prepare all you want, doesn't change the fact that you lost a full week and more -- considering the lack of efficiency during holiday seasons.
What should they have done - held off until afterwards and lost even more time?
What should they have done - held off until afterwards and lost even more time?
At $300, it doesn't even matter if it's 15% slower than a 7970. This is starting to smell like RV770 in reverse indeed.xDxD said:
If his report of Kepler being competitive didn't dampen some enthusiasm for 7970, this definitely will...
NathansFortune
Regular
If Kepler really is properly competitive with 7970, even within 10% striking distance ATi are fecked. Seriously done. AMD are losing money on processors and Nvidia will be pushing their margins negative on GPUs, I expect they won't be around for long if they can't turn at least one of their divisions around. Nvidia are fighting well in the ARM theatre with more designs coming out with Tegra 3, an area where AMD aren't even involved.
We have to remember that AMD and Nvidia almost always price their products according to their performance against competition. If GK104 will be indeed priced at $300 then you can expect ~GTX 570 performance with (much?)lower power draw and more features (so this will be the GTX 660).
What AMD did with RV770 was forced by their previous gen fiasco and financial trouble caused by ATI acquisition. They were desperately trying to regain mind and market share. Even though i loved what they did i highly doubt something like this will happen again.
Of course i also suspect that $300 GK104 will force AMD to drop prices on HD79xx but not by much.
Seriously I don't get all the doom and gloom stuff here.
GF114 wasn't that much slower than Cayman neither (with a smaller die and using slower memory).
So if the chip is somewhat smaller and somewhat slower what's the problem? In terms of power consumption (which dictates cooling, voltage regulation) everything seems to point to it being similar to a HD7950. I bet if nvidia could make profits with such a card at 300$ AMD could as well with a similar priced HD7950. Maybe AMD would have a slightly more complex pcb (due to using a 384bit memory interface instead of 256bit) but I don't think there would be a dramatic difference (at least for the 1.5GB version) in costs.
As for the die size, well I've not yet seen any credible estimate for GK104. Only if it's really small (compared to the performance it offers) it could be problematic for AMD. And noone seems to know what the heck GK110 actually is...
nvidia has no reason to slash prices just because they potentially could while still making a (smaller) profit. Their goal is to make the most money not to hurt amd, and prices will be set accordingly.
GF114 wasn't that much slower than Cayman neither (with a smaller die and using slower memory).
So if the chip is somewhat smaller and somewhat slower what's the problem? In terms of power consumption (which dictates cooling, voltage regulation) everything seems to point to it being similar to a HD7950. I bet if nvidia could make profits with such a card at 300$ AMD could as well with a similar priced HD7950. Maybe AMD would have a slightly more complex pcb (due to using a 384bit memory interface instead of 256bit) but I don't think there would be a dramatic difference (at least for the 1.5GB version) in costs.
As for the die size, well I've not yet seen any credible estimate for GK104. Only if it's really small (compared to the performance it offers) it could be problematic for AMD. And noone seems to know what the heck GK110 actually is...
nvidia has no reason to slash prices just because they potentially could while still making a (smaller) profit. Their goal is to make the most money not to hurt amd, and prices will be set accordingly.
Haha, yeah right. You do realize even if this 299 thing is true, and even if performance is close to 7970, AMD can just drop 7970 to 299 as well?
I'm sure the GPU's are similar in die size, 7970 is a pretty small chip. It's only priced high due to market slotting not BOM costs.
But as it stands now, AMD is continuing actually getting cards out with the 7950 dropping next week and Nvidia has nothing. Until that changes it's just speculation and AMD has the upper hand.
My guess/hope is GK104 will be really nice. It's a no brainer, a GTX 580 with some amount more shaders and perhaps a higher clock should be a doddle on 28nm, and such a part would obviously give 7970/50 a run for it's money. Forcing AMD to slash prices as well.
It's the 299 part that sounds great, fantastic, I would LOVE it if that happens, but possibly too good to be true. I would guess more like 399, which would still be nice but not quite so earth shattering.
But keep in mind 2GB/256 bit bus on GK104 will place some performance limits on it. Especially at higher resolutions. So it might be equal to 7970 at 1080P and less, then start falling back. That's still great though. We'll see.
