Official HD 7970 Specs

[John P. Myers]

652

Review, Benchmarking and Overclocking Results (http://www.legitreviews.com/article/1805/1/)

EDIT: Gotta say i'm really not impressed. EVGA's 560 Ti 2win beats it in every benchmark and is $30 cheaper.

[Fire$torm]

Review, Benchmarking and Overclocking Results (http://www.legitreviews.com/article/1805/1/)

EDIT: Gotta say i'm really not impressed. EVGA's 560 Ti 2win beats it in every benchmark and is $30 cheaper.

But, but, but........ The 560 is not a Radeon. :((

[coronicus]

Aye, again to much hipe behind a product. This so called 60%+ performance ended up only really being 20-30% in most cases cosidering the gaming side of things. It does make me feel better about my purchase of the two 6950's. We will have to see wether it will give this magical 60% on projects, ill believe that when i see real numbers.

[DrPop]

Don't know why I always seem like the one defending AMD - especially when I'm really not happy with them either, but... I really feel ya'll kind of missing the point a little. This is 3.7GFLOPS @ 250W. This is actually a pretty big breakthrough.

That 560Ti x2 card will run @ 340W. That means the CUDA card consumes 90W or 36% more power. If you pay SoCal business or San Fran PG&E electric rates like Zombie and I do - this is a big thing. If you live in a place with cheap, gov. subsidized or cheap hydro power, maybe it won't matter - yet. :p;)

As crunchers, we could really care less about the frame rates in games, because those vary with several parameters. What it boils down to for us, is the CUDA / Stream cores, and how fast they run.
The 7970 has 2,048 cores, compared to 1536 cores on the 6970. This is 512 more cores, a 33.33% increase on the surface; but if you look at the clock speed, it's more. 7970 clock is 925, and 6970 clock is 880, for a 5% increase on the new card. 5% doesn't sound like much, but over 2,048 cores, that's equivelant to an extra 102.5 cores!
So, crunching wise, it will equate to at least a (512+102.5=614.5/1536=) 40% gain off the top. Then, if the RAM or other things are fast, the architecture is tweaked, etc, you will be able to pull a little more out of it as well.

What does this mean for us?
If you own an O/Ced 5870 or 6970, and it takes X minutes to run a Moo! WU, it will take only .6X minutes to run the same Moo! WU on the new 7970. (EDIT: All things being equal, of course) That is a pretty big jump. And it could be even faster if they eek out some architectural and other tweaks. And this is with NO increase in power consumption over the 6970!

[John P. Myers]

I won't argue that it's not more power efficient, and it is in fact the fastest single GPU card available. No doubting that. Though the RAM speed is the same on this as it was on the 6970, so no increase in speed there, just the stream clock. My issue is that before this release, there were reports stating that it would be as fast or just a hair faster than a GTX 590. Unfortunately it turns out that it's not. So ok, let's say it is 40% faster than a 6970. This thing (in it's basic, reference design) will retail for $549. A basic 6970 goes for $310 (some are as low as $300 when you include rebates). $310 + 40% = $434. But it's another $115 on top of that. Also, the power used by this card is the same as a 6970. It just does a little more work per watt.

So you could ask: how long would you have to use a 7970 instead of a 6970 to save $115 on your bill? But since both cards have the same TDP, you can't really even ask that. But fine, you're getting a card 40% faster than a 6970 for 77.5% more money. Awesome.

[DrPop]

So you could ask: how long would you have to use a 7970 instead of a 6970 to save $115 on your bill? But since both cards have the same TDP, you can't really even ask that. But fine, you're getting a card 40% faster than a 6970 for 77.5% more money. Awesome.

hehe...well, I had no idea what the price was, and I'm not even saying anyone should get it! ;) I was just saying it seems pretty dang efficient to me. For sheer GFLOPS/Watt, the 5870 was king, and now this blows the old cards out of the water.

Certainly it won't be cheaper to run the 7970 over the 6970, but you'll get 40% more credit return for every $ you spend on the electric bill. When I crunch all my rigs 24/7 to hit ~ 1Mil / day, the portion of my electric bill that is attributed to crunching is quite significant. Would I love to hit 1.4 Mil / day for the same money? Of course I would! :D But...I agree with you - I don't have that kind of cash just laying around upfront to get that return...so I'm stuck with older generation cards. :p:o

[John P. Myers]

lol well my best (and only) ATI card is a 5870. i'd love to have a 7970 :D Another good thing about the 7970 is when the waterblocks come out, it can easily be a single-slot card. What would be really impressive is if maybe the 7990 could be single-slot too. That would enable people to pack a serious amount of crunching power in a single box. Can you imagine? :-?

