[SA] Charlie says, "stick with 7970" for the absolute highest performance

[blackened23]

http://semiaccurate.com/forums/showpost.php?p=149785&postcount=321

Charlie,

As a discerning consumer, I am not partial to any particular brand and only employ my resources to obtain the absolute most I can for my dollar. I am a recent purchaser of a 7970, it is still yet opened as I have decided to struggle with the temptation amidst stumbling across your article. I know you have sources to protect and you cannot reveal much beyond a thin, vague wall of data. I however, must ask, when do you believe the enthusiast community as a whole will be privy to tangible data, metrics, pricing and the such?
Stick with the 7970.

-Charlie

Saw this over at OCN..

Seems GK104 will be what I suspected. A well performing part that performs lower than GTX 580 / 7970, but has way better thermals and a smaller die size.

Seems that everything he's been implying is that kepler is a small die with great thermals , which is something NV has never done.

 

[MrK6]

http://semiaccurate.com/forums/showpost.php?p=149785&postcount=321



Saw this over at OCN..

Seems GK104 will be what I suspected. A well performing part that performs lower than GTX 580 / 7970, but has way better thermals and a smaller die size.

Seems that everything he's been implying is that kepler is a small die with great thermals , which is something NV has never done.

Without reading into it too much, I agree. The GK104, being mid range, hopefully continues NVIDIA's line of successful mid range parts since Fermi (GTX 460 -> GTX 560 Ti). Has the 225W TDP rumor been discounted? Otherwise, I wouldn't say it has better thermals.

 

[blackened23]

Without reading into it too much, I agree. The GK104, being mid range, hopefully continues NVIDIA's line of successful mid range parts since Fermi (GTX 460 -> GTX 560 Ti). Has the 225W TDP rumor been discounted? Otherwise, I wouldn't say it has better thermals.

I'd say the chances of it beating a GTX 580, much less the 7970, with a 256 bit memory bus are slim to none.

Still, should be a good part with thermals and small die size.

 

[dangerman1337]

He implied that the most interesting thing about the GK104 was more on the financial side on twitter here: http://twitter.com/#!/CDemerjian/status/160089353861861378

So I'm guessing that it'll be something similar to the 280/260 to the 4870 (though not as wide since the 7970 could probably be sold at 399 and profit comapred to the 280 few years ago).

I do hope it performs slightly better than a 580 at least.

 

[Red Storm]

Having the best mid range card is always better. High end performance king is a niche market and really only used for bragging rights.

 

[blackened23]

Having the best mid range card is always better. High end performance king is for niche markets and bragging rights.

High end also has the biggest profit margin, so its not so clear cut.

 

[boxleitnerb]

I'd say the chances of it beating a GTX 580, much less the 7970, with a 256 bit memory bus are slim to none.

Still, should be a good part with thermals and small die size.

You do realize that the difference between a 560Ti and a 580 is only about 40%. AMD could get this much improvement at equal power consumption. Why wouldn't this be possible for Nvidia? Depends on what you call "beating". I guess GTX580+15% is reasonable.

Computerbase measured that the bandwith of the 384bit bus netted only a 15% increase for the 7970. That still puts it ahead of the 580. Also, 6970 could approach and even beat a 580 by a slim margin in some cases...with a 256bit bus. And finally, Charlie doesn't say anything about performance. You could interpret his short comment in all kinds of ways. You can interpret it in a way that 7970 will be faster in absolute performance, but then please mark it as your own personal interpretation, not as fact. Or put words in other peoples mouths.

Sometimes I feel that you people post rather hastily without really thinking it through

 

[Grooveriding]

This was already pretty clear from way back when we were hearing nvidia wouldn't have high end cards ready any time soon and would release mid-range first. Their mid range card will not be faster than a 7970, will be at best close to a 580 and possibly a little slower.

