Are you really comparing SSE4.1 with XOP and then claim its somehow average performance of FX?
When did I say it was the average performance of Vishera? When did mrmt say anything about the "average" performance of Zen being similar to the "average" performance of Nehalem? The word average didn't come into play.
So, of course I am comparing SSE4.1 with xOP. SSE4.1 is the absolute best-case scenario for Nehalem, while xOP/FMA4 is the absolute best-case scenario for Vishera (or Kaveri). y-cruncher is an example of a benchmark being coded the way it "should be", taking advantage of every feature available on all production silicon. Again, refer to the Haswell numbers. Haswell devastates Ivy Bridge and Sandy Bridge in that benchmark, and it absolutely ruins AMD's offerings as well, thanks in no small part to AVX2.
It isn't my fault - nor AMD's nor Intel's - when code that could/should support modern ISA extensions doesn't.
There is no more fair comparison of what the hardware can do. If you want to gimp everyone by making them run legacy x87 or SSE2 software then go right ahead. Maybe you'll have a valid point, or maybe you won't. I'll let others decide that. But, in terms of the potential performance of the hardware in question, such testing is meaningless for AMD's or Intel's processors.
It is also well-documented that various "features" of AMD's Construction cores make them suffer in unoptimized benchmarks that inflict cache flushes/pipeline stalls. Construction cores have long pipelines and slow cache, so they recover poorly from such events. Zen is moving to an Intel-like cache hierarchy and (hopefully) to a shorter pipeline. I would expect that Keller is ripping off older Core variants to hell and back to bring performance of Zen closer to Intel offerings, especially in unoptimized "real world" code. The performance delta between Intel and AMD running legacy code will probably tighten up a bit.
Zen is also going to support FMA3 and AVX2 (and is dumping xOP), so unless Keller screws the pooch on the AVX2 implementation, Zen should get a nice performance boost on AVX2 code. There shouldn't be any issue with splitting AVX instructions that you can get with Vishera et al.
Also there is a newer article from June. Confirming the start of A9 at TSMC.
I would like to read that article if you've got a link handy. Or did you link it already in this thread?
Zen wont be 22nm SOI. So its completely useless to compare with. And do you think they will continue that road? Why did IBM pay a billion to get rid of it again?
You made a statement saying that you did not understand why anyone would consider GF to be a company that specializes in high-performance nodes. Since you didn't seem to think that their 32nm or 28nm processes qualified, I pointed you towards a node that GF now controls that is indisputably high performance. No, they won't be using 22nm SOI for anything, but what they do get is a bunch of engineers that implemented and/or worked with said 22nm SOI node to add to the many engineers they have already that failed to implement 28nm SOI. Do you not see that as an improvement?
IBM sold off their fabs for many reasons:
1). To shift away from hardware and towards software/services
2). To avoid the expenses and difficulties involved with future node shrinks.
You can either interpret it as a vote of "no confidence" in their POWER lineup, or you can interpret it as a vote of "no confidence" in the silicon foundry industry as a whole. Perhaps it was some of both.
Nope, this is *exactly* what I expect from AMD 40% claim. Cherry picked cases like this one reporting sizable increases, while in overall being just slightly faster. But they, if they can put up Nehalen scores with Nehalen cores while being efficient, they will have a server product better than Vishera.
And yet you can't deny that the performance is there. Stop and think about what you are saying before you accuse me (or anyone else) of cherry-picking. Look at the Haswell scores and then compare them to Vishera. Haswell destroys it completely and utterly. A 4 GHz i7-4770k @ 4 GHz puts up a y-cruncher 500M time of 63.602s. That is
more than twice as fast as Vishera (as tested on the same page) given the same clockspeed and same thread count. Hell look at the 3630m in the same benchmark, which puts up a 154.446s. There's Ivy Bridge nearly matching Vishera at a considerably lower clockspeed (2.4-3.2 GHZ vs 4.0-4.2 GHz). If there were a 4M/8T Steamroller @ 3.4 GHz as I hypothesized above, the 3630qm would be trading blows with it, again at a lower clockspeed.
AMD shows SIMD-optimized benchmark results at its own peril.
Otherwise, I'll refer you to my reply to ShintaiDK. No sense in repeating myself.