We've just witnessed the last days of large, single chip GPUs
Posted on 31st Mar 2010 at 11:15 by Richard Swinburne with 33 comments
Just as with the 65nm manufacturing process used for GT200, I'm certain Nvidia overestimated what the 40nm node would offer when it first designed Fermi and that this miscalculation has played a huge part in the fact that the GeForce GTX 480 is so hot and uses so much power.
We know each major architecture change for GPU development takes at least a few years to mash out, so fabless companies such as Nvidia need to guess where fabrication partners - TSMC in this case - will be.
As it stands, TSMC has had more than a rough year with its 40nm node and there's been considerable stress for both ATI and Nvidia - however, to ATIs advantage, it started on 40nm with the Radeon HD 4770. It's clearly not forgotten the lesson came at the expense of the HD 2900 XT, which first arrived on a massive 80nm die, before being respun into TSMCs then upcoming 55nm node at a more digestible price.
Luckily for Nvidia the GTX 480 isn't quite up to the HD 2900 XT par of failures; at least it's faster than the previous generation and, negating the lateness and practical engineering issues, the die size and power use are truly massive.
The thing is, TSMC hasn't yet demonstrated how commercially viable its next fabrication node (likely 28nm) is, so while we fully expect Nvidia to 'pull an ATI HD 3000 series' in six months time and re-do Fermi with a smaller process, resulting in a much more power efficient GPU, TSMC's troubles - and the fact Nvidia is desperate for a new process - means Nvidia is likely looking to Global Foundries, the manufacturing firm spun off from AMD last year.
The heat and power consumption of the GTX 470 and 480 mean its highly unlikely we'll see a dual GPU product any time soon and, as much as the nay-sayers claim we cannot compare the "dual GPU" Radeon HD 5970 to the "single GPU" GTX 480, the fact is that these are all products competing for the titles of "fastest graphics card" - and that's a title which isn't leaving AMD any time soon.
While there are many reasons to be pessimistic about Fermi, one lesson Nvidia has learnt is to design Fermi with more modularity in mind, so we won’t have a repeat of the GT200 series, where Nvidia struggled to produce any derivatives for the all-important mainstream market.
We will be seeing more Fermi-based derivatives in the coming months, which Nvidia absolutely needs to nail because ATI has already successfully launched a complete top-to-bottom DirectX 11 range and after a 9 months of cold turkey, 2010 will see which Nvidia partners can take the strain and come out dancing.
For a few, they might actually be better off for less competition. BFG has already left the European market along with many other smaller partners, and EVGA is set for a resurgence. Things could get interesting.
Looking to the future, I wonder if this will be the very last big GPU we ever see. After two big and hot GPUs, will Nvidia abandon this method of design in favour of the ATI-esque (or should we say, Voodoo-esque) route of multi-chip graphics card design? Dare I suggest, does it even care?
Are Nvidia’s ambitions reallyto now concentrate on CUDA applications in the HPC market where big money is to be made? It’s certainly a huge growth area compared to the relatively mature PC gaming market.
Like so many of us, I never want to see PC gaming die, but in my opinion the days of multi-billion transistor single chip graphics cards are practically over.
We know each major architecture change for GPU development takes at least a few years to mash out, so fabless companies such as Nvidia need to guess where fabrication partners - TSMC in this case - will be.
As it stands, TSMC has had more than a rough year with its 40nm node and there's been considerable stress for both ATI and Nvidia - however, to ATIs advantage, it started on 40nm with the Radeon HD 4770. It's clearly not forgotten the lesson came at the expense of the HD 2900 XT, which first arrived on a massive 80nm die, before being respun into TSMCs then upcoming 55nm node at a more digestible price.
Luckily for Nvidia the GTX 480 isn't quite up to the HD 2900 XT par of failures; at least it's faster than the previous generation and, negating the lateness and practical engineering issues, the die size and power use are truly massive.
The thing is, TSMC hasn't yet demonstrated how commercially viable its next fabrication node (likely 28nm) is, so while we fully expect Nvidia to 'pull an ATI HD 3000 series' in six months time and re-do Fermi with a smaller process, resulting in a much more power efficient GPU, TSMC's troubles - and the fact Nvidia is desperate for a new process - means Nvidia is likely looking to Global Foundries, the manufacturing firm spun off from AMD last year.
The heat and power consumption of the GTX 470 and 480 mean its highly unlikely we'll see a dual GPU product any time soon and, as much as the nay-sayers claim we cannot compare the "dual GPU" Radeon HD 5970 to the "single GPU" GTX 480, the fact is that these are all products competing for the titles of "fastest graphics card" - and that's a title which isn't leaving AMD any time soon.
