NVIDIA's GeForce GTX 580: Fermi Refined
by Ryan Smith on November 9, 2010 9:00 AM ESTThe GTX 480… it’s hotter, it’s noisier, and it’s more power hungry, all for 10-15% more performance. If you need the fastest thing you can get then the choice is clear, otherwise you’ll have some thinking to decide what you want and what you’re willing to live with in return.
The GeForce GTX 480 and the associated GF100 GPU have presented us with an interesting situation over the last year. On the one hand NVIDIA reclaimed their crown for the fastest single GPU card on the market, and in time used the same GPU to give rise to a new breed of HPC cards that have significantly expanded the capabilities of GPU computing. On the other hand, like a worn and weary athlete finally crossing the finish line, this didn’t come easy for NVIDIA. GF100 was late, and the GTX 480 while fast was still hot & loud for what it was.
Furthermore GTX 480 and GF100 were clearly not the products that NVIDIA first envisioned. We never saw a product using GF100 ship with all of its SMs enabled – the consumer space topped out at 15 of 16 SMs, and in the HPC space Tesla was only available with 14 of 16 SMs. Meanwhile GF100’s younger, punchier siblings put up quite a fight in the consumer space, and while they never were a threat to GF100, it ended up being quite the surprise for how close they came.
Ultimately the Fermi architecture at the heart of this generation is solid – NVIDIA had to make some tradeoffs to get a good gaming GPU and a good compute GPU in a single product, but it worked out. The same can’t be said for GF100, as its large size coupled with TSMC’s still-maturing 40nm process lead to an unwieldy combination that produced flakey yields and leaky transistors. Regardless of who’s ultimately to blame, GF100 was not the chip it was meant to be.
But time heals all wounds. With GF100 out the door NVIDIA has had a chance to examine their design, and TSMC the chance to work the last kinks out of their 40nm process. GF100 was the first Fermi chip, and it would not be the last. With a lesson in hand and a plan in mind, NVIDIA went back to the drawing board to fix and enhance GF100. The end result: GF110, the next iteration of Fermi. Hot out of the oven, it is launching first in the consumer space and is forming the backbone of the first card in NVIDIA’s next GeForce series: GeForce 500. Launching today is the first such card, the GF110-powered GeForce GTX 580.
GTX 580 | GTX 480 | GTX 460 1GB | GTX 285 | |
Stream Processors | 512 | 480 | 336 | 240 |
Texture Address / Filtering | 64/64 | 60/60 | 56/56 | 80 / 80 |
ROPs | 48 | 48 | 32 | 32 |
Core Clock | 772MHz | 700MHz | 675MHz | 648MHz |
Shader Clock | 1544MHz | 1401MHz | 1350MHz | 1476MHz |
Memory Clock | 1002MHz (4008MHz data rate) GDDR5 | 924MHz (3696MHz data rate) GDDR5 | 900Mhz (3.6GHz data rate) GDDR5 | 1242MHz (2484MHz data rate) GDDR3 |
Memory Bus Width | 384-bit | 384-bit | 256-bit | 512-bit |
Frame Buffer | 1.5GB | 1.5GB | 1GB | 1GB |
FP64 | 1/8 FP32 | 1/8 FP32 | 1/12 FP32 | 1/12 FP32 |
Transistor Count | 3B | 3B | 1.95B | 1.4B |
Manufacturing Process | TSMC 40nm | TSMC 40nm | TSMC 40nm | TSMC 55nm |
Price Point | $499 | ~$420 | ~$190 | N/A |
GF110 is a mix of old and new. To call it a brand-new design would be disingenuous, but to call it a fixed GF100 would be equally shortsighted. GF110 does have a lot in common with GF100, but as we’ll see when we get in to the design of GF110 it is its own GPU. In terms of physical attributes it’s very close to GF100; the transistor count remains at 3 billion (with NVIDIA undoubtedly taking advantage of the low precision of that number), while the die size is at 520mm2. NVIDIA never did give us the die size for GF100, but commonly accepted values put it at around 530mm2, meaning GF110 is a hair smaller.
But before we get too deep in to GF110, let’s start with today’s launch card, the GeForce GTX 580. GTX 580 is the first member of the GeForce 500 series, giving it the distinction of setting precedent for the rest of the family that NVIDIA claims will soon follow. Much like AMD last month, NVIDIA is on their second trip with the 40nm process, meaning they’ve had the chance to refine their techniques but not the opportunity to significantly overhaul their designs. As a result the 500 series is going to be very familiar to the 400 series – there really aren’t any surprises or miracle features to talk about. So in many senses, what we’re looking at today is a faster version of the GTX 480.
