Haswell 4c8t @2.6 Fritz Chess Benchmark <Update:2.8G Haswell>

[cytg111]

Put me under "I disagree". Just how AVX2 only benefits certain types of computation benchmarks, there are other architectural features that can skyrocket other types of traces. And it can be done without additional instructions or additional ALUs. On top of that, remember instruction level is just one level of abstraction. Consider the uop flows for complex instructions that can be improved. Same instruction... just faster.

I have never run SPECfp but I do know it composes of many FP intensive traces and I guess the final score is a big average. Has anyone attempted to see the speedup of the individual traces? You might be surprised.

Edit But in case anyone wanted to nitpick. Yes, SOME traces need new instructions to improve. When a certain trace is well understood and therefore very optimized for the CPU to minimize misprediction, stalls, etc.... then you start hitting the point that IPC improvements is going to require some creativity.

That might be, but evidence seems to suggest that IPC is scaling off for both Intel and AMD, converging at about the same level (yes, Intel is getting there sooner, rather than AMD-later).

 

[TuxDave]

That might be, but evidence seems to suggest that IPC is scaling off for both Intel and AMD, converging at about the same level (yes, Intel is getting there sooner, rather than AMD-later).

I see it as broad ipc improvements getting harder but application specific improvements still making large strides.

 

[Idontcare]

I see it as broad ipc improvements getting harder but application specific improvements still making large strides.

Evolving away from general purpose hardware towards fixed/dedicated hardware?

 

[TuxDave]

Evolving away from general purpose hardware towards fixed/dedicated hardware?

I wasn't really thinking that but that's true too. Take Ivybridge's FP divider improvements. Not a lot of programs care about FP divides performance but certain financial benchmarks love it. So I guess in a way, you can argue even an FP divide unit as a fixed/dedicated hardware and not really general purpose. What if we just delete the whole darn thing and save some area. Heh.

So finding a single feature that affects a very broad number of applications is pretty tough. It's definitely one of those things where you have to probably put in 4 features that target 4 different areas to finally see a fake "broad IPC" improvement. But if you start narrowing your view to a customer specific application, some lucky customers will see dramatic improvements.

At least that's my simplistic view of things.