As for your AMD is doomed screed, please. AMD isn't going anywhere, Bulldozer could literally not have been worse and they still aren't going anywhere, so imagine if they ever actually get a decent CPU out. There will likely always be a place for two vendors, one of them AMD being the low cost one. AMD does really well with all the APU type chips now that you arent even thinking about, just the other day I saw about sub $500 AMD based ultrabooks planned, while the Intel ones are 700+. Bet those are going to be pretty popular. And I was looking at really low end laptops a few months ago in mass market retailers (like, Wal Mart), and noticed all most all of them feature AMD chips.
Silent_Buddha
Legend
If performance is close to 7950/7970, then GK104 will be priced at a similar price to those cards.
Pricing of 299 USD suggests it will be a fair bit slower than 7950. Nvidia aren't ones to seriously discount their cards unless they are forced to as they were with GTX 260/280.
They aren't hurting for marketshare as AMD was with Rv770. They continue to generate more sales than AMD when price/performance is similar. So I see no reason why Nvidia would price something that is competitive with 79xx at a significant discount and then run into supply issues leaving huge sums of money on the table for absolutely no reason. If it's similar in performance to a 7950, they'll price it similarly to 7950. They'll still likely sell every single card they made. Pricing it 100+ USD lower won't change how many cards they sell if it's at the same performance level. The only thing such a move would do is make them make less money.
As long as it's perf/mm^2 or perf/watt is better than whatever it will be competing against it's still a win in those vague words that Charlie used in a previous article.
Of course, there's always the speculation that it's priced at that level because they are anticipating AMD dropping prices drastically once they launch GK 104.
Regards,
SB
Pricing of 299 USD suggests it will be a fair bit slower than 7950. Nvidia aren't ones to seriously discount their cards unless they are forced to as they were with GTX 260/280.
They aren't hurting for marketshare as AMD was with Rv770. They continue to generate more sales than AMD when price/performance is similar. So I see no reason why Nvidia would price something that is competitive with 79xx at a significant discount and then run into supply issues leaving huge sums of money on the table for absolutely no reason. If it's similar in performance to a 7950, they'll price it similarly to 7950. They'll still likely sell every single card they made. Pricing it 100+ USD lower won't change how many cards they sell if it's at the same performance level. The only thing such a move would do is make them make less money.
As long as it's perf/mm^2 or perf/watt is better than whatever it will be competing against it's still a win in those vague words that Charlie used in a previous article.
Of course, there's always the speculation that it's priced at that level because they are anticipating AMD dropping prices drastically once they launch GK 104.
Regards,
SB
That's great. There will be a $299 card launched in three months. Performance is unknown, feature-set is unknown, die-size is unknown.
Nvidia's RV770? I think we need to know significantly more details to make such a judgement. Nvidia had to bring significantly better performance/transistor ratio to start a price war - it would be quite silly to start a price war otherwise, because they couldn't win. Every recent architecture change (G80->GT200-GF110) brought worse perf./tr. ratio, not a better one. Will Kepler change it?
Tahiti's performance per transistor ratio isn't bad. It can be also improved by ~20 % using higher clocks, which would bring performance/transistor even above Cypress. There is also a lot of time to improve drivers. AMD enhanced performance in 5 games by 10-15 % in a month since launch. Now they have another 2-3 months till the launch of Kepler.
Is Kepler a tiler or what to outclass such a solid product?
Nvidia's RV770? I think we need to know significantly more details to make such a judgement. Nvidia had to bring significantly better performance/transistor ratio to start a price war - it would be quite silly to start a price war otherwise, because they couldn't win. Every recent architecture change (G80->GT200-GF110) brought worse perf./tr. ratio, not a better one. Will Kepler change it?
Tahiti's performance per transistor ratio isn't bad. It can be also improved by ~20 % using higher clocks, which would bring performance/transistor even above Cypress. There is also a lot of time to improve drivers. AMD enhanced performance in 5 games by 10-15 % in a month since launch. Now they have another 2-3 months till the launch of Kepler.
Is Kepler a tiler or what to outclass such a solid product?
