Sweeney and Richards on the future of 3D

[Harison]

Unfortunately this is not true. Else you would see accelerator cards for all major offline renderers. Some attempts have been made and all of them failed miserably.

Current real time 3D graphics are constrained by what the GPUs can do.

Depends on the area, if HW acceleration is feasible or not, for example professional graphics and video markets strongly depend on accelerating HW cards. Games benefit from HW a LOT as well. DirectX has its limitations, but its far better than fully software based approach at the moment. Its easier for game devs to make games based on DirectX than to create software engine top-to-bottom, yet devs are more and more behind both DirectX versions and HW.

Its not because DirectX/HW are holding devs back (limiting yes, holding back - no), its because as Panda said, we are seeing diminishing returns in graphics. Same happened in audio market, HW and better standards development are halting now simply because there is less and less need for even better solutions, just ask Creative about it. Games havent got there yet, but more and more people will be satisfied simply by integrated GPUs in CPUs, no wonder its predicted there will be only 10% market for discreet graphics in 2015, and demand for discreet will only decrease after that. DirectX is holding back? Hell, some major software powerhouses sticked with DX9, despite DX10 and now 11 are offering some very compelling features.

 

[brain_stew]

I saw these a while ago, a real interesting watch but Sweeney's utter insistence that an unabated drive towards general programmability and software rendering is the only solution in a world where performance per watt is quickly becoming the only metric that matters has always come off as seriously naive and short-sighted viewpoint to me. I realise he's way more qualified to speak on this subject than me but I can't help but feel he's missing the bigger picture. Consumers are expecting their computing and gaming devices to shrink in size, the rise of the netbook, NDS, smartphone and tablet are pretty strong evidence of that. Given those market realities is perusing an approach where performance per watt is sidelined really all that smart?

I think the 3DS vs. the iPhone4 is quite an interesting test case for this theory. The PICA200 in the 3DS has to work within much tighter power constraints and is generally considered a much lower end part compared to the SGX 535 in the iPhone4. It forgoes full unified shaders and ES 2.0 support in order to hard-wire a bunch of the most important and commonly used pixel shader effects to increase efficiency and allow the chip to run within a tighter power budget.

The result? We have numerous 3DS titles that pack more technology/shader effects/per pixel lighting than any previous generation console game and any iOS game, including Epic's own tech demo which doesn't even appear to have proper per pixel lighting. Now I realise there's a much higher abstraction layer on the iPhone and the resolution is higher (although the extra overhead required for S3D rendering kind of offsets that) but the discrepancy is pretty striking considering developers have had access to idevice hardware with ES 2.0 chips for around 18 months now and still games which have minimal usage of even the most basic pixel shader effects are as rare as hen's teeth. Surely the increase in raw throughput should be enough to overcome the higher abstraction layer considering the hardware goes through a much bigger battery, much quicker? Its not an apples-to-apples comparison at all of course but it is looking to be an interesting test case into the benefits that more efficient and fixed function hardware can bring to embedded devices which have to work within strict power constraints.

Now, I'm not suggesting that the next generation of consoles should ditch unified shaders like the 3DS has, not at all, but I do feel that the march towards something approaching the Larrabee school of designis misguided. I've long believed that this current generation of consoles were too big, too noisy, too hot and sucked way too much power for a general consumer electronics device. With that, I'd say a 100W power budget maximum for the entire machine is really at the upper limits of where Sony and Microsoft should be going next generation, at the very least the power draw won't be increased over this generation's launch hardware, that figure is only going down.

How many die shrinks would it take to fit even a less aggressive general purpose design like Fermi into a small box where that 100w figure needs to be shared between a motherboard, CPU, RAM, optical drive, HDD, cooling system and several peripherals? The current GF100 design @ 40nm isn't even remotely close to that sub 50w range and its going to be a long time before it ever is with die shrinks becoming less and less reliable. Will it be there by 2013? Perhaps, but I have my doubts. So does a design that pursues general throughput, programmability and bandwidth above all else to an even more aggressive degree than Fermi really make all that much sense? If clockspeeds are going to have to be utterly massacred in order to fit it into a <50w package? We all know what a disaster cramming GF100 into a notebook turned out to be, Nvidia barely managed to beat out the mobility (5870, basically a desktop 5770 with more modest clocks) despite the chip being manufactured on the same process and coming with a ridiculous doubling of the 5870's TDP. I have fears that adapting a design like that for an embedded device could run into the exact same problems.

