AMD Athlon (Carrizo)
x4 845 Interrogation
You could call it a performance review, I guess...

Zen Performance Estimate
Finally! What all this work was for!

Zen - Excavator + 40%
The all-too-tempting method of estimating Zen's performance is just to take Excavator's performance numbers and blindly add 40% to them.

Doing so creates the chart on the left - at 3Ghz. This puts Zen at 0.99% above Sandy Bridge on average for these benchmarks.

However, it also puts Zen 27% above in integer. An entirely unlikely reality.

So, how do we solve this? We can't just arbitrarily modify the numbers - we have to have a reason for every change...

Excavator Module and Cache Penalties
Ah, right!! Adding 40% blindly completely negates the reality of Zen's change in relative performance for multi-threaded programs. So how do we account for that on the chart?

Simple: assume Zen will have identical scaling to K10, which gives us, perhaps erroneously, 5% more performance, bringing us to a 45% improvement over Excavator

The chart on the right shows the same data, except as relative to Sandy Bridge, which should make the fact that the performance profile is out of whack very clear.

Single Instruction Multiple Data. A very important factor for application performance, and undoubtedly included as part of the 40% figure given by AMD.

If you think for a second these charts were easy to make, you're wrong. It required me to dig into as much detail of these programs' SIMD reliance as possible - and the specific instructions at play - and Zen's theoretical ability to execute them in the same cycle.

The chart on the left has the exact same number of instructions executed as the charts above - only the distribution has changed to represent what I expect should happen in the real world.

The chart on the right represents an assumed 5% performance improvement gained from the process itself. This is actually a rather safe assumption - if it doesn't, in fact, woefully understate things. AMD's own claims were that the 40% IPC improvement were purely architectural and do not include any improvements gained from the process, so this is fair game.

And, finally...