While there are many reasons to be pessimistic about Fermi, one lesson Nvidia has learnt is to design Fermi with more modularity in mind, so we won’t have a repeat of the GT200 series, where Nvidia struggled to produce any derivatives for the all-important mainstream market.
We will be seeing more Fermi-based derivatives in the coming months, which Nvidia absolutely needs to nail because ATI has already successfully launched a complete top-to-bottom DirectX 11 range and after a 9 months of cold turkey, 2010 will see which Nvidia partners can take the strain and come out dancing.
For a few, they might actually be better off for less competition. BFG has already left the European market along with many other smaller partners, and EVGA is set for a resurgence. Things could get interesting.
Looking to the future, I wonder if this will be the very last big GPU we ever see. After two big and hot GPUs, will Nvidia abandon this method of design in favour of the ATI-esque (or should we say, Voodoo-esque) route of multi-chip graphics card design? Dare I suggest, does it even care?
Are Nvidia’s ambitions reallyto now concentrate on CUDA applications in the HPC market where big money is to be made? It’s certainly a huge growth area compared to the relatively mature PC gaming market.
Like so many of us, I never want to see PC gaming die, but in my opinion the days of multi-billion transistor single chip graphics cards are practically over.
33 Comments
Discuss in the forums ReplyMakes me chuckle..only Nvidia could make a 40nm based card more power hungry and hot than its predecessor.
What has a change to multi chip GPUs got to do with PC gaming dying? If all they make are multi chip GPUs, wouldn't the support for them just get better?
More money has to be put into driver development and more money has to be put into graphics card design. With the cards stacked against PC gaming already in some respects, and each generation of graphics cards having less and less of a performance jump from the last - are we already hitting a wall?
And why is it Global Foundries to NVidia's rescue? (ATI'd rejoice I guess, GloFo is still full of "old" AMD people)
What Node is GloFo at? 45nm? (Opterons mostly I guess)
Even intel is just at 32nm for something as complex as a processor (which a GPU is...if not more complex)
how about specifically design a bus for multi-GPU? i like 4870x2's side-bus (??) and i think that's the way forward for higher/better performance.
It is unfortunate that we are unlikely to see any of this until after a new generation of consoles comes out and raises the floor under games graphics. PC gaming isn't dying, it just no longer leads.
then ATI will have no competition
and without it no real will to push the graphics boundry further.
yes i prefer a single gpu always will but i think like the blob person we have seen the last of them
Or just use a single 5870 and get 85% of the performance....
Anyways no, Nvidia wont learn from this because they still think Fermi is the best thing since slice bread. Jen-Hsun Huang is borderline delusional in his own brand fanboyism, hell he still thinks nVidia "make the best chipsets in the world"....
I think there is a change of focus coming, perhaps Nvidia is going to concentrate more on the hpc market and this will drive development and a derivative of the hpc product will be used for gaming....very much as fermi is.
Tbh perhaps thats the way it should be...it has always struck me as a little frivolous that "gaming" should be a major driving force in computer hardware development.
Infact it may even be advantageous as regardless of the ebb and flow of demand for pc gaming hardware there will always be demand from the hpc sector...so provided there is sufficient demand to make a gaming derivative of a hpc product then hardware pc gaming hardware will continue to develop.
The future of GPUs clearly lies in smaller and more efficient multi-core designs, just as it does for CPUs. Yes, it'll take a paradigm-shift and the learning curve for optimizing code for multi-core solutions might be a bit steep, but it's clearly the way ahead.
Graphic cards purely for gaming as their main raison detré? Games are (with the honourable exception of Crysis) eaten alive by by most mid+ cards of the last two years. (Think 4890/275 or higher) . It is game developersthat have to develop games that DRIVE users to want up grade...Ati/Nvidia are on a lost cause if they think people will continually up grade just to have the latest card if their exsisting card remains more than adequate. Nvidia have bought games to show off physX. Wasted effort!. Ati (and Nvidia) should be paying them to make games only properly playable on the best of the last gen cards so people buy the next gen and the next gen is worth developing.
Some of the issues for games and drivers is down to the way the drivers are coded, the rest lays in the hand of the developers and the code they write. Far too many people are too keen to point to bad drivers when a lot of the time the problem lays within the software itself and is why on the PC platform we have the need for patches, something the consoles don't get because they spend more time testing for problems.
