By: Sylvain Collange (full.name.delete.delete@this.this.gmail.com), July 10, 2015 2:03 am
Room: Moderated Discussions
Patrick Chase (patrickjchase.delete@this.gmail.com) on July 9, 2015 1:08 pm wrote:
> Microarchitectures must be compared at the operating regimes for which they
> are optimized. In many cases those operating points are so far apart that useful
> comparisons are impossible, and IMO Cyclone-v-Haswell is one of those.
I think Maynard was just trying to derive the characteristics of A8's indirect branch predictor from benchmark results here, not comparing the whole microarchitectures.
As you mention, they are so many factors at play that just looking at IPC or perf/f is not conclusive. Misprediction rate would be much more relevant. Does A8 expose any kind of performance counters?
Actually, even knowing misprediction rates, two TAGE predictors with different design parameters can still give very different results (better or worse depending on the workload). The predictor design space is large. But at least it would give an idea of what kind of predictor is implemented.
> Microarchitectures must be compared at the operating regimes for which they
> are optimized. In many cases those operating points are so far apart that useful
> comparisons are impossible, and IMO Cyclone-v-Haswell is one of those.
I think Maynard was just trying to derive the characteristics of A8's indirect branch predictor from benchmark results here, not comparing the whole microarchitectures.
As you mention, they are so many factors at play that just looking at IPC or perf/f is not conclusive. Misprediction rate would be much more relevant. Does A8 expose any kind of performance counters?
Actually, even knowing misprediction rates, two TAGE predictors with different design parameters can still give very different results (better or worse depending on the workload). The predictor design space is large. But at least it would give an idea of what kind of predictor is implemented.
