Nvidia GT300 core: Speculation

[Sound_Card]

Anybody here think they would honestly buy a $600 HD 4870?

 

[Razor1]

Anybody here think they would honestly buy a $600 HD 4870?

why would it be $600 it went up against the gtx 260 which was at $400-$450

 

[Ailuros]

I'm missing why exactly people want the ability to bounce between "scalar" and "vector". It isn't like they need separate registers like LRB. Running scalar means you're tossing out 16x or 32x or whatever of your ALU capacity, and if you want to do that then it's easy, just predicate out all but one lane of your vector. What is really needed beyond what is already there (in NVidia's GPUs) is branch/call by register.

Given that MfA was targetting autonomous branching per unit, I have the feeling that we might mean the same thing under different definitions.

Anybody here think they would honestly buy a $600 HD 4870?

Hypothetical answer to a hypothetical question: if I would have wanted a high end GPU and nothing else would had been available in that ballpark the answer is yes. It would presuppose the next best competing part to be about 30-40% slower and cost around $400.

In reality the RV770 is rather a performance than a high end GPU, but there were times were some of us paid such high if not higher amounts for a single GPU. The 8800GTX cost me in November 06' 650 Euros which by todays conversion should be ~$840.

 

[fellix]

Is RV770 MPMD, where each of the 10 clusters runs independently?

Each MP cluster has its own program counter, which determines the MIMD (MKMD in shader terminology) level of parallelism, sort of.

 

[DegustatoR]

To bad you are 1000% wrong... I mean come on, you didn't even bring any data to backup your claim.

That's because i can't disclose it.

So back then the $80-$250 market made up 88% of the revenue...
While the enthusiast was only 4%...
Hmmm... who was wrong?

What you really need to know is the profit margins of each segment. Enthusiast while being only 4% in revenue (this number seems to be too low btw because workstation being only 5% of revenue is clearly wrong for NV today) generates massive amount of profits because of much higher margins.
So who is right - the one who dumps shitty AFR cards on that market or the one who makes special products for it also while still using shitty AFR cards as a temporal solutions? Don't answer, that's a rethorical question.

 

[willardjuice]

That's because i can't disclose it.

I lol'd.

 

[Ailuros]

Each MP cluster has its own program counter, which determines the MIMD (MKMD in shader terminology) level of parallelism, sort of.

Honest question: so I assume that each cluster can branch completely independently?

 

[fellix]

Yup!

 

[Lukfi]

So who is right - the one who dumps shitty AFR cards on that market or the one who makes special products for it also while still using shitty AFR cards as a temporal solutions? Don't answer, that's a rethorical question.

How is that a rhetorical question when you have already judged the AFR cards as shitty?

 

[DegustatoR]

How is that a rhetorical question when you have already judged the AFR cards as shitty?

You have another opinion? (As in you, not AMD who may have it's reasons for bailing out of high end GPU production.)

 

[Ailuros]

Don't throw any stuff at me but to the application it doesn't matter if use a dual-chip GPU, two GPUs or a dual-GPU board, it's AFR through SLi/CF anyway. Whether crappy or not it was NVIDIA that first re-introduced the whole idea after the NV40 introduction (if memory serves well).

 

[dkanter]

What you're basically saying is this: "why earn $300 when you can earn $100". Do you really want to say something like this? Because that's clearly b.s.


Bolded part isn't true at all. I know this for sure. So all the other parts of this phrase are wrong also.

I think you misunderstood what I wrote. See the revenue breakdown someone else posted.

High-end graphics cards do jack shit for profits. You make real money on high volume shipments of cards in the sweet spot, not ridiculous $500 cards.

The pro market is insanely profitable, but that's quite different from the market for graphics cards. The pro market can be broken down into two sub-sets:
1. Traditional workstation and professional graphics
2. GPGPU

Market #1 is the GPU-world equivalent of servers. Low volume, but insane margins. And you can make a bucket load of money there.

Market #2 contributes approximately 0 to anyone's bottom line, because the volumes aren't low...they are SUPER HYPER MEGA ULTRA LOW edition (TM). The margins are lower too, since the NRE (software mostly) is amortized over even lower volumes.

Market #2 absolutely requires the highest performance monolithic die because the programming model for multiple general purpose GPUs is brain dead and retarded (to put it politely).