So what makes GTX 580 faster? We’ll start with the obvious: it’s a complete chip. All the L2 cache, all the ROPs, all the SMs, it’s all enabled. When it comes to gaming this is as fast as GF110 can be, and it’s only through NVIDIA’s artificial FP64 limitations that double-precision computing isn’t equally unrestricted. We have wondered for quite some time what a full GF100 chip would perform like – given that GTX 480 was short on texture units, shaders, and polymorph engines, but not ROPs – and now the answer is at hand. From all of this GTX 580 has 6.6% more shading, texturing, and geometric performance than the GTX 480 at the same clockspeeds. Meanwhile the ROP count and L2 cache remains unchanged; 48 ROPs are attached to 768KB L2 cache, which in turn are attached to 6 64bit memory controllers.
The second change of course is clockspeeds. The reference GTX 480 design ran at 700MHz for the core and 924MHz (3696MHz data rate) for the GDDR5. Meanwhile GTX 580 brings that up to 772MHz for the core and 1002MHz (4008MHz data rate), marking a 72MHz(10%) increase in core clockspeed and a slightly more modest 78MHz (8%) increase in memory bandwidth. This is a near-equal increase in the amount of work that GTX 580 can process and the amount of work its memory can feed it, which should offer a relatively straightforward increase in performance.
Last but certainly not least change coming from GTX 480 is in GF110 itself. NViDIA has ported over GF104’s faster FP16 (half-precision) texture filtering capabilities, giving GF110/GTX580 the ability to filter 4 FP16 pixels per clock, versus 2 on GF100/GTX480. The other change ties in well with the company’s heavy focus on tessellation, with a revised Z-culling/rejection engine that will do a better job of throwing out pixels early, giving GF110/GTX580 more time to spend on rendering the pixels that will actually be seen. This is harder to quantify (and impossible for us to test), but NVIDIA puts this at another 8% performance improvement.
Meanwhile NVIDIA hasn’t ignored GTX 480’s hot and loud history, and has spent some time working on things from that angle. We’ll dive in to NVIDIA’s specific changes later, but the end result is that through some optimization work they’ve managed to reduce their official TDP from 250W on the GTX 480 to 244W on the GTX 580, and in practice the difference is greater than that. NVIDIA’s cooling system of choice has also been updated, working in tandem with GTX 580’s lower power consumption to bring down temperatures and noises. The end result is a card that should be and is cooler and quieter while at the same being faster than GTX 480.
GF110
The downside to this is that if it sounds like a fairy tale, it almost is. As you’ll see we have a rather high opinion of GTX 580, but we’re not convinced you’re going to be able to get one quite yet. NVIDIA is technically hard-launching GTX 580 today at $499 (GTX 480’s old price point), but they aren’t being very talkative about the launch quantity. They claim it’s for competitive reasons (to keep AMD from finding out) and we can certainly respect that, but at the same time it’s rare in this industry for someone to withhold information because it’s a good thing. We really hope to be surprised today and see GTX 580s available for everyone that wants one, but we strongly suspect that it’s going to be available in low quantities and will sell out very quickly. After that it’s anyone’s guess on what the refresh supply will be like; our impression of matters is that yields are reasonable for such a large chip, but that NVIDIA didn’t spend a lot of time stockpiling for today’s launch.
In any case, with GTX 580 taking the $500 spot and GF110 ultimately destined to replace GF100, GF100 based cards are going to be on their way out. NVIDIA doesn’t have an official timeline, but we can’t imagine they’ll continue producing GF100 GPUs any longer than necessary. As a result the GTX 480 and GTX 470 are priced to go, falling between the GTX 580 and the GTX 460 in NVIDIA’s lineup for now until they’re ultimately replaced with other 500 series parts. For the time being this puts the GTX 480 at around $400-$420, and the GTX 470 – still doing battle with the Radeon HD 6870 – is at $239-$259.
Meanwhile AMD does not have a direct competitor for the GTX 580 at the moment, so their closest competition is going to be multi-GPU configurations. In the single card space there’s the Radeon HD 5970, which is destined for replacement soon and as a result AMD is doing what they can to sell off Cypress GPUs by the end of the year. The last reference 5970 you can find on Newegg is a Sapphire card, which is quite blatantly priced against the GTX 580 at $499 with a $30 rebate. Given that it’s the last 5970, we’d be surprised if it was in stock for much longer than the initial GTX 580 shipments.
For cards you do stand a good chance of getting, a pair of 6870s will set you back between $480 and $500, making it a straightforward competitor to the GTX 580 in terms of price. A pair of cards isn’t the best competitor, but CrossFire support is widely available on motherboards so it’s a practical solution at that price.
Fall 2010 Video Card MSRPs | ||
NVIDIA | Price | AMD |
GeForce GTX 580
|
$500 | Radeon HD 5970 |
$420 | ||
|
$300 | Radeon HD 5870 |
$240 | Radeon HD 6870 | |
$180 | Radeon HD 6850 |
160 Comments
View All Comments
RussianSensation - Wednesday, November 10, 2010 - link
Very good point techcurious. Which is why the comment in the review about having GTX580 not being a quiet card at load is somewhat misleading. I have lowered my GTX470 from 40% idle fan speed to 32% fan speed and my idle temperatures only went up from 38*C to 41*C. At 32% fan speed I can not hear the car at all over other case fans and Scythe S-Flex F cpu fan. You could do the same with almost any videocard.Also, as far as FurMark goes, the test does test all GPUs beyond their TDPs. TDP is typically not the most power the chip could ever draw, such as by a power virus like FurMark, but rather the maximum power that it would draw when running real applications. Since HD58/68xx series already have software and hardware PowerPlay enabled which throttles their cards under power viruses like FurMark it was already meaningless to use FurMark for "maximum" power consumption figures. Besides the point, FurMark is just a theoretical application. AMD and NV implement throttling to prevent VRM/MOSFET failures. This protects their customers.
