Haswell vs Kaveri

My 8350 is overclocked to 4600 and undervolted to 1.275 volts. Its rock solid stable.

Just out of curiosity, did you do any power measurements before and after OC+UV?

This sort of thing makes you wonder whether Vishera's somewhat mediocre performance might just be due to lazy binning.
 
Conservative binning seems more likely. Their worst-case assumptions involve crummier power supplies and worse cooling solutions than would be provided by people who know how to undervolt and OC their chips. They would also have more stringent error tests.
 
Aside from stuff we already knew (first HSA APU), he stated that Kaveri will have about 1TFlops.
That means a Cape Verde, or at least an 8CU GPU.

It doesn't look like AMD should worry that much about Iris Pro, unless they fail miserably in solving the bandwidth starvation issue.


Oh, and Ruby now has long hair and a prosthetic arm.
 
Last edited by a moderator:
I did read somewhere they was supposed/rumored to have GDDR5 support - in that case, no BW starvation at all.
 
900-1000 gflops for desktop. Problem will be mobile for them. Competitor for Kaveri mobile will be Broadwell in some segments.

I think 1 TF is CPU+GPU. GPU alone should be around 800-900.
 
Exactly. Means ~900 Gflops for the biggest Desktop version. We don't know exactly without having CPU Gflops results.
 
I think it's slightly bigger than that, but didn't take the time to measure it accurately. ComputerBase have a nice big image next to an Euro coin for those who might like to undertake that exercise.

You might be right, I've just repeated what someone else has estimated, so don't take it as some official figure.

Regardless it looks like this chip will be smaller than Trinity and Llano, which is good for AMD bottom line and more in line with Intel 'sweet spot' die size.

I'm really curious how this chip will clock, as this will be the first high performance (and high clock) core on GloFo bulk process from AMD (and first GloFo high performance bulk process since 0,18um?). Will it hit similar frequencies as 32nm Richland (assuming Steamroller has similar FO4 delay to it's predecessor), or will it be scaling much less than it's predecessor.

Perf/W will also be interesting to compare as I expect Kaveri to shine in this regard.
 
Last edited by a moderator:
I'm really curious how this chip will clock, as this will be first high performance (and high clock) core on TSMC bulk process from AMD.
I thought Kaveri was set to be fabbed at Globalfoundries, using their 28nm direct shrink of the 32nm process that Vishera/Trinity/Richland are using?
 
I thought Kaveri was set to be fabbed at Globalfoundries, using their 28nm direct shrink of the 32nm-SHP that Vishera/Trinity/Richland uses?

You've managed to quote me while I was corss checking known facts and corrected myself :p
 
So, based off Taylor's "teraflop-class device" comment, I'm assuming 512GCN cores at 900(ish) MHz max turbo clock. But with as big a die size as Trinity, the die size of Cape Verde and an assumed shrinkage of... what, 15-20%, could they also have stuffed a third steamroller module in there?
 
I think there was a footnote in some AMD slide indicating 8 compute units for Kaveri. That was a while ago, so I suppose the design might have changed since then, but probably not.


Yes you are correct. 512 SP for Kaveri should be safe.
 
Back
Top