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
SiliconDoc - Thursday, October 1, 2009 - link
Jeezus, you're just that bright, aren't you.The article is dated September 19th, and "they scored a picture" from another website, that "scored a picture".
Our friendly reviewer herer at AT had the cards in his hands, on the bench, IRL.
--
I mean you have like no clue at all, don't you.
palladium - Thursday, October 1, 2009 - link
I agree. GPGPU has come a long way, but it's still in its infancy, at least in the consumer space (Badaboom and AVIVO both had bugs).I just want a card that can play Crysis all very high 19x12 4xAA @60fps. Maybe a dual-GPU GT300 can deliver that.
wumpus - Wednesday, September 30, 2009 - link
First first reaction after reading that the cost of double multiply would be twice that of a single was "great. Half the transistors will be sitting there idle during games." Sure, this isn't meant to be a toy, but it looks like they have given up the desktop graphics to AMD (and whenever Intel gets something working). Maybe they will get volume up enough to lower the price, but there are only so many chips TMSC can make that size.On second thought, those little green squares can't take up half the chip. Any guess what part of the squares are multiplies? Is the cost of fast double point something like 10% of the transistors idle during single (games)? On the gripping hand, makes the claim that "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)". If they seriously mean that they are prepared to include all rounding, all exceptions, and all the ugly, hairy corner cases that inhabit IEEE-754, wait for Juniper. I really mean it. If you are doing real numerical computing you need IEEE-754. If you don't (like you just want a real framerate from Crysis for once) avoid it like the plague.
Sorry about the rant. Came for the beef on doubles, but noticed that quote when checking the article. Looks like we'll need some real information about what "core level at IEEE-754" means on different processors. Who provides all the rounding modes, and what parts get emulated slowly? [side note. Is anybody with a 5870 able to test underflow in OpenCL? You might find out a huge amount about your chip with a single test].
SiliconDoc - Wednesday, September 30, 2009 - link
I think I'll stick with the giant profitables greens proven track record, not your e-weened redspliferous dissing.Did you watch the NV live webcast @ 1pm EST ?
---
Nvidia is the only gpu company with OBE BILLION DOLLARS PER YEAR IN R&D.
---
That's correct, nvidia put into research on the Geforce, the whoile BILLION ati loses selling their crappy cheap hot cores on weaker thinner pcb with near zero extra features only good for very high rez, which DOESN'T MATCH the cheapo budget pinching purchasrs who buy red to save 5-10 bang for bucks...--
--
Now about that marketing scheme ?
LOL
Ati plays to high rez wins, but has the cheapo card, expecting $2,000 monitor owners to pinch pennies.
"great marketing" ati...
LOL
PorscheRacer - Wednesday, September 30, 2009 - link
Just so you know, ATI is a seperate division in AMD (the graphics side obviously) and did post earnings this year. ATI is keeping the CPU side of AMD afloat in all intents and purposes. Is there a way to ban or block you? I was excited to read about the GF300 and expecting some good comments and discussion about this, and then you wrecked the experience. Now I just don't care.Adul - Thursday, October 1, 2009 - link
silicon idiot is doing more harm than good. please ban himSiliconDoc - Thursday, October 1, 2009 - link
The truth is a good thing, even if you're so used to lies that you don't like it.I guess it's good too, that so many people have tried so hard to think of a rebuttal to any or of all my points, and they don't have one, yet.
Isn't that wonderful ! You fit that category, too.
SiliconDoc - Wednesday, September 30, 2009 - link
Do you think yhour LIES will pass with no backup ?" A.M.D. has struggled for two years to return to profitability, losing billions of dollars in the process.
A.M.D., the No. 2 maker of computer microprocessors after Intel, lost $330 million, or 49 cents a share, in the second quarter. In the same period last year, it lost $1.2 billion, or $1.97 a share.
Excluding one-time gains, A.M.D. says its loss was 62 cents a share. On that basis, analysts had predicted a loss of 47 cents a share, according to Thomson Reuters. Sales fell to $1.18 billion, down 13 percent. Analysts were expecting $1.13 billion."
---
http://www.nytimes.com/2009/07/22/technology/compa...">http://www.nytimes.com/2009/07/22/technology/compa...
ATI card sales did increase a bit, but LOST MONEY anyway. More than expected.
--
PS I'm not sorry I've ruined your fantasy and expsoed your lie. If you keep lying, should you be banned for it ?
PorscheRacer - Thursday, October 1, 2009 - link
http://arstechnica.com/hardware/news/2009/07/intel...">http://arstechnica.com/hardware/news/20...-graphic...Again, the graphics group of AMD turned a profit (albeit a small one after R&D and costs) while the other divisions lost money.
SiliconDoc - Thursday, October 1, 2009 - link
LOL- YOU'VE SIMPLY LIED AGAIN, AND PROVIDED A LINK, THAT CONFIRMS YOU LIED.It must be tough being such a slumbag.
--
" After the channel stopped ordering GPUs and depleted inventory in anticipation of a long drawn out worldwide recession in Q3 and Q4 of 2008, expectations were hopeful, if not high that Q1’09 would change for the better. In fact, Q1 showed improvement but it was less than expected, or hoped. Instead, Q2 was a very good quarter for vendors – counter to normal seasonality – but then these are hardly normal times.
Things probably aren't going to get back to the normal seasonality till Q3 or Q4 this year, and we won't hit the levels of 2008 until 2010."
As you should have a clue, noting, 2008 was bad, and they can't even reach that pathetic crash until 2010.
An increase in sales from a recent prior full on disaster decrease, is still less than the past, is low in the present, and is " A LOSS " PERIOD.
You don't provide text because NOTHING at your link claims what you've said, you are simply a big fat LIAR.
Thanks for the link anyway, that links my link:
http://jonpeddie.com/press-releases/details/amd-so...">http://jonpeddie.com/press-releases/det...ntel-and...
This is a great quote: " We still believe there will be an impact from the stimulus programs worldwide "
LOL
hahahhha - just as I kept supposing.
" -Jon Peddie Research (JPR), the industry's research and consulting firm for graphics and multimedia"
---
NOTHING, AT either link, describes a profit for ati graphics, PERIOD.
Try again mr liar.