I'm just babbling hereI know, I just don't think the usefulness of some fixed function and specialised hardware is going to vanish in a world where efficiency and performance per watt are becoming so crucial to all designs. General computing devices are getting smaller not larger. Those ~300w monsters may be fine for the high end discrete PC space but I doubt they're of much use anywhere else.

Yes, he has accomplished a lot, but it is simply absurd how Sweeney persists that a software render pipeline is the certain future of game rendering.

Maybe there will be some cases where an alternative to classic rasterization could lead to a new pipeline standard. But why in hell would this not be implemented in (a) hardware (standard)? He has completeley lost touch with reality by suggesting that even a marginal amount of people in the industry is interested to create a complete render pipeline from ground up.

Well he's in the rendering engine and middleware business. A world where only a few elite developers like his own Epic, can actually manage to create their own competitive 3D rendering engine on a given piece of hardware is a world where Epic will have an endless supply of willing licensees. The fewer developers that can actually create a modern game engine the more Epic will benefit as it would mean the vast majority of developers would have to place an even greater reliance on middleware provided by the likes of Epic.

I don't think hardware vendors and platform holders are crazy enough to bend to his will (though its naive to simply write off the influence that the industry's primary graphics middleware provider could yield) but its not difficult to see how much he'd personally benefit from a scenario like this becoming reality.

Personally I think Mr. Sweeney is still bitter that Quake 3 ate his lunch despite not offering software rendering support. IIRC there was still software rendering support hidden away in UE2, Sweeney and Epic's unhealthy obsession with software rendering is hardly a new thing.

 

[Rys]

Given those market realities is perusing an approach where performance per watt is sidelined really all that smart?

Performance per watt is never sidelined in any chip containing a graphics processor, and is a primary consideration in every design I can think of. The tradeoffs are basically what define modern graphics architectures.

I think the 3DS vs. the iPhone4 is quite an interesting test case for this theory. The PICA200 in the 3DS has to work within much tighter power constraints and is generally considered a much lower end part compared to the SGX 535 in the iPhone4. It forgoes full unified shaders and ES 2.0 support in order to hard-wire a bunch of the most important and commonly used pixel shader effects to increase efficiency and allow the chip to run within a tighter power budget.

What theory exactly? That chip designers trade off performance and power? You already said that without comparing NDS and the iPhone 4, and it's one of the basic tenets of chip design. Out of interest, do you know the PICA200 has a much lower power budget than the A4? That's interesting data that's not available as far as I'm aware.

The result? We have numerous 3DS titles that pack more technology/shader effects/per pixel lighting than any previous generation console game and any iOS game, including Epic's own tech demo which doesn't even appear to have proper per pixel lighting.

That's just false, even if you limit yourself to previous generation handhelds. I can also assure you that CastleDemo has proper per-pixel lighting, having worked on it. I take it you're just looking at screenshots and making a call on what's going on per-pixel in various situations, which isn't going to work sadly.

Now I realise there's a much higher abstraction layer on the iPhone and the resolution is higher (although the extra overhead required for S3D rendering kind of offsets that) but the discrepancy is pretty striking considering developers have had access to idevice hardware with ES 2.0 chips for around 18 months now and still games which have minimal usage of even the most basic pixel shader effects are as rare as hen's teeth.

There are more games on iOS that use the programmable hardware in a non-trivial manner than there are total 3DS games announced so far. I do agree that there's loads of room for more aggressive use of the power available on the SGX-powered devices though.

