Nvidia GT300 core: Speculation

[rpg.314]

I don't know exactly yet, but isn't the "multithreaded rendering" stuff in it even helping multi GPU cards in distributing the workload?

Nahh.

 

[Scali]

Too true. Well unless they can make multi-gpu gaming actually work just as well, instead of super patchy performance that works great in some games and crappy in others.

It will be interesting to see Intel's solution for this.
Intel has already hinted at LRB scaling very well to multiple GPUs.

 

[Scali]

I don't know exactly yet, but isn't the "multithreaded rendering" stuff in it even helping multi GPU cards in distributing the workload?

As far as I know, D3D11 only allows you to create batches of drawing operations on multiple threads in parallel, but in the end they'll all just be sent to the GPU via a single thread, in a sequential fashion.

 

[Jawed]

Nice evil wish, Mr. Jawed. I doubt it's gonna happen. After all amd sits on the committee that makes the spec. :-/

I think the TS, UAV and CS stuff all basically fragment rendering into such tiny pieces and irregular, transitory blobs of memory that AFR just becomes a bottleneck as all the sane algorithms try to make multi-frame use of all the goodies brought by this functionality.

Also, it seems to me that Microsoft has so far successfully refused to allow any AFR nonsense to pollute D3D. Could be accidental, of course.

Jawed

 

[CarstenS]

The idle usage doesn't tell you what the load usage is.

Jawed

I didn't say so, now did I?
But since there's no hard data available for RAM alone (at least not that i know of), I can only judge from Idle power consumption and this is definitely not implemented for a decent show-off of GDDR5 power saving features, is it? I mean, after all the high idle power of HD 4870/90 is generally blamed on the memory controller inside the chip.

Now, if it's doing what it is doing when doing nothing in the same manner under load, only faster, then I'd rather not think of it as very power efficient in it's current implementation - iow why should it turn into an energy efficient system under load, when it isn't when idlin?

Hence my wish, to first see a decent GDDR5-implementation.


edit:
To make a long story short: No, I have no hard data regarding the power consumption of GDDR5 alone. But since it is apparently the way to go, it has to be taken into the equation. And its supply - be it high or low - has to come from within the 300 watt power budget.

 

[Dave Baumann]

Hence my wish, to first see a decent GDDR5-implementation.

Are GDDR5 enabled devices not at similar power loads (or less, in the case of HD 4890) as their performance competetive contemporaries?

To make a long story short: No, I have no hard data regarding the power consumption of GDDR5 alone.

Although it won't necessarily give the full picture, the first place to start is look at the voltages of the devices. GDDR5 typically runs lower voltages than GDDR3.

 

[rpg.314]

I think the TS, UAV and CS stuff all basically fragment rendering into such tiny pieces and irregular, transitory blobs of memory that AFR just becomes a bottleneck as all the sane algorithms try to make multi-frame use of all the goodies brought by this functionality.

Sorry. Didn't get it.

 

[DegustatoR]

It is my opinion that from now on, multi-GPU solutions will always win over monolithic (assuming similar manufacturing costs of both competing products). The current situation has a lot to do with GT200 being not so good. But generally, performance scales linearly to transistor count, while the function of yields is concave. The only problem of today's multi-GPU solutions is software. Monolithic GPU means performance is guaranteed in any game, multi-GPU means living in uncertainty.

Don't forget that while you may save some bucks on the smaller GPUs in mGPU card you're still burning much more on double memory size and more complex PCB/power/cooling solution.
Generally "one big GPU" cards are simply more effective than any mGPU solution that we saw up until today. And that's true not only in cost of producing cards, but in performance and features too.
RV770x2 vs GT200 is like an opposite of RV670x2 vs G80 -- the second scenario was a bit unfair to mGPU cards because RV670 was kinda bad. And now we have a scenario which is in my opinion unfair to "big single GPU" cards because now GT200 is kinda bad.

But maybe this problem can be solved by adding special multi-GPU logic that would make the whole solution work more like a single-GPU system. Yeah that sounds like Hydra... perhaps future solutions will use this "dispatch processor" model? I don't know, but I feel these are areas worth to explore.

I'm quite sure that nothing will bring mGPU solutions to the level of single GPU solutions in terms of flexibility and efficiency ever.
An interesting area for mGPU cards lies a bit higher than where AMD is putting it's RV670/770 GPUs -- let's say that you have a GPU with performance between middle and high-end class. There may be a window where you can make an mGPU card with two such GPUs which cannot be challenged with one single big GPU simply because you won't be able to make such GPU (due to technical limitations). Anything higher will be too complex to use in mGPU cards, anything lower will be beaten by a single big GPU boards.
This window is where AMD is with RV770x2 essentially, but only WRT performance levels, not die complexity.

