ATI Radeon HD 4000 Series discussion

[HKPolice]

RV770 is both impressive and disapointing at the same time.

RV670 Specs:
16 TMUs
320 SP arranged in 4 SIMD arrays of 80sp each
16 ROPs
256Bit DDR3/4 Ram
666Million Transistors on 55nm

RV770 Specs:
32 TMUs
800 SPs arranged in 5 SIMD arrays of 160sp each
16 ROPs
256Bit DDR4/5 Ram
~830Million transistors on 55nm

Now the impressive part is that ATi managed to more than double the number of SPs and double the amount of TMUs and only add about 200Mill transistors. The disapointing part is that overall performance is only expected to be 50% higher than RV670. One would expect more than that especially since RV770 will finally offload AA to the ROPs.

There is no "true next gen" chip for the DX10 generation, even the GT200 is just a revamped G92 with more shaders/ROPs and TMUs. Things in the GPU world will be boring until DX11 comes.

[AliG]

RV770 is both impressive and disapointing at the same time.

RV670 Specs:
16 TMUs
320 SP arranged in 4 SIMD arrays of 80sp each
16 ROPs
256Bit DDR3/4 Ram
666Million Transistors on 55nm

RV770 Specs:
32 TMUs
800 SPs arranged in 5 SIMD arrays of 160sp each
16 ROPs
256Bit DDR4/5 Ram
~830Million transistors on 55nm

Now the impressive part is that ATi managed to more than double the number of SPs and double the amount of TMUs and only add about 200Mill transistors. The disapointing part is that overall performance is only expected to be 50% higher than RV670. One would expect more than that especially since RV770 will finally offload AA to the ROPs.

There is no "true next gen" chip for the DX10 generation, even the GT200 is just a revamped G92 with more shaders/ROPs and TMUs. Things in the GPU world will be boring until DX11 comes.

Thing is I don't think ati is offloading aa to the ROPs, that's probably the reason for the high shader increase and no ROP increase, and that could very well mean that there's a need for all those shaders. However I don't think that transistor count is right, there's no way that's possible to more than double the shader count with only adding 200 some transistors

[NH|Delph1]

RV770 is both impressive and disapointing at the same time.

RV670 Specs:
16 TMUs
320 SP arranged in 4 SIMD arrays of 80sp each
16 ROPs
256Bit DDR3/4 Ram
666Million Transistors on 55nm

RV770 Specs:
32 TMUs
800 SPs arranged in 5 SIMD arrays of 160sp each
16 ROPs
256Bit DDR4/5 Ram
~830Million transistors on 55nm

Now the impressive part is that ATi managed to more than double the number of SPs and double the amount of TMUs and only add about 200Mill transistors. The disapointing part is that overall performance is only expected to be 50% higher than RV670. One would expect more than that especially since RV770 will finally offload AA to the ROPs.

There is no "true next gen" chip for the DX10 generation, even the GT200 is just a revamped G92 with more shaders/ROPs and TMUs. Things in the GPU world will be boring until DX11 comes.

The first story about 50% faster was posted by Fudzilla, I think he beat me with a few hours on that one. I'm not sure what his source told him, but mine was very vague and after hearing some more I was thinking that the "50% faster" was based on the fact that RV770 was believed to have 50% more SPs, I.e. 480 all in all. I can't say that for certain though, it's just a speculation on my behalf.

I've been hearing the figure 480 over and over again, but the context has been a bit different every time. Right now it seems that we're looking at a total of 800 SPs, which means 480 added SPs. If I'm right about my first speculation, we can throw the early stories of 50% faster out the window.

But there's one more thing to it. (I wrote this is in the original "50% faster" article) If you go back and check how much faster each [true] generation than the last one has been, you won't see much higher figures than 50%. With each generation that has passed, the last generation high-end has performed like the new generation mid-range. But with the lack of a genuine high-end chip from AMD/ATI, I can surely understand that most people are terrified by this.

And then we have the fact that AMD/ATI has come up with a way for two chips to share a memory buffer, which at least makes me really eager to see how the dual-GPU R700 will perform, not to mention two cards with four GPUs and just two memory buffers.

I'm very skeptical toward whether AMD/ATI will ever make a new super GPU. It just makes more sense to spread the workload over several GPUs instead, to them. And considering how well the concept of R680 worked out, I can only imagine that AMD/ATI will keep heading down that road. The fact that the RV670 was a poor performer in AA/AF situations shouldn't cloud your opinion of the R680 concept. Just plug it in and it works. No hassle.

//Andreas

[flopper]

And then we have the fact that AMD/ATI has come up with a way for two chips to share a memory buffer, which at least makes me really eager to see how the dual-GPU R700 will perform, not to mention two cards with four GPUs and just two memory buffers.

I'm very skeptical toward whether AMD/ATI will ever make a new super GPU. It just makes more sense to spread the workload over several GPUs instead, to them.
Just plug it in and it works. No hassle.

//Andreas

what is the advantage of that solution with memory buffers?

and the cards with ati/amd seems to be a better driver support due to the program they run for a while.

however, rumours aside, we want facts

[NH|Delph1]

what is the advantage of that solution with memory buffers?

and the cards with ati/amd seems to be a better driver support due to the program they run for a while.

however, rumours aside, we want facts

Well, in the case of CrossFire, it would double the amount of memory.

Today the data is mirrored in each buffer, which means that the 3870X2 1GB, really only uses 512MB, and two 3870X2 cards have the same data mirrored four times.

With two GPUs sharing one 1GB buffer, the entire 1GB would be used in CrossFire. Link two of these cards and they would still have just 1GB, even though the buffer is 1+1GB. But we've doubled the buffer, or halved the loss of effective memory. The data would only be mirrored once, instead of four times.

That's theory and rumor though

//Andreas

[RedBull78]

Well, in the case of CrossFire, it would double the amount of memory.

Today the data is mirrored in each buffer, which means that the 3870X2 1GB, really only uses 512MB, and two 3870X2 cards have the same data mirrored four times.

With two GPUs sharing one 1GB buffer, the entire 1GB would be used in Crossfire. Link two of these cards and they would still have just 1GB, even though the buffer is 1+1GB. But we've doubled the buffer, or halved the loss of effective memory. The data would only be mirrored once, instead of four times.

That's theory and rumor though

//Andreas

Nice. That's like eliminating the middle man. I just hope you're right. No mirrors at all and just one main memory buffer. Huge improvement in latency as well and no more slowdowns in multi-GPU solutions that was talked about a few months ago...

[Unbornchild]

No more "mini-stuttering" (i suppose) you mean. Which would be awesome.