Surely the increase in raw throughput should be enough to overcome the higher abstraction layer considering the hardware goes through a much bigger battery, much quicker? Its not an apples-to-apples comparison at all of course

The capabilities of each device are completely different, and each one has a battery befitting the use cases, power and cost.

but it is looking to be an interesting test case into the benefits that more efficient and fixed function hardware can bring to embedded devices which have to work within strict power constraints.

Not really, given the entirety of the rest of the market is ES2 or higher. Nobody else is going to go backwards to fixed function after seeing the 3DS, and only Nintendo could get away with not going with something more programmable in 2010.

Now, I'm not suggesting that the next generation of consoles should ditch unified shaders like the 3DS has, not at all, but I do feel that the march towards something approaching the Larrabee school of designis misguided.

It's not misguided at all, it's completely natural. Give the programmer control where it makes sense to, to let them efficiently express the solution to the problem they're trying to solve. That's why fixed function got out of the way in the first place, because it was the right thing to do to naturally advance graphics. Nintendo don't give two hoots about advancing graphical quality on their consoles beyond the bare minimum required to keep their customers interested. They're several steps behind at this point.

I've long believed that this current generation of consoles were too big, too noisy, too hot and sucked way too much power for a general consumer electronics device. With that, I'd say a 100W power budget maximum for the entire machine is really at the upper limits of where Sony and Microsoft should be going next generation, at the very least the power draw won't be increased over this generation's launch hardware, that figure is only going down.

Well said and I almost totally agree, bar the upper bound being something as arbitrary as 100W. I'm not smart enough to divine that number, but I do agree that it'll go down next time.

How many die shrinks would it take to fit even a less aggressive general purpose design like Fermi into a small box where that 100w figure needs to be shared between a motherboard, CPU, RAM, optical drive, HDD, cooling system and several peripherals? The current GF100 design @ 40nm isn't even remotely close to that sub 50w range and its going to be a long time before it ever is with die shrinks becoming less and less reliable. Will it be there by 2013?

Yes it will. It's at 100W total board power now (106 if memory serves) with a 15x16mm die at 40nm.

So does a design that pursues general throughput, programmability and bandwidth above all else to an even more aggressive degree than Fermi really make all that much sense?

Above all else never happens anyway, so I'm not sure what you're saying. Bandwidth required is just a function of throughput for the most part.

I'll stop there, but I think your view of why the hardware in the 3DS is a good idea for that product are skewed by Nintendo's goals for graphics, which are counter to almost everyone else. Your argument about power efficiency is good though (although a bit obvious) and that should always strive to increase (and almost always has for the last 10 years, generation after generation after generation). But it's never really going to happen at the expense of programmability. Even Nintendo haven't made the 3DS more restrictive for the programmer, but they've not caught up either.

 

[Richard]

My problem with Tim's view on software rendering is not the arguments he puts forth on why it's better. Yes, no more convoluted code to work on specialised hardware. Easier to maintain, and update, more flexible, yes, yes, yes.

So why didn't it work back in 1998? We had the two major engines coming out with software renderers, scratch that, games coming out with rendererers optimised for the CPU archs of the time instead of the hw accelerators exactly because most people didn't have a 3D Accel. Back then, assembly was also heavily used and the sequence of CPU instructions was carefully profiled. The data types, application model, precision, everything was made with the CPU in mind.

The first 3D accelerators didn't even provide that much of a performance boost, if they provided one at all (see GLQuake and Verite). Yes, you got 16-bit colour and filtering "for free" but still with the immature drivers of time, fragmented market, OS incompatibilies, why didn't software rendering succeed?

Because it was bloody slow that's why. Tim seems to believe things have changed in this regard, or will change soon. You don't need sound hardware nowadays except for professional editing or you don't have a digital output for your onboard. This doesn't mean dedicated sound hardware can't sound better, it just means the worst case of sound taking 10% of CPU time is fairly reasonable for what you get. Not so with graphics.