Hmm, that's an interesting point. For AMD, this strategy proved useful. On the other side, people kind of expect a performance bomb, not a middle class.
Maybe they'll launch a middle-class GPU alongside with a SLI-on-a-stick version that would claim the performance crown and create positive publicity to bolster single-GPU card sales.

Well, people tend to expect too much. But why isn't GT30x middle with GT200+ performance and DX10.1/11 support at $250 price point won't be a performance bomb anyway? After all that's exactly what's expected from AMD's RV870.

Well for that you'd need to keep power in check. It's kinda hard to see nv doing that while making an uber gpu in the first place. Why would they put constraints on themselves. gtx 285/295 is kind of a freak thing since it is not always that you can launch a chip and it's shrink 6 months later.

What's the problem with doing everything that AMD does PLUS doing a single big GPU AFTER you've done what AMD did? I don't see how's any kind of power usage may be a problem here.

Shouldn't GT212 have arrived by now?

End of 2nd Q, start of 3rd Q was a target for GT212 as far as I remember.

How big is GT200b if shrunk to 40nm? Does that count as a mid-range GPU then? Surely, by 2009Q4 GTX285 performance is what we'll be calling "mid-range", so that would give us a good idea of a "mid-range" GT3xx.

Here's a funny rumour: I've heard that NV already "shrunk" GT200 to 40G (512-bit GDDR3 remains, yeah). Sounds fishy, I know.
But maybe they've decided to scrap all new GT212 in favour of this "quick'n'dirty GT200@40nm" GPU PLUS bring some GT30x middle closer to GT300 launch? I guess we'll know something more solid about what's going on in a couple of months.

 

[rpg.314]

Well for that you'd need to keep power in check. It's kinda hard to see nv doing that while making an uber gpu in the first place. Why would they put constraints on themselves. gtx 285/295 is kind of a freak thing since it is not always that you can launch a chip and it's shrink 6 months later.

What's the problem with doing everything that AMD does PLUS doing a single big GPU AFTER you've done what AMD did? I don't see how's any kind of power usage may be a problem here.

My point was that 285 is a shrink of a giant gpu, so it is unusual.

Scaling up a midrange chip? That would leave the high end (with new features) until you can scale it up.

 

[Lukfi]

Don't forget that while you may save some bucks on the smaller GPUs in mGPU card you're still burning much more on double memory size and more complex PCB/power/cooling solution.

PCB, power and cooling depend on the total power draw. Who says monolithic GPUs can't be power-hungry? Seems more like a general issue than multi-GPU-specific.
Memory, that's a different story. Yeah sure the current way is ineffective, but that's exactly what I'm talking about, we need something better than AFR, something that can take advantage of the extra memory. This is still up there for grabs.

Generally "one big GPU" cards are simply more effective than any mGPU solution that we saw up until today. And that's true not only in cost of producing cards, but in performance and features too.

We don't know what the cards cost to make. But, suppose HD 4770 will retail for $100 and HD 4890 will cost $200. With a $200 budget, what will you buy? The HD 4890, sure, but if CrossFire didn't suck you'd go for a couple of HD 4770's. Anyway most likely you've got a motherboard with two PCIe x16 slots and CF support even if you don't use it.

I'm quite sure that nothing will bring mGPU solutions to the level of single GPU solutions in terms of flexibility and efficiency ever.
An interesting area for mGPU cards lies a bit higher than where AMD is putting it's RV670/770 GPUs -- let's say that you have a GPU with performance between middle and high-end class. There may be a window where you can make an mGPU card with two such GPUs which cannot be challenged with one single big GPU simply because you won't be able to make such GPU (due to technical limitations).

That's one way to look at it. But the holy grail of multi-GPU tech is to make a GPU that will cover several segments alone. And if you can somehow get it to scale in all cases, even though it won't be perfect, that's exactly what you need. Of course, there's a big if about whether that can be done at all... but hey, strange stuff happens.

Well, people tend to expect too much. But why isn't GT30x middle with GT200+ performance and DX10.1/11 support at $250 price point won't be a performance bomb anyway? After all that's exactly what's expected from AMD's RV870.

While I think R600 wasn't that bad in itself, it certainly was a big flop. ATI fell to the bottom, people didn't expect them to launch anything nearly as powerful as the RV770. But nVidia is in a different situation now: they have the most powerful monolithic GPU, arguably the most powerful graphics card (GTX 295). That is their reputation. Although, if they launch something, call it the GT 350, pitch a fair price and say "this is the mainstream, there's yet high-end to come", people will buy it. But doing that they either destroy GTX 200 sales, or the new card will get thumbs down for being expensive. Maybe that's why you launch high-end first. You get time to sell old stock and prepare ground for your new mainstream models. Some say that AMD could do a bigger RV770 based chip, but they don't want the RV870 (with worse performance/die size ratio because of DX11) to look weak in comparison. When ATI was launching the RV770, there wasn't a whole lot to cannibalize as there was nothing faster than RV670.