While FurMark can be great for stability/overclock testing, the power consumption tests from it are completely meaningless since it is not something you can achieve in any videogame (can a videogame utilize all GPU resources to 100%? Of course not since there are alwasy bottlenecks in GPU architectures).
techcurious - Wednesday, November 10, 2010 - link
How cool would it be if nVidia added to it's control panel a tab for dynamic fan speed control based on 3 user selectable settings.1) Quiet... which would spin the fan at the lowest speed while staying just enough below the GPU temperature treshold at load and somewhere in the area of low 50 C temp in idle.
2) Balanced.. which would be a balance between moderate fan speed (and noise levels) resulting in slightly lower load temperatures and perhaps 45 C temp in idle.
3) Cool.. which would spin the fan the fastest, be the loudest setting but also the coolest. Keeping load temperatures well below the maximum treshold and idle temps below 40 C. This setting would please those who want to extend the life of their graphics card as much as possible and do not care about noise levels, and may anyway have other fans in their PC that is louder anyway!
Maybe Ryan or someone else from Anandtech (who would obviously have much more pull and credibility than me) could suggest such a feature to nVidia and AMD too :o)
BlazeEVGA - Wednesday, November 10, 2010 - link
Here's what I dig about you guys at AnandTech, not only are your reviews very nicely presented but you keep it relevant for us GTX 285 owners and other more legacy bound interested parties - most other sites fail to provide this level of complete comparison. Much appreciated. You charts are fanatastic, your analysis and commentary is nicely balanced and attention to detail is most excellent - this all makes for a more simplified evaluation by the potential end user of this card.Keep up the great work...don't know what we'd do without you...
Robaczek - Thursday, November 11, 2010 - link
I really liked the article but would like to see some comparison with nVidia GTX295..massey - Wednesday, November 24, 2010 - link
Do what I did. Lookup their article on the 295, and compare the benchmarks there to the ones here.Here's the link:
http://www.anandtech.com/show/2708
Seems like Crysis runs 20% faster at max res and AA. Is a 20% speed up worth $500? Maybe. Depends on how anal you are about performance.
lakedude - Friday, November 12, 2010 - link
Someone needs to edit this review! The acronym "AMD" is used several places when it is clear "ATI" was intended.For example:
"At the same time, at least the GTX 580 is faster than the GTX 480 versus AMD’s 6800/5800 series"
lakedude - Friday, November 12, 2010 - link
Never mind, looks like I'm behind the times...Nate007 - Saturday, November 13, 2010 - link
In the end we ( the gamers) who purchase these cards NEED to be be supporting BOTH sides so the AMD and Nvidia can both manage to stay profitable.Its not a question of who Pawns who but more importantly that we have CHOICE !!
Maybe some of the people here ( or MOST) are not old enough to remember the days when mighty
" INTEL" ruled the landscape. I can tell you for 100% fact that CPU's were expensive and there was no choice in the matter.
We can agree to disagree but in the END, we need AMD and we need NVIDIA to keep pushing the limits and offering buyers a CHOICE.
God help us if we ever lose one or the other, then we won't be here reading reviews and or jousting back and forth on who has the biggest stick. We will all be crying and complaining how expense it will be to buy a decent Video card.
Here's to both Company's ..............Long live NVIDIA & AMD !
Philip46 - Wednesday, November 17, 2010 - link
Finally, at the high end Nvidia delivers a much cooler and quiter, one GPU card, that is much more like the GTX 460, and less like the 480, in terms of performance/heat balance.I'm one where i need Physx in my games, and until now, i had to go with a SLI 460 setup for one pc and for a lower rig, a 2GB 460 GTX(for maxing GTA:IV out).
Also, i just prefer the crisp Nvidia desktop quality, and it's drivers are more stable. (and ATI's CCC is a nightmare)
For those who want everything, and who use Physx, the 580 and it's upcoming 570/560 will be the only way to go.
For those who live by framerate only, then you may want to see what the next ATI lineup will deliver for it's single GPU setup.
But whatever you choose, this is a GREAT thing for the industry..and the gamer, as Nvidia delivered this time with not just performance, but also lower temps/noise levels, as well.
This is what the 480, should have been, but thankfully they fixed it.
swing848 - Wednesday, November 24, 2010 - link
Again, Anand is all over the place with different video cards, making judgements difficult.He even threw in a GTS 450 and an HD 4870 here and there. Sometimes he would include the HD 5970 and often not.
Come on Anand, be consistent with the charts.