nVidia wins this round - Charlie D.

[3DVagabond]

Speculation, what else are we doing in this thread?

I'd just assume Nvida isn't going to go backwards in per core cuda performance.

But yes, until we know exactly what we're looking at as far as specs such as TMU/ROP/Poly everyone is just stating opinions based on information they got from their nether regions.

Well, considering the rumors that show the BigGK as having 2048 cores (4x Fermi) and also the possibility of dropping hot clocks, we can't assume that per core performance is going to be the same or go up.

I realize we all are speculating. That doesn't mean anything goes, though. We need something to base it on. I'll be glad when we have more than industry rumors. Just making random comparisons though isn't really a reasonable way to go.

 

[blastingcap]

Which 28nm process?

Sounds like a problem for everybody. http://www.techpowerup.com/159189/28-nm-struggles-TSMC-amp-GlobalFoundries.html

 

[BallaTheFeared]

Well, considering the rumors that show the BigGK as having 2048 cores (4x Fermi) and also the possibility of dropping hot clocks, we can't assume that per core performance is going to be the same or go up.

I realize we all are speculating. That doesn't mean anything goes, though. We need something to base it on. I'll be glad when we have more than industry rumors. Just making random comparisons though isn't really a reasonable way to go.

Well if it has 2000 on the top, than the mid ranged GK104 wouldn't have 600-700 cores, would it?

Since that would put it's effective performance closer to the GTX 460 in cuda apps.

But again who really knows, what's going on besides Nvidia and some select people I'm sure at this point.

However Nvidia needs to keep their design progressing forwarded in GPGPU which means a 760 is going to go backwards in CUDA performance compared to the 560TI if you assume 2000 on the top end and 600-700 on the mid ranged. So either it's more or they didn't drop hot clocks.

 

[MagnusTheBrewer]

Just to add a little fuel, check out this blind pc test.

 

[Lonbjerg]

It's Char-lie...who cares what he says?

 

[meloz]

Nvidia fanboys taking Charlie/SA rumours as gospel.

Charlie / SA have not been, uh, charitable to nvidia these past few years. You should read some of his editorials slamming nvidia and their 'failed' products. Quiet legendary, almost read like troll posts you find on certain forums.

This is why everyone is giving his article so much consideration this time. If nvidia's self-proclaimed public enemy No. 1 is giving them kudos, the new product must be really good.

I do think people are getting a bit carried away: a $300 card from Team Green will not outperform a $550 card from Team Red. It makes no business sense for nvidia. For nvidia it is sufficient to beat AMD in performance/price ratio by a small but convincing amount (at all price points) to win this battle, they gain nothing by depressing their own revenue by offering an extremely great deal to consumers. They love cash as much as AMD and the rest of us all.

 

[Lonyo]

Charlie / SA have not been, uh, charitable to nvidia these past few years. You should read some of his editorials slamming nvidia and their 'failed' products. Quiet legendary, almost read like troll posts you find on certain forums.

This is why everyone is giving his article so much consideration this time. If nvidia's self-proclaimed public enemy No. 1 is giving them kudos, the new product must be really good.

I do think people are getting a bit carried away: a $300 card from Team Green will not outperform a $550 card from Team Red. It makes no business sense for nvidia. For nvidia it is sufficient to beat AMD in performance/price ratio by a small but convincing amount (at all price points) to win this battle, they gain nothing by depressing their own revenue by offering an extremely great deal to consumers. They love cash as much as AMD and the rest of us all.

Really?
Tell that to the GTX280/260 vs HD4800 series.
Maybe in that case not outperforming, but certainly making the price look stupid.

 

[boxleitnerb]

It could be if the 550$ performance card from Team Red is ridicilously overpriced. I doubt it, though. My guess is 7970 -10% at 160-170W TDP and 300$.

 

[lifeblood]

So you're of the beleif that AMD wouldn't gouge their customers because they were first out of the gate despite nobody actually knowing what kind of performance Nvidia's new cards would bring?

Physics apply to both the same way. They both are using the TSMC 32nm process, their engineers are just as smart, the cost of all the other components are roughly the same. I accept AMD is price gouging, I accept nVidia has more experience at designing a compute architecture, but NVs midrange card beating AMDs high end card while costing $200 less? That is too much of a reach.