 

[Jawed]

End of 2nd Q, start of 3rd Q was a target for GT212 as far as I remember.

One quarter before GT300? That would be like launching 7900GTX in July 06 ~4 months ahead of 8800GTX.

Here's a funny rumour: I've heard that NV already "shrunk" GT200 to 40G (512-bit GDDR3 remains, yeah). Sounds fishy, I know.
But maybe they've decided to scrap all new GT212 in favour of this "quick'n'dirty GT200@40nm" GPU PLUS bring some GT30x middle closer to GT300 launch? I guess we'll know something more solid about what's going on in a couple of months.

Hmm, how much smaller can GT200 get and still have a 512-bit bus?

How big is G94b's die? Double that for 512-bit?

Or maybe we should just say that R600 is as small as a 512-bit bus can get - though with some fudge factors relating to power.

But there's still going to be some GT2xx in 40nm between now and GT300, isn't there?

Jawed

 

[Jawed]

Sorry. Didn't get it.

Say you build a "global illumination" algorithm, you might use low refresh rates on the computations, or even stagger the computations over successive frames operating on different areas/LODs in a round robin fashion.

Or you have a particle system built upon an effects-physics simulation.

Or reprojection caching of pixels, using final shaded pixel results from prior frames.

Any kind of technique that makes stuff persist over multiple frames is AFR-unfriendly.

Jawed

 

[DegustatoR]

PCB, power and cooling depend on the total power draw. Who says monolithic GPUs can't be power-hungry? Seems more like a general issue than multi-GPU-specific.

Power effiency of sGPU cards should always be higher than that of the mGPU card. You either should have more performance at the same power or less power at the same performance. In reality it's not that simple of course because of the differences in architectures and processes.

Memory, that's a different story. Yeah sure the current way is ineffective, but that's exactly what I'm talking about, we need something better than AFR, something that can take advantage of the extra memory. This is still up there for grabs.

Why, that's single big GPU =)
The problem of any mGPU config with some kind of shared memory lies in cost of such system compared to single GPU card. You'll eventually burn everything you gained from smaller GPUs on all this complex technology which allow them to work together in the same way a single GPU does.
So I'm thinking that while some kind of mGPU evolution is possible it all comes down to "cheap AFR card vs costly single GPU card" scenario. If an AFR card can be better than single big GPU card from the ttm/price/performance point of view, a more complex mGPU card with shared memory and who know what else will probably end up being more expensive than any single big GPU solution.
So I won't hold my breath for anything better than AFR in the next generation of mGPU cards again.

While I think R600 wasn't that bad in itself, it certainly was a big flop. ATI fell to the bottom, people didn't expect them to launch anything nearly as powerful as the RV770. But nVidia is in a different situation now: they have the most powerful monolithic GPU, arguably the most powerful graphics card (GTX 295). That is their reputation. Although, if they launch something, call it the GT 350, pitch a fair price and say "this is the mainstream, there's yet high-end to come", people will buy it. But doing that they either destroy GTX 200 sales, or the new card will get thumbs down for being expensive.

So they'll change everything based on GT200b to this GT350 while lowering prices and adding DX11 support. I don't see how's that bad for them considering that they'll promise a high-end solution soon anyway.
That's exactly what AMD's done two times already -- i don't understand why you think that what's good for AMD will be bad for NV.

Maybe that's why you launch high-end first. You get time to sell old stock and prepare ground for your new mainstream models. Some say that AMD could do a bigger RV770 based chip, but they don't want the RV870 (with worse performance/die size ratio because of DX11) to look weak in comparison. When ATI was launching the RV770, there wasn't a whole lot to cannibalize as there was nothing faster than RV670.

GT200b isn't exactly a stellar perf/mm GPU, it's on 55nm while all GT30x will use 40G and have a seriously updated architecture. I'd say that NV have much more room for improvements than AMD had with RV670->RV770 transition. After all GT212 was expected to have +30% of GT200b shader performance and GT30x middle class GPU should be near that performance position.

One quarter before GT300? That would be like launching 7900GTX in July 06 ~4 months ahead of 8800GTX.

Maybe that's why we won't have a pleasure of seing GT212? It was too late to make any kind of sence.

But there's still going to be some GT2xx in 40nm between now and GT300, isn't there?

Low end OEM parts (GT218, GT216) and some kind of RV740 competitor (GT215). It looks like everything higher than that will be GT30x-based.

 

[Lukfi]

The problem of any mGPU config with some kind of shared memory lies in cost of such system compared to single GPU card. You'll eventually burn everything you gained from smaller GPUs on all this complex technology which allow them to work together in the same way a single GPU does.

That's also a possibility. Well, it's cool that we know that it is possible to manufacture and sell a 576 mm2 GPU, but how much longer can this go on?