Also back in 1998 Moore's Law when applied to CPU speed, rather than transistors as it should, was very much alive. CPU speed basically doubled every 18 months so it stands to reason a software renderer would have legs to grow. Since 2004 we've hit a wall and are getting more cores rather than speed. Shouting RAYTRACING at that doesn't help Tim's case because this is the best ray traced implementation of an actual game I've seen and apart from the textbook rt effects the IQ improvement is marginal.

If you do go SR but bring up performance by introducing hacks/work arounds/smoke & mirrors how is that different from today's hw featureset requiring those? Yes, you have the option, but it's a false one if you can't hit 30fps without them.

So, software rendering is still as slow as it was before, relatively speaking, yet the market for it is much harder to penetrate now than it was, there's 12 years of game developer knowledge built on top of hw renderers so inertia will play a big role, and finally, if Tim continues to think SR is the way forward why hasn't he rolled up his sleeves and did something about it in the past 12 years?

perf/watt is a good argument that is demolished by the fact enough people still buy NVIDIA's top of the line cards for them to keep going OR the fact AMD (and now NV according to the recent slides) are/will be improving perf/watt significantly. Heck, if perf/watt is _that_ important that we're willing to overlook something as basic as being able to hit 30fps, why go with X86 at all? Let's all do software renderers for ARM. Performance, efficiency, flexibility. Pick any two.

Anyway, I mostly want him to release something. I want to see a new tech demo.

 

[brain_stew]

Thanks for the confirmation about the per pixel lighting in the Citadel demo Rys, could you say if it varies per platform? The impressions of the demo seem to be all over the place depending on what platform its played on. It sure impresses me but its the sort of thing I was expecting from the hardware for a while now, I just hope it serves as a watershed moment as the hardware seems to be so underutilised in most games. At least Apple seem to really be taking gaming seriously now, so hopefully the masses of games that seriously underutilise the hardware will become a thing of the past.


I assume the 106w figure applies to GF106 right? Well that's hardly the simple die shrunk GF100 which I was talking about and its power efficiency still strikes me as seriously mediocre at best. It can barely beat out the old 55nm RV770 in terms of performce/watt and performance/mm2. I dunno, I expect a new architecture built on a new process node to improve in these areas by a significant degree, not take baby steps like that. As such I don't really see what this new architecture could potentially bring to lower end and enbedded markets (like next generation consoles), relying purely on process shrinks to increase the performance per watt just seems very disapointing to me.

Genuine question but where do you feel the march towards programmability ends? Does fixed function hardware really have no future at all in markets where efficiency is so damn important. Is aiding developers to make breakthroughs in software efficincey by creating more programmable hardware the way we're going to make all of the real breakthroughs in terms of performance/watt in the medium term future?

I genuinely appreciate learning the perspective of someone like yourself and I guess my big problem is that I'm placing too much emphasis on Nvidia and Intel's approach to this problem. Both yourselves (you work for/with IMG, right?) and AMD are inching towards greater programmability with each new generation and yet still managing to make some seriously impressive strides in area/power efficiencey along the way so it should really be clear to me that these things aren't mutually exclusive. Perhaps the balooning TDP of Larabee and GF100 actually have less to do with the strides they've made towards programmability and more down to the fact that AMD/IMG are just better at designing efficient GPUs! If Nvidia's big focus is on targetting markets where large discrete cards are suitable then I guess it makes sense to put priorites elsewhwere.

Yeah, that 100w figure was completely arbitrary and obviously consoles aren't designed around arbritrary limits like that. Its certainly the sort of ballpark I'd expect/hope for anyway, anything significantly past that point is getting into seriously dodgy terriotry. I'd imagine the whole RROD fiasco may encourage movement in this direction. Targetting a lower TDP doesn't prevent something like that from happening again but it should at least make it less likely.

Oh and I apologise for my noobness and general naivety, don't be offended when I jump to incorrect conclusions. I only started learning about CPU and GPU architecture design and game technology about 18 months ago and only then in my spare time, so I'm still making some stupid and obvious mistakes. Be gentle!

 

[Andrew Lauritzen]

And its good, as much as I respect Sweeney and Carmack, I would hate if market would totally depend on fully software approach, this would only make them richer, but also make very diverse market with different engines.