It is a misconception that the sheer variety of hardware on a pc platform is to blame, but in essence this simply comes down to drivers and the software that uses it. For a simplified understanding consider Adobe's Flash and that it is capable of running on any PC regardless of hardware.
I do agree that multi-core GPU's are going to be the way forward, though ATI have shown that there is still some life left in in single core GPU's for a little while longer yet.
graphics processing is already an extremely parallelizable task. That's why GPU's have hundreds of stream processors and shader units. Those channels already act as a "multi GPU" solution with shared resources like memory links and the like. hence any modern GPU already is the equivalent of a 100+core processor (mind you with a specific instruction sets and shared memory)
correct me if I'm wrong, but what is the benefit of going to multiple discrete dies instead of using one die with twice the transistors? unless we are talking about heat and yields?
multigpu = 4870x2, 3870x2, etc.
multiple discrete dies lets pretty much everyone in the manufacturing process save a bit of time and money as its easier doing it once and slapping two of them onto a PCB instead of doing 2 totally different GPU's. Although its possible to just do this. Use Cypress of example.
Currently its this:
5970 = 2 5870's on the same PCB
5870 = 20 processor clusters (right name) enabled
5850 = 18 clusters enabled.
5830 = 16 or something.
and so on. And theyre all the same GPU with different parts disabled (except 5970). If they wanted to they could call a 5870 a 5970, a 5850 a 5870, and so on and achieve a similar result to what you were talking about. But the reason why they dont take the transistor count and shader count from the 5970 and force it into a 5870 die is because the stupid thing would pull a GTX480/470. (cook itself and blackout the whole on new york).
It seems like ATI waits for a die shrink before doing that. Since the 5870 matches the 4870x2 for shader count but runs cooler and draws less power.
Now on to the topic of the blog.
Big GPU's wont be going anywhere for a while. As long as there are people, stuff like Fermi will keep happening simply because everyone will be like "i hope that new process can keep this thing from flopping" when the process is a few years off to begin with. Still there will be big single GPU's to power our playstations and xbox 360's of the future. However a demand for cool and light-bill friendly GPU's will become necessary especially when flash/internet browsing becomes gpu accelerated.
Fermi + 28nm > 5970?
Not only are the processes getting increasingly smaller, but the industry is stagnating in terms of graphical prowress in games. Of course many I think already foresaw that the GPU size could only be as big until there was no more room, and in this case there isn't.
Think of the CPUs now, 2 cores on 1 die..
multi-core is where there is a number of cores share the same cache, with the same memory controller.
multi-processor is where there are a number of memory controllers, each have their own memory, nothing is shared except for the IO of the system.
due to shared nature of multi-core, current single GPU's can have very effecient scheduling and data can be exchanged on the very fast cache. but with multi-core, data must travel through a form of bus connecting the cores, thus creating bottleneck.
natal and the playstation eye I think it's called are there next big ideas
Sony said they have a 10 year product life cycle does that mean it's 10 years before ps4. Nvidia and Sony have a decent relationship so you would assume they would stick together
graphics can't go that much higher before they are life like there's only so much detail you can add before the game hits reality
Crysis is pretty close to it already. Chracters can already show emotion
3d is the next big thing but i don't think 3d is for pc Market. Most people are still on. 21 inch screen if not below. How much further can graphics truly be pushed. Wipeout HD is still the best looking game out there with god of war 3 pretty close up
both running on a 4-5 year old gpu.
No game bar crysis really requires the 5850 and above unless you go into the crazy resolutions which a bit tech review showed very few use
most people are still at the 1680 mark or even 1280x1024 Tilll everyone is at 1080p/i I dout graphics hardware will ever be pushed
The PS2 wasnt 10 years before the PS3.
+1 To natal.
IF OLED's ever get around to being sold like LCD's it just may be possible to have a whole wall as your monitor. That oughta keep GPU's busy. Who needs Eyefinity? I got a wall!
We haven't even scratched the surface of the possibilities! This is some of what I see in the future of gaming.
Games physics is still incredibly immature. simulation of liquids and gasses and their interaction with the rest of the world is immature or non-existent. Where it exists, the methods are not "physics" based, but done with a few formulae that captures the general behaviour, but without much of the subtly of the real world. Particle systems are still using fairly coarse approximations to model gases and the simulation of chemical behaviours is only just appearing. Think about simulating game physics at the atomic level...