I suspect AMD has cut Charlie off and he's trying to suck up to nVidia in hopes are reestablishing his relationship.

 

[Lonbjerg]

Physics apply to both the same way. They both are using the TSMC 32nm process, their engineers are just as smart, the cost of all the other components are roughly the same. I accept AMD is price gouging, I accept nVidia has more experience at designing a compute architecture, but NVs midrange card beating AMDs high end card while costing $200 less? That is too much of a reach.

I suspect AMD has cut Charlie off and he's trying to suck up to nVidia in hopes are reestablishing his relationship.

You mean 28nm, not 32nm right?

 

[SirPauly]

Physics apply to both the same way. They both are using the TSMC 32nm process, their engineers are just as smart, the cost of all the other components are roughly the same. I accept AMD is price gouging, I accept nVidia has more experience at designing a compute architecture, but NVs midrange card beating AMDs high end card while costing $200 less? That is too much of a reach.

I suspect AMD has cut Charlie off and he's trying to suck up to nVidia in hopes are reestablishing his relationship.

It depends on what context! AMD's performance may look like premium based when compared to 40nm product but why would anyone think that is premium based when the 28nm competition comes into play? Especially in the context of AMD raising MSRP?

Maybe the rumored 299 dollar price-point isn't being compared to AMD but nVidia's potential future products. nVidia isn't just competing with AMD but with themselves to give their customers reasons to upgrade as well.

 

[lifeblood]

You mean 28nm, not 32nm right?

Sorry, yes. 28nm.

 

[boxleitnerb]

It depends on what context! AMD's performance may look like premium based when compared to 40nm product but why would anyone think that is premium based when the 28nm competition comes into play? Especially in the context of AMD raising MSRP?

Maybe the rumored 299 dollar price-point isn't being compared to AMD but nVidia's potential future products. nVidia isn't just competing with AMD but with themselves to give their customers reasons to upgrade as well.

This. I will have a laugh when people realize this and GK104 is not priced at 500$.

 

[lifeblood]

It depends on what context! AMD's performance may look like premium based when compared to 40nm product but why would anyone think that is premium based when the 28nm competition comes into play? Especially in the context of AMD raising MSRP?

Maybe the rumored 299 dollar price-point isn't being compared to AMD but nVidia's potential future products. nVidia isn't just competing with AMD but with themselves to give their customers reasons to upgrade as well.

I'm sorry, I'm not sure what your trying to say.

A 28nm wafer costs X dollars. That cost is then spread across how ever many GPUs you can get out of it (I'm generalizing costs. I know their are some differences in what AMD pays versus NV). In general, a 300mm NV GPU will be as fast as a 300mm AMD GPU. So, for NV to make a $300 GPU that's faster than a $500 AMD GPU, AMD must be price gouging by $200 or NV has indeed made a CPU that's significantly smaller (therefore cheaper) yet faster than AMDs GPU.

I accept AMD is price gouging, but by $200? I accept NV, being more experienced at compute, can build a faster GPU per square mm, but by that much?

Again, if I'm wrong I will be more than pleased to eat my words. I'm in the market for a new GPU and this has definitely got my interest.

If its true, how could it happen?
1. AMD is price gouging (given)
2. NV has really really created a faster architecture (NV does have more compute experience).
3. AMD released a GPU that wasn't fully ready-for-prime-time but was ready-enough (see apoppin's article on ABT).

It would really have to be a combination of all three things in order to create a smack-down of that magnitude.

 

[boxleitnerb]

Yeah, I think it is a combination, too.

Tahiti is 389mm2. If GK104 is around 300mm2 for instance (and NV maybe has a better deal with TMSC) and AMD charged a heavy premium, why not?

 

[Olikan]

now charlie says to ....Stick with the 7970.

 

[MrK6]

now charlie says to ....Stick with the 7970.

Well, it seems a lot of people got trolled for page clicks. I have no doubt GK104 will be an interesting part. In fact, I think the best parts NVIDIA's delivered since Fermi have been its mid range (GTX 460 and then the GTX 560 Ti). But a 7970-killer for $300 doesn't even make physical sense, nevermind business sense.

 

[AtenRa]

Yeah, I think it is a combination, too.

Tahiti is 389mm2. If GK104 is around 300mm2 for instance (and NV maybe has a better deal with TMSC) and AMD charged a heavy premium, why not?