What would be ironic is if this is all they are going to have in single gpu land and go for a dual-gpu card to get something with halo performance out. I seriously doubt that though, they need something more powerful than a mid range card for their compute customers.

Once this mid range card is out I'm sure we'll suddenly be hearing a whole lot about overclocking if nvidia can get the sort of OC headroom out of 28nm that AMD has. Of course the 7970 hitting 1200core and higher will still lay waste to this card.

Still looking like nvidia cannot release a high end single-gpu halo part on time anymore to compete with AMD. We'll be lucky to get the 680 by the Summer, which is a shame as I would like to compare it to the 7970 when I buy, but I'm not interested in waiting that long to compare it to the AMD 7970 refresh. Too long to wait.

 

[Nintendesert]

If it's close to the 580 in performance and in the $299 range I'll be all over it. If it starts getting up towards the $400 range then I start looking at a 7950 instead. We'll see. But I'd like to see some real numbers from it and a 7950 right now though.

 

[Lonyo]

You do realize that the difference between a 560Ti and a 580 is only about 40%. AMD could get this much improvement at equal power consumption. Why wouldn't this be possible for Nvidia? Depends on what you call "beating". I guess GTX580+15% is reasonable.

Computerbase measured that the bandwith of the 384bit bus netted only a 15% increase for the 7970. That still puts it ahead of the 580. Also, 6970 could approach and even beat a 580 by a slim margin in some cases...with a 256bit bus. And finally, Charlie doesn't say anything about performance. You could interpret his short comment in all kinds of ways. You can interpret it in a way that 7970 will be faster in absolute performance, but then please mark it as your own personal interpretation, not as fact. Or put words in other peoples mouths.

Sometimes I feel that you people post rather hastily without really thinking it through

(http://www.computerbase.de/artikel/grafikkarten/2011/test-amd-radeon-hd-7970/20/ source on the Computerbase comment, for anyone interested as I was).

In addition, the GF100 series and its ilk were SERIOUSLY memory handicapped with respect to GDDR speeds, and this is a fact.

The TOP GDDR5 speed for NV was 4008MHz. For AMD? 5500MHz

*IF* NV have fixed their memory controller, then a 384-bit 4008MHz bus would give 192.4GB/s bandwidth. 256-bit at 5500MHz? 176GB/s.
That means that a speculative AMD performing memory controller would have 91.5% of the memory bandwidth of a GTX580, even if it was only 256-bit vs 384-bit.

5500MHz is something which can be done, AMD have done it more than once.
Whether NV is able to do it yet remains to be seen, but don't get hung up on the memory bus when it doesn't necessarily equate to bandwidth.

 

[Genx87]

This is actually good news for me. I am planning on going two mid range cards in SLI this round just to say I have tried it.

 

[96Firebird]

If it's close to the 580 in performance and in the $299 range I'll be all over it. If it starts getting up towards the $400 range then I start looking at a 7950 instead. We'll see. But I'd like to see some real numbers from it and a 7950 right now though.

Same here... Except I am sticking to $300 and won't go $100 above that, maybe a couple of bucks though. If GK104 is too much over $300, I will continue to wait.

 

[blastingcap]

High end also has the biggest profit margin, so its not so clear cut.

Fixed cost of R&D can't be made up by selling small numbers of high-end products at high cost; you need volume.

High end doesn't sell nearly as many units as low- and-mid. And now that APU/integrated graphics are strong enough for video decoding and even light gaming, the relative importance of midrange is magnified.

Nvidia can sell their chips in pro/HPC garb like Tesla units for much, much more than gaming-grade cards, anyway; that is their real high end and allows them potential for cross-subsidizing their gaming cards. So Nvidia has much greater pricing latitude than AMD when it comes to gaming cards.

Given the above, from a long-term financial perspective, Nvidia would probably do best by grabbing as much midrange as possible and matching AMD on the high end, in price/perf. This results in maintaining or increasing market share in those critical areas, which weakens AMD in the long run as they have fewer units with which to spread fixed costs. Market share is very important with such high fixed costs.