It has yet to be demonstrated to me why Market #1 (which is the only one that's financially relevant) requires the highest performance monolithic die. In fact, I suspect that a large swath of the market would be happy with the highest performance card, whether it's one GPU die or two...or 16 (maybe you've heard of Infinite Reality). One of the nice things about most pro apps is that since they were originally written for SGI systems, I strongly suspect most of them scale quite nicely across multiple GPUs...

Doing more with less resources is uninteresting? Doing things that are impossible on AFR system is uninteresting? That's certainly an interesting point of view. Maybe we should go back to Voodoo days since all that flexibility and programmabililty is uninteresting?

Nothing is tricky in the inefficiencies of AFR. The tricky part is when you try to avoid them. And for that you often loose that flexibility.

I think you missed my point. Using two smaller GPUs can be more efficient than a single big GPU in many cases (if nothing else you get to reduce the pin, power and heat density on the board). Also those two smaller GPUs yield better, and have lower absolute NRE and even lower per-die NRE (thanks volume economics - you are awesome!).

The problem, as we both agree is that AFR is a ugly application visible hack that sucks. I suspect that most graphically intense games have AFR profiles...if you have proof to the contrary, I'd love to see it.

While you're at it, could you give some examples of things that are 'impossible on AFR systems'?

However these servers don't use middle class CPUs to achieve that and they certainly aren't selling in mainstream market. Why? If anything we're seeing the opposit process with CPUs: more cores are getting integrated into one big chip. Have you ever thought about this?

Get your facts straight.

Plenty of mid-range chips are sold into 2S servers. I'm sure you've heard of this product called Barcelona...And Intel sells quite a few Harpertown systems these days too.

Dual chips will always have some logic that isn't needed in dual chip configuration and that means that their efficiency will always be less than the efficiency of single chip. Single chip will always have some algorythms where it will beat dual chips because of the limitations of AFR mGPU scheme.

That's the weakest argument ever. The advantage of multi-chip scalability vastly outweighs the overhead.

So you've saved some bucks on the die and you've wasted nearly 2x bucks on the price of the card. Are you in the green after that?

Dude you don't get it. GPUs are semiconductors, semiconductors are about volume economics. Extremely large die sizes are antithetical to volume economics. Look at successful CPU strategies, you only sell reticle limited chips when your margins are >70% and the volume supports it. The pro market has the margins and the volumes, the high-end gaming market has neither.

What if you've missed the sweet spot and even one GPU competitor's card is faster than your mGPU card? If you have some GPU faster than you're using in your mGPU card you may be able to use it in the new mGPU card (GTX295 is an example although not the best one), if not -- you're truly fucked.

Not having the highest performance card is fine. What matters is that your single die does a great job of addressing the market for $100-300 cards. If someone else has to match your pricing with a much bigger die, they lose period.

I think what you are saying is that if you make your single die too weak, then you will be in trouble, which I agree with. What I'm talking about is instead of having a single GPU hit 100% highest performance, you have a single GPU that hits 70% of maximum performance, and then you use two GPUs to hit 140%. I'm not saying go design a half-assed 30% GPU, since that would make no sense.

AMD is leaving it's high end dangerously open for a possibility like that.

It's possible, but not as long as NV is fielding reticle limited chips.

It's a question of having full line up. AMDs line up is missing high end at the moment. Were NV will use two chips in Quadro/Tesla market AMD might need to use four with appalling efficiency and costs. That's a possibility that you should think about when you're speaking of multi-CPU servers.

AMD's high-end is fine, they have the x2 line, which does quite well, and the RV770 is earning money in the market that matters.

You don't get it. A board with 4xRV770s is probably cheaper for AMD than a board with 2xGT200s for NV.

And can someone explain me why ATI earns zero on all these great small GPUs and NV earns nearly the same on that big ugly GT200 now selling in cards for less than $200? I've always had a problem with that pricing argument since it's kinda always was "assumed" that RV770 is much better for AMD than GT200/b for NV from the pricing point of view but in reality i'm not seeing any results of this "greatness" in AMDs balance sheets -- ATI earned less in the 1Q09 than it did in the 1Q08 when all they had was RV670 against G92 and G80.

I'm not sure you noticed, but the economy in 1Q09 wasn't the best in the world. Why don't we compare AMD's gaming GPU results for the last three months against NV's gaming GPU results for the last three months when they are available.

Nobody is have to do anything. NVIDIA is doing what they believe will earn them money. AMD is doing the same. Whose way is the best -- we don't know. But what everyone should consider is that NV's way is essentially nothing more and nothing less than AMD's way plus big GPU dies for high-end/workstation/server markets. AMD has simply left that market segment.


