If you're not talking per-clock and you aren't specifying benchmarks or specific model numbers, then specific numbers become rather meaningless.
Even per-clock performance comparisons are highly benchmark-specific, since any SSE-heavy workloads got a huge boost from the doubling of their internal vector ALUs from 64 to 128-bits. So, a SSE-centric benchmark should easily show near 2x IPC gains. But, that wasn't all workloads.
Only a handful had to do with SSE, most couldn't care less. Go back and look at the benchmarks. The gains are pretty uniform. Actually Anandtech bench shows less gains for encoding benchmarks than it does in something like Sysmark 2014 which is general purpose.
There was a really good site back then that did all the comparisons. It came out to be 90% per clock in average over a suite of 30 something applications. Of course Pentium D clocked higher and had HT hence the advantage was less.
That's not how I remember it. Again, somewhat pointless to litigate.
It was fine if you were after productivity or something but back then games barely took advantage of dual cores. Most usages weren't that multi-threaded either.
When I look for uarch advantages I do everything to isolate it so its truly a uarch test. One site turned HT and Turbo off and the average was only 5-10% in ST. Very few gained 15%+ probably due to the increased memory bandwidth. Realworldtech guys did testing and even found regressions for ST workloads.
The surprise for Nehalem in client was that the integrated memory controller brought small gains. For Athlon 64 the gains were massive, something like over 2/3rds of the 30% perf/clock gains were due to the IMC.
It's probably because the Core uarch hid memory latencies so well that there wasn't much to gain by moving to the IMC. Indeed the memory latency tests showed just that - noticeable, but not big gains.
That was the disappointment. A lot of us took the figure from Athlon 64 and thought immediately we'd get 20-30% gains in ST. No, we got a fraction of that. We COULD get that but only with with disclaimers and fine prints.
In servers the workload is much larger so its harder to fool and hits the memory subsystem, hence it was fantastic.
Conroe was for client, Nehalem was for servers.
Another fantastic client chip was Sandy Bridge. 15-20% faster per clock(uarch), lowered power, increased clocks, Turbo improved vastly(especially in laptops) and overclocked super well. The much improved graphics making it usable was an icing on the cake. And while Nehalem regressed on the battery life side, Sandy Bridge more than made up, and even made high end laptops look efficient.
Facebook
Twitter
Instagram