People forget sometimes that Nvidia's premium pricing was mostly done when they had double the market share of their rival. AMD managed to battle Nvidia to a draw with 5xxx/6xxx vs. 4xx/5xx but Nvidia wants that market share back, and if they fixed Fermi then they have a chip with which to do it. (Even during the Fermi days, when Nvidia had a flawed chip, Nvidia was willing to semi-match AMD on pricing. Recall that GTX 480 was priced at $500. Nvidia knew if they kept to their traditional $600+ price point, they would be laughed at. With such a flawed and expensive chip, though, it's understandable why Nvidia didn't want to outright undercut or even match AMD on pricing... but if Kepler if a fixed Fermi then it will be much easier to match or undercut AMD on pricing.)

Even if Nvidia makes a little less money in the gaming sector, they can make up the difference in pro graphics/HPC where Nvidia has a de facto monopoly. Invest proceeds in mobile where the real action will be, and of course back into graphics R&D for gaming, pro, and HPC markets.

Oh wait, that's exactly what Nvidia has been doing. Nvidia's CEO is a belligerent jerk sometimes, but he isn't stupid.

 

[Arkadrel]

Having the best mid range card is always better. High end performance king is a niche market and really only used for bragging rights.

^ this.

Sounds like nvidia will have really good mid-high range cards this round.
So what if it doesnt beat a 580 or 7950, if its close and alot cheaper, its gonna sell like hot cakes.

Also nvidia will still have bigger cards comeing than this one right? so just gotta wait for those if you want something that ll have a chance to beat the 79xx cards.

 

[bunnyfubbles]

(http://www.computerbase.de/artikel/grafikkarten/2011/test-amd-radeon-hd-7970/20/ source on the Computerbase comment, for anyone interested as I was).

In addition, the GF100 series and its ilk were SERIOUSLY memory handicapped with respect to GDDR speeds, and this is a fact.

The TOP GDDR5 speed for NV was 4008MHz. For AMD? 5500MHz

*IF* NV have fixed their memory controller, then a 384-bit 4008MHz bus would give 192.4GB/s bandwidth. 256-bit at 5500MHz? 176GB/s.
That means that a speculative AMD performing memory controller would have 91.5% of the memory bandwidth of a GTX580, even if it was only 256-bit vs 384-bit.

5500MHz is something which can be done, AMD have done it more than once.
Whether NV is able to do it yet remains to be seen, but don't get hung up on the memory bus when it doesn't necessarily equate to bandwidth.

was going to point this out to those who were writing off the potential for GK104 to outperform the GTX580

the GTX285 had some of the fastest GDDR3 possible on a 512bit bus for a bandwidth of 159GB/s, meanwhile the GTX460 and 560 TI had 256bit bus with GDDR5 and resultant bandwidths of only 115GB/s and 128GB/s respectively, the 460 was just as fast if not a tad faster and the 560 is clearly faster despite this apparent handicap. Basically nVidia was extremely behind in GDDR5 development last round. Its always possible they're still woefully behind, but its just as possible that they've caught up.

Perhaps for some of the most demanding high resolution games and multi monitor setups the GK104 might not cut the mustard, however I wouldn't be shocked at all if it cause a major ruckus at 1920 resolutions and could be a major sweet-spot product, especially if it is going to undercut the 7900s in price it would fit right in line with the 1920 demographic.

 

[Arkadrel]

Im wondering how much TPD AMD uses on its extra ram?
And how much it costs them in $ to add extra ram on them?

Because I honestly still dont think theirs really much reason for more than 1.5GB on the cards.
I think AMD went overboard going with 3GB on their 79xx cards.

Im kinda hopeing they do some 7950 and 7970's with only 1.5gb ram, if that allows them to sell the cards 30-40$ or so cheaper, that would be a win for most people that dont want to do eyefinity setups.

