AMD Radeon 7000-Series 28nm (Southern Islands) | 7990 7970 7870 7770 | Discussion

[Zargon]

ah man I am probably going to save and get dual 7950's if they are >30% increase over 6950's and similar power use

 

[roedtogsvart]

how do you get 64 and 60 ROPs with 384 bit?

Wondering this as well.

 

[bunnyfubbles]

What's wrong with these prices? If correct, AMD is clearly pricing it higher than what the 6970 launched at (and very close to what the GTX580 sells at). It seems to me, that if anything, they expect this to be the fastest card for a while, and probably pretty competitive with Nvidia's fastest single GPU.

Maybe I'm missing what you are saying, then. I don't think this is overly aggressive pricing. I believe AMD is pricing these right in line with what the expected performance will be. They have to price them at what the market will allow... selling these for $250 and making a penny of profit won't help them, even if they sell 10 million a quarter. Likewise, pricing them at $10k and selling 500 a quarter won't do much for them. I think the 7970 will be much faster than the GTX580, and be priced pretty close to what we have seen top single GPU parts sell for lately. I'm sure AMD expects price cuts from Nvidia if the GTX5xx cards are what Nvidia has for the immediate future, or they expect Nvidia to have their next gen out in the not too distant future.

At any rate, the 7950 may be my next upgrade... I am looking to get my power use down to respectable levels from two 5870's, and at 1920x1200 1GB is going to be a limiting factor soon.

You are missing what he's saying, he's saying these prices are too good, many were expecting $399 and $499 for the 7950 and 7970 respectively, not $50 cheaper. His point is that if they're priced this "low", then maybe performance won't meet the common high expectations of anywhere from at least 50% faster to nearly twice as fast.

 

[JBT]

That 7950 is looking very enticing.

 

[Stoneburner]

Finally more texture units and ROP's. Maybe this will be far better balanced than the last two ATI generations.

 

[skipsneeky2]

Hope this is a true release date which i doubt.

Gotta wonder about their power consumption.

Might be out of the running with my 600 watt psu with even a 7950.so mind as well stick to the gtx570 i planned today.

Betting most places will raise their prices of a 7950 into the $400 range demand should be high .

 

[ArchAngel777]

Why am I doubting this? Don't get me wrong, I will be thrilled if true... But...

 

[badb0y]

My body is ready.

 

[toyota]

Finally more texture units and ROP's. Maybe this will be far better balanced than the last two ATI generations.

I am still not buying the 64 and 60 ROP claims. that makes no sense as I do not see how you can end up with that with six 64bit memory controllers. it would suely have to be something like 24, 48, or 96 and so on with a 384 bit bus. and how would you just disable 4 of those for the 7950 since it would still have the same number of memory controllers? I am no expert here but that just sounds odd.

 

[badb0y]

I am still not buying the 64 and 60 ROP claims. that makes no sense as I do not see how you can end up with that with six 64bit memory controllers. it would suely have to be something like 24, 48, or 96 and so on with a 384 bit bus. and how would you just disable 4 of those for the 7950 since it would still have the same number of memory controllers? I am no expert here but that just sounds odd.

Nice catch your right it doesn't add up.

 

[GrumpyMan]

What a coincidence they are coming out just in time for tax refund season.

 

[Vesku]

Are these confirmed as GCN or is this VLIW4+?

 

[Regk]

Hopefully these cards will be sub 10.5" so I can actually fit one in my CM Strom scout case!

 

[Concillian]

Interesting, but we'll see.

 

[Vesku]

Going off this pricing I'd think it would be 7850 at $200 and 7870 at $250.

 

[Grooveriding]

Going off this pricing I'd think it would be 7850 at $200 and 7870 at $250.

I'm wondering what sort of performance gap there is going to be between 7850/70 and 7950/70. With the mid-range being VLIW4 and the high-end being GCN, there may be a large performance disparity between the two.

7850/70 could wind up being a 6970 on 28nm with higher clocks and a few tweaks. Still would be a good card at $250.

 

[Zanovar]

prices if true seem reasonable to me,but ive seen the rumoured expected prices somewhere else(cant remember where)and both cards where coming in at alot higher prices.

 

[Silverforce11]

On specs alone, Tahiti XT is a good 50% (more sp, texture, rop, bandwidth and faster core speed) faster than Cayman XT, not factoring in architecture changes and improvements. So i see that as worst case performance.

But wait for real partner briefing leaks.

 

[gnuelxela]

thought they the 79XX should be using XDR2 instead of GDDR5?

 

[RussianSensation]

Assuming the various shaders/processors perform similarly to older units, about how fast might this be?

Ripped this form HardForums:

7970XT against 6970 would be:
33% More Shaders
33% More Texture Units
13,6% Higher Clock Speed
65% Higher Memory Bandwidth
51% more Gflops

That looks pretty decent.

264 GB/sec vs. 176 GB/s is a 50% increase in memory bandwidth. To arrive at 65%, the poster probably used the incorrect 160 GB/sec memory bandwidth of the 6950.

It's a little difficult to say how much faster these cards will be since we are now comparing 2 entirely different architectures: GCN vs. VLIW-4. Since this is an entirely new architecture, SPs, TMUs and ROPs may have been significantly improved. Outside of ROPs, everything seems to have improved by about 50%. If GCN adds another 15-20% improvement over VLIW-4 architecture, a single HD7970 might be within 10% of the HD6990

Still, unlike say 8800GTX --> GTX280 or HD4870 --> HD5870 where the architecture remained the same, and we could simply compare pixel, texture fillrates and memory bandwidth, comparing cards on paper vs. 2 different architectures such as going say GTX280 to GTX480 can often lead to erroneous results.

For example, on paper, GTX280 should smash a GTX460 1GB, but it doesn't.