So they'll change everything based on GT200b to this GT350 while lowering prices and adding DX11 support. I don't see how's that bad for them considering that they'll promise a high-end solution soon anyway.
That's exactly what AMD's done two times already -- i don't understand why you think that what's good for AMD will be bad for NV.

When did AMD do this? People didn't want R600 anyway, so RV670 didn't cause much harm. RV770 was faster than RV670 so they didn't interfere at all (except single RV770 vs. RV670 X2 which was an exotic solution anyway). RV730 offered similar performance and features as the RV670, for most people this was a "potato, po-tah-to" situation.
If they would've been different, the solution would be simply adjusting prices. The problem with a GT 350 is, how do you adjust the price for a DX11 part versus the DX10 ones? I certainly don't expect the perf/transistors ratio to increase, as long as I can remember this hasn't happened with any Direct3D version. How do you sell your old DX10 stock when your marketing is all about "DX10 is crap, buy our new DX11 products"? Of course, the classic way around this is to lower prices of the current-gen cards while delaying the new ones - that would however depend on whether there will or won't be an RV870 to pressure nVidia to launch a DX11 part (similar to Radeon X800 GT/GTO/GTO2 vs. X1600 situation).

 

[Domell]

You are saying everything higher than mainstream GPUs will be most likely based on GT3xx architecture, but the question is WHEN? According to latest rumours GT300 is supposed to be released at least in October or maybe even in January/February 2010. Moreover NVIDIA almost always releases highend GPUs at first so Middle-end GT3xx are going to be seen on the market even later.

Another thing is latest rumours about die size of GT300. Bigger die in 40nm than 2xGT200B die in 55nm??????!! What the hell is going on here?? It`s sick. If it`s true (but i hope is not) then what die size will have middle end parts?? 500-600 mm^2?

 

[Lukfi]

You are saying everything higher than mainstream GPUs will be most likely based on GT3xx architecture, but the question is WHEN? According to latest rumours GT300 is supposed to be released at least in October or maybe even in January/February 2010. Moreover NVIDIA almost always releases highend GPUs at first so Middle-end GT3xx are going to be seen on the market even later.

We expect the first GT300 parts in October or November. We don't know, however, whether those will be high-end or mainstream parts. My bet stays on the high-end.

Another thing is latest rumours about die size of GT300. Bigger die in 40nm than 2xGT200B die in 55nm??????!! What the hell is going on here?? It`s sick. If it`s true (but i hope is not) then what die size will have middle end parts?? 500-600 mm^2?

Where did you hear that?

 

[CarstenS]

Are GDDR5 enabled devices not at similar power loads (or less, in the case of HD 4890) as their performance competetive contemporaries?

When comparing against HD 4850 - no. It's about 40-50 watts higher under load and I do not want to attribute all that to the GPU alone, albeit it runs at a higher voltage.

Although it won't necessarily give the full picture, the first place to start is look at the voltages of the devices. GDDR5 typically runs lower voltages than GDDR3.

As was the case with GDDR2 (FX5800 Ultra anyone?) and GDDR4, which failed to impress also. That's basically what I meant when talking about the superior power characteristics on paper.

You/AMD have shown that PP 2.0 can work really well in the case of HD4670. All I'm saying is, that I'd love to the an equivalently impressive implementation on HD 4770.

 

[CarstenS]

Another thing is latest rumours about die size of GT300. Bigger die in 40nm than 2xGT200B die in 55nm??????!! What the hell is going on here?? It`s sick. If it`s true (but i hope is not) then what die size will have middle end parts?? 500-600 mm^2?

I think, Charlie D. did write that a bit unfortunate. IMO he referrred to a sinlge GT200b chip.

 

[Ailuros]

You are saying everything higher than mainstream GPUs will be most likely based on GT3xx architecture, but the question is WHEN? According to latest rumours GT300 is supposed to be released at least in October or maybe even in January/February 2010. Moreover NVIDIA almost always releases highend GPUs at first so Middle-end GT3xx are going to be seen on the market even later.

Another thing is latest rumours about die size of GT300. Bigger die in 40nm than 2xGT200B die in 55nm??????!! What the hell is going on here?? It`s sick. If it`s true (but i hope is not) then what die size will have middle end parts?? 500-600 mm^2?

You're referring to Theo's ramblings aren't you? It's damn hard to find at least one assumption in that newsblurb that makes sense, but if you'd calculate how much die area a single GT2x0@40nm would capture, I wouldn't be suprised if the result is roughly half the die size of GT3x0.

If a GT3x0 would be as big as you are thinking, the result would exceed the 5B transistor mark which I don't think is the case. Something rather =/> 2x times the transistor count of GT200 sounds far more reasonable for this generation.

 

[Jawed]

If a "GT350" is going to be any use in X2 configuration then it needs double-precision.

Jawed