HD6xxx launches ...Oct 12?!

[extra]

Well if what you are saying is in fact true, then NV is toast for at least 3 months.

Looks like another awesome launch by AMD to look forward to.

Well, it remains to be seen how these new parts will be priced. =p

 

[GodisanAtheist]

I thought that was already factored in, hence the 4D-is-95%-as-fast-as-5D thing. If that wasn't factored in, then I suppose it is indeed possible for Barts XT to match Cypress XT without having to be clocked at 1GHz+

I always interpreted the "easier to feed" aspect of the 4D cluster implying that its practical performance might actually outstrip the practical performance of the 5D cluster, despite its theoretical peak output dropping.

For example, say the 5D's peak performance is 1, while the 4D's peak performance is 90% of the 5D's peak or 0.9. In the real world however, the 5D arch can only extract 50% or .5 of its theoretical peak while the 4D arch can hit 75% or .75 of its theoretical peak output.

5D cluster's practical output would be 1 x .5 = .5 = 50% of 5D peak
4D cluster's practical output would be .9 x .75 = .675 = 67.5% of 5D peak

So in the given scenario, despite the 4D cluster's 10% lower theoretical peak it would actually outperform the 5D cluster by 17.5% in real world applications.

Unless I horribly failed at basic maths or my logic is horribly broken it could be possible for the Barts XT in any incarnation to turn out extremely powerful, let alone Cayman XT.

That being said, I do think the hype for the 6xxx series is getting a little out of hand and people should really par down expectations. This is a refresh on the same node, so room for improvement is a little limited.

 

[blastingcap]

Dang I had just edited my inquiry a moment before you posted this, too, after I answered my own questions: Ah never mind, I am reading at S|A about efficiency improvements all around. Seems like every little bit adds up to a nice speed bump, plus the smaller 4D shaders allow more to be packed into the same die area. One of the posters there who has historically had some degree of accuracy says that Barts XT will be in-between HD5870 and HD5850 in performance but didn't say which one it would be closer to.

Yes the setup and triangle rates are reportedly improved. Also the shaders at 1280 would be a very nice bump up from the 800 shaders of Juniper, made possible by the 4D architecture shift. Possible clockspeed bump up, too, as yields have improved on 40nm.

I don't know if there is that much hype on the HD6xxx series, it's expected to be 30-50% faster than HD5xxx which is perfectly reasonable on a somewhat larger die + various uncore efficiencies + new shaders + clockspeed bump + reclaiming some unused die space from the Cypress layout and/or removing now-unneeded redundancies.

Personally it all boils down to pricing now for me. If AMD screws up pricing, it won't matter how well their chips perform.

I always interpreted the "easier to feed" aspect of the 4D cluster implying that its practical performance might actually outstrip the practical performance of the 5D cluster, despite its theoretical peak output dropping.

For example, say the 5D's peak performance is 1, while the 4D's peak performance is 90% of the 5D's peak or 0.9. In the real world however, the 5D arch can only extract 50% or .5 of its theoretical peak while the 4D arch can hit 75% or .75 of its theoretical peak output.

5D cluster's practical output would be 1 x .5 = .5 = 50% of 5D peak
4D cluster's practical output would be .9 x .75 = .675 = 67.5% of 5D peak

So in the given scenario, despite the 4D cluster's 10% lower theoretical peak it would actually outperform the 5D cluster by 17.5% in real world applications.

Unless I horribly failed at basic maths or my logic is horribly broken it could be possible for the Barts XT in any incarnation to turn out extremely powerful, let alone Cayman XT.

That being said, I do think the hype for the 6xxx series is getting a little out of hand and people should really par down expectations. This is a refresh on the same node, so room for improvement is a little limited.

 

[Martimus]

Dang I had just edited my inquiry a moment before you posted this, too, after I answered my own questions: Ah never mind, I am reading at S|A about efficiency improvements all around. Seems like every little bit adds up to a nice speed bump, plus the smaller 4D shaders allow more to be packed into the same die area. One of the posters there who has historically had some degree of accuracy says that Barts XT will be in-between HD5870 and HD5850 in performance but didn't say which one it would be closer to.

Yes the setup and triangle rates are reportedly improved. Also the shaders at 1280 would be a very nice bump up from the 800 shaders of Juniper, made possible by the 4D architecture shift. Possible clockspeed bump up, too, as yields have improved on 40nm.

I don't know if there is that much hype on the HD6xxx series, it's expected to be 30-50% faster than HD5xxx which is perfectly reasonable on a somewhat larger die + various uncore efficiencies + new shaders + clockspeed bump + reclaiming some unused die space from the Cypress layout and/or removing now-unneeded redundancies.

Personally it all boils down to pricing now for me. If AMD screws up pricing, it won't matter how well their chips perform.

I still don't think you quite understand. Try reading this first: http://www.anandtech.com/show/2841/4

As you can see, each cluster of shaders has 4 simple Streaming Processors (SPs), and one complex SP in the current setup. The vast majority of applications never used the complex SP, so AMD decided to make the 4 simple SPs slightly more powerful (or maybe only 2 of them from recent rumors), and cut the complex SP completely. Since it was rarely used, losing it doesn't affect performance much.

So you need to divide the Cypress SPs by 5 to get the number of SP clusters (1600/5 = 320 shader clusters), and the Barts SPs by 4 to get the number of SP clusters (1280/4 = 320). As you can see, the two have the same number of clusters, and should have nearly equal performance at the same clock speed, and if all other components are the same. Like you said, however, clock speed may have increased, and the other components may have gotten better as well. This might make it actually faster than the current generation 5870, even though it has fewer "SPs". It is all because the current architecture does not efficiently use all of the SP's available.

 

[Silverforce11]

There's a rumor that the 2+2 setup no longer has all simple SPs, but all four are medium. So you end up with 1/2 DP performance as 2 SPs combined to process DP, rather than 1/5 in evergreen (~2.7 Tflops sp, ~500 Gflops dp). But if it's true, Cayman XT may have >1 Tflops DP. That would be insane.

I can't confirm this leak though.

 

[Keysplayr]

There's a rumor that the 2+2 setup no longer has all simple SPs, but all four are medium. So you end up with 1/2 DP performance as 2 SPs combined to process DP, rather than 1/5 in evergreen (~2.7 Tflops sp, ~500 Gflops dp). But if it's true, Cayman XT may have >1 Tflops DP. That would be insane.

I can't confirm this leak though.

If that were true, there would be no point in the 2+2 nomenclature. 2+2 indicates there are differences. Else it would just be 4. Unless of course the 2+2 leak was bogus.

 

[Tempered81]

If that were true, there would be no point in the 2+2 nomenclature. 2+2 indicates there are differences. Else it would just be 4. Unless of course the 2+2 leak was bogus.

that slide with the 'leaked barts info' contradicts the 4d stream processor design, unless each simd is 4 x 20 instead of 5 x 16.

IMO, either the barts slide or the 4d rumor is untrue. One or the other. but I havent read any convincing info that would tend lean either way.

 

[formulav8]

I haven't been keeping up on the rumor mill much...

Did AMD change the shader engine to have its clockspeed independent of the core clockspeed and the shader clocks are considered the uncore now? (Similar to nVidia's shaders?) I may be completely missing it...

 

[RavenSEAL]

So has anyone been able to confirm this date or what? I just RMA'd my 5670 in order to grab a 4870, but if the 6xxx series is coming out this soon, i might as well save up for another 3 weeks.