[zombie67]

That 560Ti x2 card will run @ 340W. That means the CUDA card consumes 90W or 36% more power. If you pay SoCal business or San Fran PG&E electric rates like Zombie and I do - this is a big thing. If you live in a place with cheap, gov. subsidized or cheap hydro power, maybe it won't matter - yet. :p;)

Okay, I am going to have to be "that guy". *Tax Payer Subsidized.* There is no such thing as "gov. subsidized" anything. The government has no money of its own. It's just some of us, paying for others of us....at the point of the IRS gun.

Okay, everyone groan, and go back to what you were doing...

[DrPop]

heh...well, suppose I'd be labeled a "conservative" in darn near any argument, so I agree with you . . . it's a bunch of B.S. but I didn't know what other way to put it. "Tax payers, especially small business owners, being robbed blind to artificially maintain your standard of living" is probably the best way to put it. :mad::p:o But then...I suppose we don't want to get everyone all up in arms at us, now do we?:rolleyes: :cool:
EDIT: at some point we also have to bring in the "selling our future generations' souls to the Chinese and anyone else who will buy our debt to artificially maintain our standard of living..." But that's probably getting too deep for some... ;)

Meanwhile...back at the ranch...JPM is laughing all the way to the bank with his CUDA card...:D

[DrPop]

Back to the 7970: it has 384bit memory bus instead of 256bit, so even though the RAM clock speed is the same, the bandwidth is considerably higher now.

Also, just did some good reading on this - it is the first GPU to sport the new GCN architecture, which is AMD's new arch to do what CUDA does for NVIDIA. This is actually excellent news for us crunchers. It will be interesting to see how AMD does with the server market now and computing on the GPU. If they can get the performance per watt high enough, they could pose a good challenge to NVIDIA, which has been dominating this segment since CUDA introduction.
With the new architecture, I'm not surprised at the lower than expected results with this new GPU. They are going to have to recompile some software to be optimized for the new GCN before we can see the true raw crunching Horse Power. Why is no one bringing this up in the reviews? Seems like the software always lags the hardware, but that's just tue way it has to be, I guess.

[Fire$torm]

Also keep in mind that TDP in and of itself does not equal high efficiency. It is a major factor of the formula but not the only contributor. Vregs can be huge power wasters. Add to that bus controllers, memory and its controller, interlink controllers and overall component count are just some of the things to include in such an analysis. These are the types of things companies like AMD, Intel, IBM and others have begun to take more seriously then in years past.

[John P. Myers]

Also keep in mind that TDP in and of itself does not equal high efficiency. It is a major factor of the formula but not the only contributor. Vregs can be huge power wasters. Add to that bus controllers, memory and its controller, interlink controllers and overall component count are just some of the things to include in such an analysis. These are the types of things companies like AMD, Intel, IBM and others have begun to take more seriously then in years past.

All very true, so it should be pointed out the 7970 has 3GB RAM standard, which is more than any other of AMD's single-GPU cards. If they had left it at 2GB like the 6970, the power draw would have been a bit less. Also they used analog VRMs instead of digital. That is another loss of efficiency.

Yes, they did increase the memory bandwidth, but as you know, it takes little bandwidth to crunch. As hi-tech as this card is supposed to be, it should be pointed out that ATI has used 512bit bandwidth before, as far back as 2005 with the Radeon X800 XT. What happened? But again, not that us crunchers really need it. For those that are also gamers though, this is a good thing :) Except that gamers *really* hate that the card only has 32 ROPs. That seems to be the most common complaint i've seen on various boards.

[Tank_Master]

Though the RAM speed is the same on this as it was on the 6970, so no increase in speed there, just the stream clock.

That's a bit incorrect. Its the same frequency, but 384 bit bus instead of 256, so it's bandwidth is 1.5x that of the 6970.

■176 GB/s memory bandwidth (maximum) (6970)
■264 GB/s memory bandwidth (maximum) (7970)

[John P. Myers]

That's a bit incorrect. Its the same frequency, but 384 bit bus instead of 256, so it's bandwidth is 1.5x that of the 6970.

■176 GB/s memory bandwidth (maximum) (6970)
■264 GB/s memory bandwidth (maximum) (7970)

Right but when crunching, if you wanted to move a 64bit double precision floating point number to/from the VRAM, it takes the same amount of time whether it's 256bit or 384bit :) It's only when you need to move more than 256bits of data to the VRAM, all at the exact same time that 384bit is actually faster.