And as a business if they were not supply constrained nothing suggests they would not target $300. If they could have sold the 580 for $300 and made a goodly profit they would have loved to. It would allow them to sell far more cards. While profit margin might decline and Nvidia has been keen about that I think they would gladly trade a bit for market share especially given the other factors in the marketplace.
Granted I think it is crazy talk myself simply b/c I doubt they could get such a fast card for cheap enough, but if they did it somehow then more power to them.
Is Kepler a tiler or what to outclass such a solid product?
ROFL I've used that joke already; however jokes aside it would be interesting to see if Kepler's caching mechanisms are more efficient this time and by how much bigger caches are compared to the current Fermi generation. The more data you can keep on chip and avoid reads/writes to external memory (and therefore saving bandwidth) the closer you get to any hypothetical TBDR philosophy.
It at least would be a sufficient explanation as to why theoretically a GK104 with potentially less bandwidth than Tahiti can at least scratch its back. If the explanation should not be within the above I'm at my wits (and limited knowledge) end.
Prices are dictated by what the market will bear. Even if a 580 at $300 would have been profitable, it would be more profitable at $500. Lowering its price would force AMD to lower prices but it would also cannibalize their own lower performance parts sales.
If the $299 part is a 560 which is close to Tahiti in performance, they are only selling it at that price because they expect to launch something faster in short order. More likely I expect that they are selling it at $299 because that's what the market will bear for its performance level, so a substantial upgrade to a 560, but nothing that will infringe on the halo cards.
That's not quite true I think. It is different to compare (different manufacturing process, memory bandwidth etc.) and G80->GT200 definitely added lots of transistors without really improving much on features or performance Fermi seems to have at least kept the perf/transistor ratio of GT200 (while adding features and improving some performance areas a lot).
Keeping the same perf/transistor for Kepler looks quite doable to me (even more so because virtually no new features got added).
The key here is the assumed supply issues.
Otherwise your argument and that of some others here seem to assume that there is no demand vs. price relationship, and that these cards will sell solely according to their relative performance vs. AMDs offerings. For devices such as this where there is no actual need other than upgradeitis, that is a dangerous supposition. For some I'm sure buying the latest and greatest is a compulsion (and I'm not joking here) but I think it's a fair assumption that the majority look at what they have and make some sort of cost/fun analysis and decide to buy or not based on that. Thus, unless you're supply bound, it will be a larger volume-and-lower profit/card vs. lower volume-and-higher profit/card balance to be struck. The problem with going lower volume-and-higher profit/card is that while it seems great, it reduces the total size of your market. The more extreme tech entusiasts will upgrade and the less extreme will tend not to. As long as the less compulsive wait around for the next round, all is relatively well, but if they loose interest in the whole thing due to lack of activity, you've managed to shrink your customer base and move deeper into a downward spiral, which interestingly can look fairly good on the balance sheet. For a while.
It would be interesting to see how volumes have changed, say during the last decade. I doubt we'll ever have other than circumstantial evidence.
There was a formula for that?
Other than the generic logic<cache<ideal, there are so many unverifiable variables that I don't see a good way calculating a sound answer.
edit:
And even in that case, it is possible that if density scaling was especially limited from the node prior, the next transition could lead to greater than expected scaling for the next.
Last edited by a moderator:
Well scaling factor should be (new process size)² / (old process size)².
Frankly though I don't actually know what "40nm" etc. refer to (transistor width maybe?). It isn't gate length though (there are quite a bit more parameters which will influence the actual density you can achieve). So this formula is just an approximation (and in any case IIRC tsmc claimed around twice the gate density for 28nm over 40nm).
Well it was the theoretical, perfectly scaling example on perfect processes, the formula was for
Anyway, if we ditch that, how big would one assume for example GF100/110 or Cypress or Cayman would be at 28nm?
I'm trying to do my own guesstimates on what kind of chip GK104 could be, and are $299 price point & Tahiti challenging or beating performance possible in same chip
(Useless speculation, but hey, that's what speculation threads are for!
)edit: I think it was as simple as (xx/yy)^2, for (40/28)^2, so theoretical 100mm^2 chip on 40nm would be ~49mm^2 on 28nm (assuming everything is perfect etc etc)
Last edited by a moderator:
