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.
415 Comments
View All Comments
palladium - Monday, October 5, 2009 - link
Not quite:http://www.dailytech.com/article.aspx?newsid=16410">http://www.dailytech.com/article.aspx?newsid=16410
Scroll down halfway thru the comments. He re-registered as SilicconDoc and barks about his hatred for red roosters (in an Apple-related article!)
johnsonx - Monday, October 5, 2009 - link
that looks more like someone mocking him- Sunday, October 4, 2009 - link
According to this very link http://www.anandtech.com/video/showdoc.aspx?i=3573...">http://www.anandtech.com/video/showdoc.aspx?i=3573... AMD already presented a WORKING SILICON at Computex roughly 4 months ago on June 3rd. So it took roughly 4 and a half months to prepare drivers, infrastructure and mass production to have enoough for the start of Windows 7 and DX11. However, Nvidia wasnt even talking about W7 and DX11 so late Q1 2010 or even later becomes more realistic than december. But there are much more questions ahead: What pricepoint, Clockrates and TDP. My impression is that Nvidia has no clue about this questions and the more I watch this development, the more Fermi resembles to the Voodoo5 Chip and the V6000 card which never made into the market because of its much to high TDP.silverblue - Sunday, October 4, 2009 - link
Nah, I expect nVidia to do everything they can to get this into retail channels because it's the culmination of a lot of hard work. I also expect it to be a monster, but I'm still curious as to how they're going to sort out mainstream options due to their top-down philosophy.That's not to say ATI's idea of a mid-range card that scales up and down doesn't have its flaws, but with both the 4800 and 5800 series, there's been a card out at the start with a bona fide GPU with nothing disabled (4850, and now 5870), along with a cheaper counterpart with slower RAM and a slightly handicapped core (4830/5850). Higher spec single GPU versions will most likely just benefit from more and/or faster RAM and/or a higher core clock, but the architecture of the core itself will probably be unchanged - can nVidia afford to release a competing version of Fermi without disabling parts of the core? If it's as powerful as we're lead to believe, it will certainly warrant a higher price tag than the 5870.
Ahmed0 - Saturday, October 3, 2009 - link
Nvidia wants it to be the jack of all trades. However, they are risking with being an overpriced master of none. Thats probably the reason they give their cards more and more gimmicks to play with each year. They are hoping that the cards value will be greater than the sum of its parts. And that might even be a successful strategy to some extent. In a consumerist world, reputation is everything.They might start overdoing it at some point though.
Its like mobile phones nowadays. You really dont need to have a radio, an mp3-player, a camera nor other such extras in it (in fact, my phone isnt able to do anything but call and send messages). But unless you have these features, you arent considered as competition. It gives you the opportunity to call your product "vastly superior" even though from a usability standpoint it isnt.
SymphonyX7 - Saturday, October 3, 2009 - link
Ahh... I see where you're coming from. I've had many classmates who've asked me what laptop to buy and they're always so giddy when they see laptops with the "Geforce" sticker and say they want it cause they want some casual gaming. Yes, even if the GPU is a Geforce 9100M. I recommended them laptop using AMD's Puma platform and many of them ask if that's a good choice (unfortunately here, only the Macbook has a 9400M GPU and it's still outside many of my classmates' budgets). Seems like brand awareness of Nvidia amongst many consumers is still much better than AMD/ATI's. So it's an issue of clever branding then?Lifted - Saturday, October 3, 2009 - link
A little late for any meaningful discussion over here as AT let the trolls go for 40 or so pages. I doubt many people can be arsed to sort through it now, so you'd be better off going to a forum for a real discussion of Fermi.neomocos - Saturday, October 3, 2009 - link
if you missed it then here you go ... happy day for all of us :quote from comment posted on page 37 by Pastuch
" Below is an email I got from Anand. Thanks so much for this wonderful site.
-------------------------------------------------------------------
Thank you for your email. SiliconDoc has been banned and we're accelerating the rollout of our new comments rating/reporting system as a result of him and a few other bad apples lately.
A- "
james jwb - Saturday, October 3, 2009 - link
Some may enjoy it, but this unusual freedom that blatant trolls using aggressive, rude language are getting lately is making a mockery of this site.I don't mind it going on for a while, even 20 pages tbh, it is funny, but at some point i'd like to see a message from Gary saying, "K, SiliconDoc, we've laughed enough at your drivel, tchau, banned! :)"
That's what i want to see after reading through 380 bloody comments, not that he's pretty much gotten away with it. And if he has finally been banned, i'd actually love to know about it in the comments section.
/Rant over.
Gary Key - Monday, October 5, 2009 - link
He is gone as are a couple of others. We have a new comments system in final development now that should take care of this problem in the future.