 

[bunnyfubbles]

better to go overboard than the under, AMD already established a new standard for buffer capacity @ 2GB with their 6900s, they would have looked extremely foolish to have put out 1.5GB parts after already coming out with 2GB parts. Really their only choice was 2GB or higher, and a 384bit bus limited that choice to 3GB.

We'll probably see some 1.5GB cards in a 7950 or 7930 flavor, but likely never for the 7970.

 

[SlowSpyder]

Im wondering how much TPD AMD uses on its extra ram?
And how much it costs them in $ to add extra ram on them?

Because I honestly still dont think theirs really much reason for more than 1.5GB on the cards.
I think AMD went overboard going with 3GB on their 79xx cards.

Im kinda hopeing they do some 7950 and 7970's with only 1.5gb ram, if that allows them to sell the cards 30-40$ or so cheaper, that would be a win for most people that dont want to do eyefinity setups.

I wouldn't be shocked if those eventually came about. Remember 256MB 8800GT's and 2900's? 320MB 8800GTS cards, etc.

 

[Arkadrel]

they would have looked extremely foolish to have put out 1.5GB parts after already coming out with 2GB parts.

I dont understand that part, what if 1gb was to little for modern games = had to do 2gb because of 256bit bus.

With 384bit bus, can reach 1.5gb, doesnt have to do 3gb, if not needed.

I dont think it would look stupid going backwards, in size of ram.
There is such a thing as too much..... its too much when 90% dont need it.

(pulled outta my arse that #, however I remember there being a thread here about 1.5gb vs 2gb, and someone did some tests and it basically amounted to there was no reason for 2gb atm)

I wouldnt blink a eye at a 1.5gb version if it was 30$ cheaper or so, thatd be the version that I bought.

 

[mak360]

Im wondering how much TPD AMD uses on its extra ram? And how much it costs them in $ to add extra ram on them? Because I honestly still dont think theirs really much reason for more than 1.5GB on the cards. I think AMD went overboard going with 3GB on their 79xx cards.

there are hd7950 1.5gb cards listed to come from AIB partners. both nvidia and amd are bulk buyers of memory, so its cheap for AIB/small OEM partners etc.

Even if Nvidia makes a little less money in the gaming sector, they can make up the difference in pro graphics/HPC where Nvidia has a de facto monopoly. Invest proceeds in mobile where the real action will be, and of course back into graphics R&D for gaming, pro, and HPC markets.

GCN is to counter exactly that, which nvidia has enjoyed pretty much without challenge, i think that may change with the next 2 years as software becomes more viable with GCN architecture


I think the battle will be where the gtx580 and 7950 is, for the gk104, i think its going to be cheaper (thats the smoking gun folks)

 

[Arkadrel]

Gonna be intresting with 580, GK104, 7890 and 7950 all in very close range of each other.

 

[BTRY B 529th FA BN]

I have a GTX 480 and I'm waiting

 

[Panopticon]

Im wondering how much TPD AMD uses on its extra ram?
And how much it costs them in $ to add extra ram on them?

Because I honestly still dont think theirs really much reason for more than 1.5GB on the cards.
I think AMD went overboard going with 3GB on their 79xx cards.

Im kinda hopeing they do some 7950 and 7970's with only 1.5gb ram, if that allows them to sell the cards 30-40$ or so cheaper, that would be a win for most people that dont want to do eyefinity setups.

Whats the point of having a card capable of running textures on ultra without the vram to push it. I game at 1080p and have seen usage biting at the heels of 2 gb. A little future proofing never hurts and theres no point in having a card that is bottlenecked by vram.

 

[wahdangun]

its just like HD 5870 all over again.

remember when nvdia launched GT240 ? although AMD have launched HD 5870 for like 2 month earlier

 

[AdamK47]

You're reading into things. I don't see where the person is asking if he should stick with the 7970 for absolute highest performance.