Well to be fair you have to assume that it would transition similar to how FF T&L -> shaders happened. i.e. someone (Microsoft, an IHV or otherwise) will provide a free implementation of the legacy pipeline in software. No one is suddenly going to be left out in the cold who wants to just use a DX-style pipeline regardless of what happens beneath the pipeline.

It has been PROVEN with Larrabee that hardware tailored to a certain render pipeline is a least two times more efficient than generalistic hardware (Larrabee even had texture units).

That's just not true. It's fair to say that it remains to be proven that software rendering pipelines on Larrabee-like architectures can be in the same ballpark as conventional GPUs for conventional rendering pipelines (DX, OGL, etc). Certainly nothing has proven it impossible that I know of, but neither has it been proven possible.

Let's make a distinction here though... Tim is talking about future rendering pipelines that may not have the same characteristics (performance or programmability) as current pipelines. In fact what we expect these pipelines to look like is key to this entire discussion so I find comments that attempt to generally dismiss either approach (hardware or software or anywhere in-between) pretty naive. You have to define the workload before you can determine whether fixed-function hardware makes sense.

 

[Rys]

Thanks for the confirmation about the per pixel lighting in the Citadel demo Rys, could you say if it varies per platform?

The main lighting shaders are pretty much identical on the platforms I've analysed it on. There are other things that vary per-platform more than the shaders (mostly texture resolution and compression, and the scaler on the device), that can change the look of the thing quite a bit.

The impressions of the demo seem to be all over the place depending on what platform its played on.

I wouldn't say they're all over the place, but that's pretty much correct depending on where you're playing it. The changes seen are more a consideration for screen resolution and the amount of memory than the raw shader throughput though. And remember it's just a tech demo and not completely optimised yet.

It sure impresses me but its the sort of thing I was expecting from the hardware for a while now, I just hope it serves as a watershed moment as the hardware seems to be so underutilised in most games. At least Apple seem to really be taking gaming seriously now, so hopefully the masses of games that seriously underutilise the hardware will become a thing of the past.

You're right, there's a lot of GL ES1.1 content still prevalent in the iOS games market just now, but that's mostly market forces at work I think, combined with the types of games that are popular on little touch-screen devices. There's a lot of MBX-powered iPhones and iPod touches out there, and you need 1.1 to address those, and a lot of the games are quite simple too. That will change going forward, and Apple themselves are pushing developers towards using SGX and ES2.0 now that they're the majority for iOS. That'll have knock on effects with Android gaming too (most phones there are 2.0 capable now too).

I assume the 106w figure applies to GF106 right? Well that's hardly the simple die shrunk GF100 which I was talking about and its power efficiency still strikes me as seriously mediocre at best. It can barely beat out the old 55nm RV770 in terms of performce/watt and performance/mm2. I dunno, I expect a new architecture built on a new process node to improve in these areas by a significant degree, not take baby steps like that.

It's not a die shrunk GF100, but I thought you were talking about Fermi in general rather than one chip. Beating the old 55nm RV770 is pretty impressive actually, given RV770 was really awesome in that respect and a real leap in terms of perf/watt and perf/area for the industry. NV's problem there is that RV8xx made similar gains too, but they're not doing terribly in isolation IMHO. A lot of people are interested in GTX 460 because it's got good perf/watt, so they've done OK with GF104.

As such I don't really see what this new architecture could potentially bring to lower end and enbedded markets (like next generation consoles), relying purely on process shrinks to increase the performance per watt just seems very disappointing to me.

They wouldn't just rely on manufacturing if they were bringing it to a console design. There are definite area-led things they'd do to bring it into a certain size and power envelope without changing it so much you couldn't call it Fermi. The question is whether they could hit any vendor's targets for that (and whether they'd want to anyway, maybe the console market isn't something they can focus on right now).

Genuine question but where do you feel the march towards programmability ends? Does fixed function hardware really have no future at all in markets where efficiency is so damn important. Is aiding developers to make breakthroughs in software efficiency by creating more programmable hardware the way we're going to make all of the real breakthroughs in terms of performance/watt in the medium term future?

