AMD to Launch GPU Codenamed Hawaii in Hawaii Sept 25

[mAJORD]

If tahiti scaled better than keplar, the 7970 would have gotten out performed by the gtx 680 in the first place. Remember, the gtx 680 is 20 percent smaller than a 7970.

Also a 7970 and 7950 at the same clocks are 3-5 percent difference in performance(adding more shaders wouldn't do much with the current GCN architecture). Tahiti at the upper end scales worse. I think GCN scales better at the lower end, but reaches its ceiling quicker and scales worse at the top end, hence why gtx 680 and 7970 are so competitive and gtx770 and 7970 ghz are so close.

Improving power consumption is hardly trivial.

Pointless comparison / speculation.. GK104 scaled poorly between 670 and 680 also.. and it should have scaled better than 7950/7970 since there's also a difference in TMU count. By your logic, Titan should have been a complete flop.

Correct me if I'm wrong, but as it stands:

A. We don't know the concrete specs..
B. We don't know architecture detail (I don't believe?)
C. Unlike GK104 to GK110 , Tahiti is already compute orientated (i.e "comprimised") so even if you were to try and speculate, you'd have to somehow take this into account.

If Nvidia are getting x performance from Titan on 28nm, to assume AMD can not (if that there's goal) based on a comparison of old technology is SILLY.

[Darxide]

Will this September thing be a paper launch or will we actually have some cards to buy?
I've just put my new rig together and decided to delay my GPU decision as the new 9xxx are so close to launch.
I bought a cheapo VTX 7750 to tide me over. I can game on it, but only just lol.

[MaddMuppet]

Will this September thing be a paper launch or will we actually have some cards to buy?
I've just put my new rig together and decided to delay my GPU decision as the new 9xxx are so close to launch.
I bought a cheapo VTX 7750 to tide me over. I can game on it, but only just lol.

With the way they have dropped the price on the 7990, I think we may well see some cards on launch. I certainly hope so and I also hope they are able to at least match Titan. With AMD's aggressive pricing, this could be a good win for us the consumer.

[Lanek]

Will this September thing be a paper launch or will we actually have some cards to buy?
I've just put my new rig together and decided to delay my GPU decision as the new 9xxx are so close to launch.
I bought a cheapo VTX 7750 to tide me over. I can game on it, but only just lol.

In reality, articles about it bring 2 different possibilities, some write will launch.. for other, the words spread are " technical presentation " of the architectures, cards etc..

But whatever if this is not the last week of september, this will be few week after.

[DefStar]

Pointless comparison / speculation.. GK104 scaled poorly between 670 and 680 also.. and it should have scaled better than 7950/7970 since there's also a difference in TMU count. By your logic, Titan should have been a complete flop.

Correct me if I'm wrong, but as it stands:

A. We don't know the concrete specs..
B. We don't know architecture detail (I don't believe?)
C. Unlike GK104 to GK110 , Tahiti is already compute orientated (i.e "comprimised") so even if you were to try and speculate, you'd have to somehow take this into account.

If Nvidia are getting x performance from Titan on 28nm, to assume AMD can not (if that there's goal) based on a comparison of old technology is SILLY.

Exactly...there is some kind of myth out there that AMD is somehow 2nd rate GPU manufacturers when the fact is, they have consistently pulled more G-flops with a smaller architecture over-and-over. If anything, AMD does a lot more hardware wise then Nvidia. And now with AMD's huge push to fix their drivers and console games being naturally optimized for AMD, things are only looking better for the red team...GPU wise that is. This isn't fan-boism...this is simply recognizing who is doing more for my money.

If Nvidia really believes i'm gong to pay $300 more because their logo is on it...they better think again.
There are things like 2D Light Boost which are really nice features of Nvidia, but if Nvidia wants to be competitive they need to price their cards similar to AMD's. My Samsung monitor can mimic 2D Light Boost so I'm not missing out on anything by saving $300 with AMD and getting equal if not better performance.

And don't get me wrong, I love you Nvidia fan-boys...you guys are the reason I can buy an AMD card for so cheap. Keep it up...please for the sake of my wallet, keep it up.