AMD Tahiti is 365mm2

 

[AtenRa]

Just to get this straight,

If NV just shrink GF110 at 28nm TSMC HKMG then it would be close to 300mm2. So it could have a 300mm2 GTX580 performance card. Now we know that GF104/114 is close to 330mm2 and they sell it at ~$250 with 1GB memory. I will add $50 for higher manufacturing cost (28nm) and 1GB more memory and we come to $299

 

[Quantos]

Just to get this straight,

If NV just shrink GF110 at 28nm TSMC HKMG then it would be close to 300mm2. So it could have a 300mm2 GTX580 performance card. Now we know that GF104/114 is close to 330mm2 and they sell it at ~$250 with 1GB memory. I will add $50 for higher manufacturing cost (28nm) and 1GB more memory and we come to $299

Well, it also depends what's the architecture used. Is it a shrink or a new architecture? The cost of a new architecture is probably greater. Also, you assume a $50 premium, which is very in fact an unknown number. We have the 7950 between the 7970 and 580 at about $450 (maybe 400 if there's a reference, I'm not sure?). A card performing at the 580's level is therefore worth, depending on the 7950's price, either 350 or 400 bucks. The question that remains is: can AMD simply lower the price of their products accordingly? If it's just a $50 difference, perhaps that's something AMD planned. However, it's entirely possible that what they sell their cards at is the lower they can get to. The latter does seem less probable to me, though...

 

[VulgarDisplay]

Just to get this straight,

If NV just shrink GF110 at 28nm TSMC HKMG then it would be close to 300mm2. So it could have a 300mm2 GTX580 performance card. Now we know that GF104/114 is close to 330mm2 and they sell it at ~$250 with 1GB memory. I will add $50 for higher manufacturing cost (28nm) and 1GB more memory and we come to $299

A 300mm2 card with gtx580 performance would not bode well for nvidia since Tahiti is 365mm2 and 24% faster on average.

 

[boxleitnerb]

AMD Tahiti is 365mm2

Thx. So GK104 could actually be around 300-320.

 

[Ajay]

now charlie says to ....Stick with the 7970.

Wow, that brings everything into focus

With statements like that, it's looking like CD is only looking for web hits.

 

[SirPauly]

Well, it seems a lot of people got trolled for page clicks. I have no doubt GK104 will be an interesting part. In fact, I think the best parts NVIDIA's delivered since Fermi have been its mid range (GTX 460 and then the GTX 560 Ti). But a 7970-killer for $300 doesn't even make physical sense, nevermind business sense.

And each of them offered more performance than their high-end counterparts from their past generation -- GTX 285. Why wouldn't this mid-range chip offer more than the GTX 580?

It's a pretty tough sell to offer nVidia's mid-range to completely and utterly destroy the raw performance of the HD 7970 and be a 7970 killer, but can see it compete like the HD 4870 did to the GTX 280.

It may not be a killer in raw performance but may be a killer in efficiency metrics.

 

[Grooveriding]

Pretty easy to troll people when you tell them you might get the performance of a $550 card for $300. After the way nvidia screwed up their 40nm intro, this shows they learned how to deal with that mistake and not do it again as they are obviously behind on their single gpu halo part again.

Last time they mucked around and delayed for 6 months trying to fix GTX480/470 and ended up rushing out a card that was not quite there. Now on 28nm nvidia is again behind with a single gpu halo, so instead of delaying any 28nm release and getting their halo chip ready they will release a mid range card. Probably what they should of done on 40nm - release the 460 and worry about fixing the 480/470 afterwards and getting it out. This way they at least have something out.

This GK104 will likely play out like the GTX460. Faster and equal to a 580 sometimes, slower at other times. Likely dependent on the game and if it likes ROPS, shaders, bandwidth etc. It makes sense since AMD has no mid-range 28nnm out yet and has put out 7970/7950 top-end only so far.

Buyers wanting to spend $300 to $350 will buy a mid-range GK104, buyers wanting the best performance available will buy a $550 7970. The 7950 could lose some sales to this card though. The smart move by AMD would be to have AIBs roll out factory OCed 7950s like crazy. The 7970 will stand on its own at stock, but the 7950 would look a lot better at $450 with factory OC models putting out stock 7970 performance.