Looking forward to seeing benches. Prices are better than I expected. Without competition from Kepler, I thought AMD might price 7970 at $499, at least.

 

[toyota]

264 GB/sec vs. 176 GB/s is a 50% increase in memory bandwidth. To arrive at 65%, the poster probably used the incorrect 160 GB/sec memory bandwidth of the 6950.

It's a little difficult to say how much faster these cards will be since we are now comparing 2 entirely different architectures: GCN vs. VLIW-4. Since this is an entirely new architecture, SPs, TMUs and ROPs may have been significantly improved. Outside of ROPs, everything seems to have improved by about 50%. If GCN adds another 15-20% improvement over VLIW-4 architecture, a single HD7970 might be within 10% of the HD6990

Still, unlike say 8800GTX --> GTX280 or HD4870 --> HD5870 where the architecture remained the same, and we could simply compare pixel, texture fillrates and memory bandwidth, comparing cards on paper vs. 2 different architectures such as going say GTX280 to GTX480 can often lead to erroneous results.

For example, on paper, GTX280 should smash a GTX460 1GB, but it doesn't.

Looking forward to seeing benches. Prices are better than I expected. Without competition from Kepler, I thought AMD might price 7970 at $499, at least.

you said that in another thread too but you seem to never acknowledge that the gtx460 has 336sp compared to 240sp of the gtx280.

 

[Stuka87]

thought they the 79XX should be using XDR2 instead of GDDR5?

That was rumored for a while. But then I think it was somebody from AMD that said this would not be the case.

 

[RussianSensation]

you said that in another thread too but you seem to never acknowledge that the gtx460 has 336sp compared to 240sp of the gtx280.

Of course I acknowledge that GTX460 has more SPs. However, I think you are placing way too much emphasis on just 1 aspect of a videocard. SP = shader/stream processors. What about the efficiency of ROPs, or TMUs, overabundance of memory bandwidth? For instance, the ROPs in HD6900 series are improved over HD5800 series. But on paper it's still the same 32 ROPs.

Having more specs doesn't necessarily allow the card to utilize them effectively if it's bottlenecked in another area. For instance, HD5870 has a bucketload of texture fill-rate performance vs. GTX480, but isn't faster. You can't compare SPs across new architectures. You can try but it will tell you almost nothing.

I used GTX280 vs. 460 because it's the perfect example of why comparing 2 entirely different architectures on paper is meaningless.

In the case of GTX280, you have higher memory bandwidth and higher texture fill-rate than GTX460 and almost the same pixel fill-rate, but the 200b chip still loses.

GTX 280 vs. 460
Pixel Fillrate (a wash) = 19264 MPixels/sec vs. 21600 MPixels/sec
Texture Fillrate (GTX280 wins) = 48160 MTexels/sec vs. 37800 MTexels/sec
Memory Bandwidth (GTX 280 wins) = 141.696 GB/sec vs. 115.2 GB/sec

On the other hand, GTX280 is far faster in older OpenGL titles than GTX460 is. But we'd never be able to guess that based on paper specs.

Based on this, GTX280 should destroy the 460. Looking at paper specs across 2 different GPU architectures is meaningless. Just look at HD6870 with 1120 SPs vs. HD5870 with 1600 SPs. You can't just compare SPs (clock speed is important too). Or look at HD5770 with nearly half the memory bandwidth of HD4870. You can't just compare memory bandwidth either. What about HD5830? It should be way faster on paper than HD4890, but it isn't, because it's ROP limited. HD4890 has pretty much the same memory bandwidth as an HD6870, but it's far slower. The extra memory bandwidth is just wasted on the 4870/4890 cards, while the 5870 has way too much texture and SP performance - an unbalanced design.

AMD is adopting a smart compute approach. Graphic Core Next is a true MIMD (Multiple-Instruction, Multiple Data) architecture. With the new design, the company opted for "fat and rich" processing cores that occupy more die space, but can handle more data. So comparing 2048 SPs of GCN architecture to VLIW-4 is probably meaningless.

In general, comparing cards on paper can work if you are comparing across the same architecture/generation, but it's almost always meaningless across 2 entirely different architectures imo. What if GCN is 20-30% more efficient than VLIW-4? What about vastly improved geometry/tessellation performance? NV cleans up in Lost Planet 2, Hawx 2, Crysis 2, partly due to superior Tessellation performance. Also, what if AMD introduces a multi-threaded DX11 driver for HD7900 series allowing it to surprass NV in games like CIV5?

All I am saying is these specs appear to be "only" about 50% faster than HD6970 (outside of ROPs), but the performance improvement might be far greater since the specs don't tell us anything about how much better (or worse) GCN is vs. VLIW-4.

 

[bunnyfubbles]

What if GCN is 20-30% more efficient than VLIW-4?

20-30% in improved efficiency on top of 33-100% increase in core components on top of slightly higher clocks = instant nerdgasm if true :awe:

 

[Concillian]

I'm wondering what sort of performance gap there is going to be between 7850/70 and 7950/70. With the mid-range being VLIW4 and the high-end being GCN, there may be a large performance disparity between the two.

7850/70 could wind up being a 6970 on 28nm with higher clocks and a few tweaks. Still would be a good card at $250.

Based on the shader counts, I'm guessing any performance differences on GCN are primarily going to be in tesselation and GPGPU type loads and that pure graphics will be pretty similar to VLIW4/5.

I don't see them dropping 6970 performance to $250 until they have to. Everyone expects the 7870 to be essentially a die-shrunk 6970, and AMD is past price wars. It's trying to squeeze out profits now, and I think they'll be slower to drop prices at a given performance. I see the 7850 / 7870 coming out at ~$250 / 300 or so (give or take, basically only slightly cheaper than the 6950 / 6970 are now) and then dropping prices slowly.