Is it even remotely possible for us to agree on IPC increases from past generations? And for those of you who don't like the word "IPC" we mean rate of work done at equal frequencies, or throughput at equal frequencies, etc...
A couple of years ago I compiled some benchmarks, normalized clocks, and calculated some percentages. Yes, I know normalizing clocks isn't great as memory doesn't scale, and yes I know there are a million benches out there, I tried to mainly stay with Anandtech. This is what I got. Support attached. You can see Intel pretty much works on one end then the other...
Intel Generational Work Rate Comparison Results | Geomean | Average | | |
P4 to Conroe | 82.7% | 83.4% | Complete redesign | |
Conroe to Nahalem | 20.2% | 22.4% | Memory increases, add L3 | |
Nahalem to Sandy Bridge | 11.8% | 12.2% | Smarter OoO larger registers | |
Sandy Bridge to Ivy Bridge | 6.7% | 6.9% | Prefetcher improvements | |
Ivy Bridge to Haswell | 8.7% | 8.9% | add 2 execution ports | |
Haswell to Skylake | 8.9% | 9.5% | add simple decoder | |
Skylake to Sunny Cove | 21.0% | 21.3% | add 2 execution ports | |