'Fixed function' will always have a place in real-time 3D graphics, full-stop. There will always be a part of any agreed on model of RT 3D that makes a non-generally programmable block on a piece of 3D silicon a good thing to do. It's just the relative area on the chip for those things will always go down. Breakthroughs in perf/watt are always going to be more the job of the materials scientists and the hardware architects, rather than programmers. That's orthogonal to wanting to give the programmer more freedom of expression in codes (but the latter can help the former).

I genuinely appreciate learning the perspective of someone like yourself and I guess my big problem is that I'm placing too much emphasis on Nvidia and Intel's approach to this problem.

Both IHVs have embedded graphics products that have completely different perf/watt and perf/area considerations to the big chips nobody seems to like any more. Both are not completely focused on massive 500mm^2 chips at 250W. They both have very low power embedded graphics products.

Both yourselves (you work for/with IMG, right?)

Yep, I work for Imagination in the PowerVR group :smile:

and AMD are inching towards greater programmability with each new generation and yet still managing to make some seriously impressive strides in area/power efficiency along the way so it should really be clear to me that these things aren't mutually exclusive. Perhaps the ballooning TDP of Larabee and GF100 actually have less to do with the strides they've made towards programmability and more down to the fact that AMD/IMG are just better at designing efficient GPUs! If Nvidia's big focus is on targeting markets where large discrete cards are suitable then I guess it makes sense to put priorities elsewhere.

LRB and GF100 are just one GPU each from a myriad that each IHV produce. In both cases they were designed for a power budget. You might not like the power budget, but both are really impressive GPUs IMHO. One is the fastest GPU in the world, the other is the most flexible piece of graphics silicon ever made. That's pretty great in both respects, right? I love both designs a lot, personally.

Yeah, that 100w figure was completely arbitrary and obviously consoles aren't designed around arbritrary limits like that. Its certainly the sort of ballpark I'd expect/hope for anyway, anything significantly past that point is getting into seriously dodgy territory. I'd imagine the whole RROD fiasco may encourage movement in this direction. Targeting a lower TDP doesn't prevent something like that from happening again but it should at least make it less likely.

Agreed, since GPU heat was the RRoD core problem.

Oh and I apologise for my noobness and general naivety, don't be offended when I jump to incorrect conclusions. I only started learning about CPU and GPU architecture design and game technology about 18 months ago and only then in my spare time, so I'm still making some stupid and obvious mistakes.

Not at all, it's good discussion and that's what these forums are for. As for being gentle....nah, where's the fun in that Christ, long quotey post is long.

 

[brain_stew]

I seriously appreciate the feedback Rys, thanks for spending your time giving your thoughts as i find its one of the best ways to learn about this stuff.

As for GF100 being an impressive design, I guess I am too hard on it, the TDP and heat figures just really made my eyes roll and I always found it hard to look past that. I have to admit it really is a silly approach to look at that stuff in isolation though, it has made some serious strides and breakthroughs that will likely prove important as we move on (the parallel rasterisation setup, the tessellation performance, the cache hierarchy and the general strides they've made on both the software and hardware side which makes GPGPU so much more accessible and useful for a wider range of markets) and with GF104 and GF106 they've proved that the architecture isn't a complete lame duck when it comes to perf/watt and perf/mm2.

One of Nvidia's products I've been most impressed with recently is actually that little iGPU they just developed for the Macbook (320m?). Considering the utterly insane battery life of that machine, that iGPU packs a serious punch and just about delivers on much of what Llano is hoping to do next year. Popular multiplatform engines like UE3 and Source run great on that chip @ ~medium settings in HD, its actually a genuinely viable gaming platform and its very rare that you could say that of machines with intergrated graphics in the past. I know all of the praise for the battery life in that machine doesn't rest solely at Nvidia's door but their chip is playing a big part, that machine wouldn't be nearly as attractive if it used a standard core i3/i5 chip with Intel integrated graphics. I guess its a shame that the march towards CPU/GPU integration has turfed them out of this area almost completely going forward, as the iGPU in the Macbook is currently very far ahead of the competition and is the benchmark that Llano, Ontario and Sandy Bridge all have to first surpass before they can be recognised as the game changers many are hyping them up to be.

