pure troll, nothing more, in every single post you've made here.
    Of course I know you're lying.
  • rennya - Friday, October 2, 2009 - link

    You are the one who are lying by claiming that 5870 is a paper launch, when availability at my place is pretty good. Then you claim that I do not actually come from a SE Asia country, but admins in this site can easily verify my IP and see where I come from. Accusing people of lying will not make you look good.
  • SiliconDoc - Thursday, October 1, 2009 - link

    Neither does europe, nor africa, nor SA, nor the ME, apparently the only spot is your walk to work. Congratulations, you're at ground zero. Just think how lucky you are.
  • rennya - Friday, October 2, 2009 - link

    I am not the who claims GT300 is available, you do with your fudzilla link. And that picture may only have mock-ups because nVidia doesn't have any working demo.

    At least Intel with its Larabee did showcase their unimpressive raytracing demo with Larabee in IDF.

Log in

Don't have an account? Sign up now