It's funny that you say this right after you've said why single big GPUs are neccessary after all.

You clearly don't understand ATI's strategy or NV's.

NV's way is exactly what they've been doing for the last 5 years or so. Build a big GPU in a cutting edge process and sell it to the high-end where there's little volume. After a process shrink or two, sell a medium size version to markets that have volume.

ATI's way is build a medium sized GPU and sell it to markets that have volume, and then sell 2 of them to the high-end markets.

It really cannot get any more blinding obvious than that.

David

 

[LordEC911]

That's because i can't disclose it.


What you really need to know is the profit margins of each segment. Enthusiast while being only 4% in revenue (this number seems to be too low btw because workstation being only 5% of revenue is clearly wrong for NV today) generates massive amount of profits because of much higher margins.
So who is right - the one who dumps shitty AFR cards on that market or the one who makes special products for it also while still using shitty AFR cards as a temporal solutions? Don't answer, that's a rethorical question.

Wow... I don't even know what to say to that...
That is sad though. Keep your BS to yourself.

 

[Scali]

The 8800GTX cost me in November 06' 650 Euros which by todays conversion should be ~$840.

Yea, sometimes these can be great investments though.
Even today an 8800GTX is still quite a decent performer, so the card has a very long lifespan in that sense.
Last time there was a card like that, it was the Radeon 9700.

 

[Panajev2001a]

David, not to contest your points (well made ), but I'd say that we should not undervalue the importance of having the performance crown...

Sure, bankrupting your company to achieve it would not be a sound strategy, but having such crown can do a lot of good to your image and help your marketing teams sell your profitable mid-range and low-range segments than tons of TV and magazine ads could ever do IMHO (gives the customers confidence in all products sold by that manufacturer).

 

[Jawed]

Each MP cluster has its own program counter, which determines the MIMD (MKMD in shader terminology) level of parallelism, sort of.

I'm pretty sure each cluster in ATI has a program counter per thread. I wonder what's the marketing name?

Jawed

 

[LordEC911]

David, not to contest your points (well made ), but I'd say that we should not undervalue the importance of having the performance crown...

Sure, bankrupting your company to achieve it would not be a sound strategy, but having such crown can do a lot of good to your image and help your marketing teams sell your profitable mid-range and low-range segments than tons of TV and magazine ads could ever do IMHO (gives the customers confidence in all products sold by that manufacturer).

Not necessarily...
Low-end, yes.
Anywhere else, not noticeably IMO.

OEM's can also even change that "Halo effect" in the lowend. Give them a $2 savings vs the competition and they could care less about overall performance, as long as it meets their minimum requirements.

Edit- Wow Deg, even when you are proven wrong again and again, you play the "source card," play dumb and ignore the obvious.
Welcome to fantasy land tinged green... (not because of cash)

 

[DegustatoR]

Don't throw any stuff at me but to the application it doesn't matter if use a dual-chip GPU, two GPUs or a dual-GPU board, it's AFR through SLi/CF anyway. Whether crappy or not it was NVIDIA that first re-introduced the whole idea after the NV40 introduction (if memory serves well).

4-way AFR is way crappier than 2-way AFR and if we assume that AMDs 2-way AFR is going against NVs 1 GPU in high end then we are in position where AMDs 4-way AFR is competeing against NVs 2-way AFR.
AMD doesn't have GPU for the high-end at the moment. It's that simple.
NVs problem right now is that it doesn't have a competitive GPU for middle end. We're still in G92 vs RV770 situation, almost a year later. That's the real problem for NV, not GT200 being slow or big or some kind of abstract strategy.

dkanter, i won't argue with you in such tone. Sorry, but i don't see any point in discussing anything with you atm. It's you who need to get your facts straight, not me.

 

[Lukfi]

You have another opinion? (As in you, not AMD who may have it's reasons for bailing out of high end GPU production.)

I agree with you, but your rhetorical question seemed to be pointed to another direction, AMD and nVidia. And you obviously know that AMD had its reasons for this.

 

[LordEC911]

AMD doesn't have GPU for the high-end at the moment. It's that simple.
NVs problem right now is that it doesn't have a competitive GPU for middle end. We're still in G92 vs RV770 situation, almost a year later. That's the real problem for NV, not GT200 being slow or big or some kind of abstract strategy.

Contradict much?
Nvidia is using G92, a 1.5year old chip, because they simply cannot earn a profit selling a G200 chip at those prices. G200 is too big and costly to sell for less than ~$150, i.e. G92 prices.