A direct comparison of three generations of GCN

[CAT-THE-FIFTH]

Not seen this done before,so it is useful to see the performance uplift from each generation:

https://translate.google.com/transl...n-polaris-architektur-performance/&edit-text=

 

[D.P.]

The problem with these comparisons is it completely ignores what clocks can be achieved. IPC is a relatively pointless metric for most purposes. Setting GPUs as fixed clock speeds they were not designed to run at causes all sorts of unintended artifacts in testing, e.g. increased missed cache costs, memory interface timing issues.


Really, fi you want to compare the performance compare them as they were sold or overclocked to their limit.

 

[CAT-THE-FIFTH]

The mobile Polaris parts probably have similar or slightly lower base clockspeeds,so its still an interesting comparision anyway IMHO.

 

[pandem0nium]

Yeah, good comparison. Would be good to see comparisons at stock and max overclock as well.

I think lowering the memory frequency on the tahiti to match bandwidths makes it less fair since the later cards have texture compression implemented to offset the smaller bus.

 

[andybird123]

The problem with these comparisons is it completely ignores what clocks can be achieved. IPC is a relatively pointless metric for most purposes. Setting GPUs as fixed clock speeds they were not designed to run at causes all sorts of unintended artifacts in testing, e.g. increased missed cache costs, memory interface timing issues.


Really, fi you want to compare the performance compare them as they were sold or overclocked to their limit.

Thats only a problem if the person doesnt understand what they are reading. Of course its interesting to compare IPC as a thing in its own right to compare the approach taken in regards potentially giving up IPC to achieve higher clocks.

For most buyers they are interested in how a particular card will run X game, but for people interested in technology a simple 3 page article like this is a fine read.

If someone gets confused that easily then maybe they shouldnt be reading tech articles.

 

[D.P.]

Thats only a problem if the person doesnt understand what they are reading. Of course its interesting to compare IPC as a thing in its own right to compare the approach taken in regards potentially giving up IPC to achieve higher clocks.

For most buyers they are interested in how a particular card will run X game, but for people interested in technology a simple 3 page article like this is a fine read.

If someone gets confused that easily then maybe they shouldnt be reading tech articles.

The IPC is just irrelevant though. It would be much more interesting to look at IPS normalized by power consumption for example.