Larrabee was always going to struggle to keep pace with the competition with the current reliance on DX9 throughout the industry, perhaps if the transition towards DX10 and DX11 was more swift then it would be a more competitive design. I'm sure its an architecture that performs very well in engines that make heavy use of compute shaders but they're only just starting to appear and be worked on now, consumers can't afford to buy a product based on how it might perform 2-4 years down the line by which point they'll probably want a new discrete card anyway.

If the industry really does make a wholesale transition to DX11 engines with heavy reliance on compute shaders within the next year or two then a product like Larrabee may just hit the mark. I doubt the fact that it would be slower in legacy DX9 titles would matter all that much so long as its "fast enough" in that area (and every <$200 GPU since the 8800GT has managed that) and there's a demonstrable advantage when its paired with engines built from the ground up to properly utilise DX11. I'm sure the fact that AMD have consistently been bringing their A game like never before in the last couple of generations didn't help matters either, I don't think many expected AMD to execute so damn well for two (now looking like 3/4!) generations off the belt right around the time Larrabee was set to hit the market. If even Nvidia with their wealth of experience and previous position of market dominance have seriously struggled to keep up then perhaps it was a bit much to expect a newcomer like Intel to keep up.

I guess at the end of the day we should be pleased that there's plenty of companies that are looking to tackle this problem in different ways. If everyone was just taking the exact same approach then we'd never know what was the "correct" approach. Designing a new GPU architecture is always going to involve a huge slice of guesswork as to where the market is going to be by the time the part is taped out. Its utterly impossible to accurately predict where the market will be just a couple of years down the line and developing something based on where things might go is always going to be a huge leap of faith. If you get that prediction all wrong through no fault of your own then in the end its going to matter little how well you execute on a particular design. If its not the right product for the current market then its going to fair poorly, its not a job I envy!

The way G71 and R580 have faired as time went on has been a real interesting curiosity for me. At launch those parts were pretty much neck and neck with most sites actually giving the nod to the G71 parts. Well, as time has went on, the gap has become huge, to the point where for the last few years most titles will run as much as 2x as fast (or better!) on ATI's part. In this case, I think its pretty safe to retrospectively give the design win to the red camp but int he end it counted for so very little. If the software market was just 12-18 months further along it would have been a massacre at launch but it wasn't so the reality was that AMD made very little headway in that generation despite clearly designing the better GPU.I guess the point I'm getting at is that even when you design a fantastic GPU, if the status of the software market isn't very tightly aligned with the vision you had at the early deign phase then you're SOL. A successful GPU design seems to rely on "luck" as much as anything else! I'll be looking back on RV870 and GF100 in a couple of years to see if a similar pattern emerges, perhaps Nvidia really will end up being proven to be right all along (the huge advantage they have in tessellation performance alone could be key) even if that didn't translate to success in the marketplace.

 

[Davros]

A lot of people are interested in GTX 460 because it's got good perf/watt, so they've done OK with GF104.

Really ? My findings suggest they are interested in perf/$ I dont think anyone I know knows how many watts it uses as long as they dont need a new psu wattage isnt a consideration

Yep, I work for Imagination in the PowerVR group :smile:

nevermind posting here go and bring out a fermi killer , you can do it

 

[Rys]

Really ? My findings suggest they are interested in perf/$ I dont think anyone I know knows how many watts it uses as long as they dont need a new psu wattage isnt a consideration

In the last few years you can't escape power measurement and discussion in graphics card reviews. As long as your friends are researching properly, I don't think they can really avoid knowing that power efficiency is important.

 

[Davros]

They know better power efficiency is desirable, but still don't think about it much unless power usage is huge and they need to worry if their psu can handle it