blackkstar :
I really wish people would stop calling Windows close sourced benchmarks compiled with unknown settings "absolute performance."
In my opinion, that's simply how well the chip performs in Windows in given workloads.
I real test of gross CPU performance would be to compare chips with compilers the CPU manufacturer has optimized (or made commits to if it's FLOSS), or created, in the most ideal environment.
I'm gonna get an A4-5000 subnotebook here pretty soon, and I plan on spending as little time on Windows as possible and as much in Gentoo with btver2 enabled in GCC as possible.
I saw a more than 2x speedup in Windows compared to Gentoo in Blender. A4-5000 scores 1.5 in Cinebench r11.5. In gentoo, that'd be the same as bumping my A4 5000 score to 3. Which actually makes it better than an i5.
So, I'd really appreciate it if we'd stop using precompiled close sourced benchmarks in Windows as absolute measure of CPU performance. It's Windows performance running on unknown compilers. Even if you are aware of what compiler a benchmark is compiled with, it's still going to call DLLs and do all sorts of stuff.
Disassembling Cinebench r11.5 has been a treat. I think I find the anti-AMD code and then I find that there's more DLLs files and stuff not playing fair that needs to be looked at.
Agree 100%.
It is evident that many 'professional' reviews of Kaveri were biased. Several sites used W7 (optimized for Intel) instead W8.1 (with better CMT support) or Linux (SteamOS is coming).
HSA software was ready but ignored by most sites.
One of the review sites used an x264 binary which is optimized for Haswell chips. Then they compared Kaveri to i3/i5 Haswell using that binary. They didn't add any Sandy/Ivy chip to the comparison because Kaveri would be competitive. Of course Kaveri had difficulties to match a Haswell i3 when running that binary optimized for Haswell. To put thing in perspective, the new x264 binary that they used in their review runs up to 45% faster than the older x264 binary in a Haswell chip.
I predicted that Kaveri would be at the SB i5 level and openbenchmarking tests show Kaveri is at that level. But some few people don't accept it because open/verifiable results don't agree with closed/inverifiable results from 'professional' sites. Of course, if one underclocks the Kaveri 2400MHz memory to 1600MHz (as one site did) drops W8.1/Linux by W7 (as many did), and run binaries exclusively optimized for Intel then Kaveri will be far from Intel.
And there are other tricks often used in the 'professional' reviews. E.g. if you are measuring power consumption, select one of the more power hungry mobos for AMD and one of the more powersaving mobos for Intel. Some 'professional' review of Kaveri selected a Gigabyte F2A88XN-WIFI which consumes near 50W more than the Asus A88X-Pro with the same A10 APU under Cinebench (another Intel optimized bench). Imagine what mobo was selected for Intel: yes, a
AsRock Z87E-ITX, which is one of the more efficient mobos for Intel. The same site (it is not Anand) used different PSUs, with the Intel system receiving the more efficient PSU (subtract another 5-10% to power consumption measurement).
Don't misinterpret me. The same open benchmarks show that AMD quad-core Steamroller is still far from Haswell i5 performance in average (only in one of the test Kaveri was faster than i5-4670), but Kaveri clearly outperforms Haswell i3 and competes with SB i5.