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
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.
Facebook
Twitter
RSS
415 Comments
View All Comments
Zingam - Thursday, October 1, 2009 - link
Perhaps next gen consoles would be C++ based and not API based (what a great terminology). So in that sense DirectX won't matter like it won't matter on Larrabee because it will be emulated in software. Larrabee would not have any silicon dedicated to OpenGL or DirectX I think.Once GPUs get fast enough I guess they won;t be called Graphics processors anymore and will support APIs as software implementations.
marc1000 - Thursday, October 1, 2009 - link
well, perhaps... I was searching the web yesterday for info on the new consoles, it was kinda sad. if we do not get a new minimun-standard (a powerful console), then the PC games will not be that hard to run on PCs... then my old Radeon 3850 is still capable of running almost ALL games with "good enough" permformance (read: average 30-50fps on most of the console ports to PC).and so: no reason to upgrade! :-(
Dobs - Wednesday, September 30, 2009 - link
Personally I think Eyefinity will be remembered as the master stroke as well as first to implement DirectX 11. Nvidia may get 10 fps more in DirectX 11 in Q2 2010 but will still struggle until it has it's own version of Eyefinity.Current uber-cool for the cashed-up is Eyefinity and once you have 3 or 6 monitors you will only buy hardware that will support it. These 'cashed-up' PC gamers are usually Nvidia's favorite customers.
Nvidia needs flexible wrap-around OLED mega resolution monitors to come out yesterday, but I'm pretty sure that didn't happen... 5850 which supports 3 monitors came out yesterday. :P
SiliconDoc - Thursday, October 1, 2009 - link
Nvidia has supported FOUR monitors on say for instance, the 570i sli, for like YEARS dude.Just put in 2 nv cards, plug em up - and off you go, it's right in the motherboard manuals...
Heck you can plug in two ati cards for that matter.
---
Anyway on the triple monitor with this 5870/50, the drivers are a mess, some having found the patch won't be for a month, then the extra $100 cable is needed, too, as some have mentioned, that ati has not included.
They're pissed.
Dobs - Thursday, October 1, 2009 - link
I'll be avoiding the extra $100 cable by getting DisplayPort monitors from the start. Also want to get ips monitor (i think) so that it will support the portrait mode.If I already had 3 non-DisplayPort monitors, I wouldn't mind shelling out for the DisplayPort adapter if that was my only expense. But if the adapter was flakey I'd be upset as well. I know multi-monitors have been around for years, but they've never been this easy to set-up... Even I could do it :P And the drivers will get better in time, and no doubt future games will look to include Eyefinity as well.
SiliconDoc - Thursday, October 1, 2009 - link
Well I do hope you have good luck and that you and your son enjoy it ( no doubt will if you manage get it), and it would be nice if you can eventually link a pic (likely in some future article text area) just because.I think eyefinity has an inherent advantage, cheaper motherboard possible, 3 on one card, and with 3 from the same card, you can have the concave wrap view going with easy setup.
I agree however with the comment that it won't be a widely used feature, and realize most don't even use both monitor hookups on their videocards that are already available as standard, since long ago, say the 9600se and before.
(I use two though, and I have to say it is a huge difference, and much, much better than one)
yacoub - Wednesday, September 30, 2009 - link
Wake up: 99% of people don't give a crap about Eyefinity. Not only do the VAST, VAST majority of customers have just one display, but those who do have multiple ones (like myself) often have completely different displays, not multiple of the same model and size. And then, even when you find that 0.1% of the customer base that has two or more identical monitors side-by-side, you have to find the ones who game on them. Then of those people, find the ones who actually WANT to have their game screen split across two monitors with a thick line of two display borders right in the middle of their image.Eyefinity is relevant to such an infinitesimally small number of people it is laughable every time someone mentions it like it's some sort of "killer app" feature.
Jamahl - Thursday, October 1, 2009 - link
thats why eyefinity has millions of youtube views already right.yacoub - Thursday, October 1, 2009 - link
because purchases are measured in YouTube views. wow, just... wow.ClownPuncher - Thursday, October 1, 2009 - link
It clearly means people are interested enough to look. Wow, just...wow.You don't like it, other people do. Get over it.