Modelling is still at the level feature films were a decade or more ago. They are using models with billions of polygons. Gaming models are a joke in comparison. The use of tessellation in the new generation chips is a great step, but only one more step along a very long path. The sort of model detail seen in Avatar is where gaming will hopefully be in the medium term.
Game AI is also in it's infancy, even compared to the current state of the art, but research in AI in general is accelerating at least as fast as the rest of the information industry, and game designers will keep racing to use whatever the researchers come up with. We will see smarter, more subtle opponents and allies. They'll have more detailed memories and intelligent behaviours, both individually and cooperatively and the ability to come up with novel solutions to problems.
I can picture something like a simple field in spring in a future game. You'd be seeing literally billions of individually animated blades of grass, leaves, flowers and so on. All of them would move independently in the breeze as they do in the real world. There would be dew on the grass, and if you looked closely, all of it would be refractive. The soil the plants are growing in would consist of separate grains of sand, bits of organic matter and small stones. Your footprints in the soil and the crushed bits of grass would be the same, with the blades of grass slowly straightening up after you pass. The dew would be making your trouser legs damp, causing them to cling to your legs a bit as you walked. The field would also be populated with the variety of insects and other small animals you would see in a real field. These would all act like their real cousins, eating the grass or each other.
That's just what I came up with in 5 minutes, without even considering actors, their behaviour and their interactions with the environment, each other and the player(s). I'm sure the designers and other creative types that build our games could add far more.
We are at the dawn of gaming and virtual environments, compare Crysis with Wolfenstein and extrapolate a couple of decades with the exponential increases we've seen so far in all aspects of gaming. It's also likely that entirely new kinds of game will be developed given the possibilities that vast increases in processing power will available. After all, before Wolfenstein, there were no essentially FPS's as we know them today. The power of the '386 generation made something entirely new possible.
1080 HD is the standard right now, but future standards are already being developed in the labs with 5-10 times the resolution we have now. Think about a wrap around screen about twice the size of a 30", but with resolution like a photo, maybe a 0.025 mm dot pitch or less.
We'll also need two frames for whatever type of 3D is used! That was the future only a couple of years ago and now lots of new products are in 3D, media and hardware. It'll be standard in another couple.
Big hi-res screens/3D displays, models with billions or trillions of polygons, accurate simulation of reflection, refraction, diffraction, diffusion, radiosity, and scattering of light will need enormous amounts of processing power from future generations of GPUs. Other advances will need physics processors perhaps AI processors and more.
Future developments in fabrication and design will allow more gates to be built on the same area of a chip. Moving to new materials and technologies will provide more grunt still.
And, still on the far horizon, but definitely out there, is quantum computing of various sorts. Some of the concepts are pretty esoteric, but there are others, based on single electron gates and circuits(which use single electrons to represent a bit) and other physical phenomena like spin, that are much closer to being realised.
Will this ultimate processor be a single chip/object? Perhaps not, and certainly there will be generations where multiple chips are used, probably LOTS of chips. But, I think there will be many single chip solutions in our future.
One reason is that it's always going to be faster to send a signal across a single chip that off one, across a wire/PCB trace/optical fibre, and back into another chip. As speeds increase, distance becomes a serious issue. The smaller your processor, the faster it can work.
Scientists and engineers have been developing the idea of truly 3D processor/chip structures where you stack up many layers of circuitry vertically for years, and some chips now used multi-level structures where several transistors are stacked on top of one another. Potentially there could be as many layers as there are gates across the width of a chip. So, tens of thousands of layers might be possible. I'd call this a single GPU.
We will never stop inventing and developing newer smarter technology and applying it to entertainment. Gaming is, and I believe will remain, one of the primary motivating factors of this progress. Maybe just behind sex!
People love to play!
Consoles have fixed graphical capabilities and usually have a longer life span that your average PC hardware. This is a problem, because when developing for consoles with their limited capabilities there's no incentive to invest time and effort in pushing the envelope for PC hardware.
Ergo, as things stand, you won't need state-of-the-art graphics hardware on the PC to play a simple console port. Which is what most games these days are.
Maybe ATI and/or Nvidia need to promote PC gaming? How about some commercials on TV? PC gaming could use a champion, it only has it's fans.
Eventually some devs are going to wake up and tap into top of the line PC hardware. The console market is getting oversaturated IMHO. They will have to advance the games physics and graphics if they want to survive and the PC will be their vehicle.
As matter of fact, 5970 = 2 5850's, not 2 5870. That, probably, will be a 5980 or 5990 if they get their 2 core graphics under the 5900 label. We're yet to see an ATi 2x5870.