[boxleitnerb]

GK104 didn't scale poorly, the 680 was just more bandwidth constrained.
To compare GFLOPs is only helpful up to a point. Until Kepler, Nvidia had a lot less GFLOPs than AMD yet they always were faster in the end. And even now, AMD gets less gaming performance (about 10%) out of a certain amount of GFLOPs (when all other factors are the same or very similar, compare Tahiti LE vs GK104 or GK106 vs Pitcairn) than Nvidia. If AMD improves their frontend, that will change.

[LordEC911]

GK104 didn't scale poorly, the 680 was just more bandwidth constrained.

Tahiti didn't scale poorly, the 7970 was just more frontend constrained.

[TRANCEFORMER]

http://www.amd.com/us/products/deskt...es/8970.aspx#3

[AMDDeathstar]

http://www.amd.com/us/products/deskt...es/8970.aspx#3

Just 7xxx series rebranded for OEM's

[TRANCEFORMER]

Just 7xxx series rebranded for OEM's

yes sorry didnt realize that

[boxleitnerb]

Tahiti didn't scale poorly, the 7970 was just more frontend constrained.

So? It's quite easy to give a GPU more bandwidth. To efficiently distribute work to the ALUs is much more difficult.

[LordEC911]

So? It's quite easy to give a GPU more bandwidth. To efficiently distribute work to the ALUs is much more difficult.

Both are relatively substantial changes...

I was just pointing out that you can't ignore scaling inefficiencies just because the bottleneck is different...

[boxleitnerb]

Both are relatively substantial changes...

I was just pointing out that you can't ignore scaling inefficiencies just because the bottleneck is different...

Nope, just put 7 or 8 Gbps memory on the card and you're good to go. In the other case, there isn't such an easy solution.
Kepler scales perfectly with more units. 680 is 2.3x as fast as the 650 Ti (50% of everything, a tad lower clocks). GCN doesn't achieve that, the 7870 is not twice as fast as the 7770, "only" like 90% faster.

[LordEC911]

Nope, just put 7 or 8 Gbps memory on the card and you're good to go. In the other case, there isn't such an easy solution.
Kepler scales perfectly with more units. 680 is 2.3x as fast as the 650 Ti (50% of everything, a tad lower clocks). GCN doesn't achieve that, the 7870 is not twice as fast as the 7770, "only" like 90% faster.

Ah, so in your world the MC's are on the GDDR5 ICs...
Kepler's MCs would need a rework to guarantete +7ghz speeds.

Easy solution for FE is to add more ACE's and have them split up the CUs. Just as easy as adding more MCs or having them work at +7ghz speeds.

And you should be comparing 7970-7770 and GTX680-650ti boost.

[boxleitnerb]

Ah, so in your world the MC's are on the GDDR5 ICs...
Kepler's MCs would need a rework to guarantete +7ghz speeds.

Easy solution for FE is to add more ACE's and have them split up the CUs. Just as easy as adding more MCs or having them work at +7ghz speeds.

And you should be comparing 7970-7770 and GTX680-650ti boost.

It's not even clear if the memory controller of GK104 on the 770 was tweaked for 7 Gbps or if it was already capable of it.

And no to your comparison. It's much easier to compare the SKUs I mentioned since everything is doubled.

[LightSpeed]

Nope, just put 7 or 8 Gbps memory on the card and you're good to go. In the other case, there isn't such an easy solution.
Kepler scales perfectly with more units. 680 is 2.3x as fast as the 650 Ti (50% of everything, a tad lower clocks). GCN doesn't achieve that, the 7870 is not twice as fast as the 7770, "only" like 90% faster.

Clock speeds are pretty different, 925mhz for the 650Ti versus 1100 for the 680 at boost. Equalise them and the margins should be around 2.1x for the 680 and 1.9x for the 7870. That isnt really all that big a difference and you have to take architectural factors into account (i.e. at which size is GCN/Kepler most efficient, ROP/Bandwidth limitations etc)