What is your definition of "special"? I think we need to agree on a definition before taking this discussion further
However, this just reaffirms my belief that these benches are not very useful cross platform.
Or rather, that 3dmark physics is closest thing to a real world benchmark that's ever regularly tested. Apple's broken useless Geekbench and the js tests, but not yet 3dmark.
As opposed to Geekbench which can barely get 1 DP flop/cycle on a modern Intel core?3dmark physics????? Is the benchmark that gives the A8 and A8X SAME score?
3dmark physics????? Is the benchmark that gives the A8 and A8X SAME score?
Are you kiding me?
BTW, 95% people don't know the history of DEC P.A. branch.
That's why they don't understand that P.A.(accquired by Apple) is the most dangerous opponent to Intel.
As opposed to Geekbench which can barely get 1 DP flop/cycle on a modern Intel core?
It's a tribute to just how powerful Intel cores are when it can score so high on a program that's optimized to make Apple cores look good.
And then morphed into what it is today; to make Apple cores look good.They are not optimised for Apple cores. Geekbench started on PC and Mac, not on mobile devices. There are a few tests in there that could be better, especially the ones that does uses x87 instead of using AVX.
Gotta love the bizarre conspiracy theories.And then morphed into what it is today; to make Apple cores look good.
Are you sure? I don't see any x87 code in the 32-bit Linux while I do see SSE instructions.There are a few tests in there that could be better, especially the ones that does uses x87 instead of using AVX.
[Geekbench-3.3.2-Linux]$ objdump -d geekbench_x86_32 | grep fmul
[Geekbench-3.3.2-Linux]$ objdump -d geekbench_x86_32 | grep mul | grep '%x' | head
804e37a: f3 0f 59 05 38 22 20 mulss 0x8202238,%xmm0
804fe99: f3 0f 59 05 38 22 20 mulss 0x8202238,%xmm0
804ff10: f2 0f 59 05 c0 23 20 mulsd 0x82023c0,%xmm0
8063491: f2 0f 59 05 c0 23 20 mulsd 0x82023c0,%xmm0
80636f2: f2 0f 59 05 c0 32 20 mulsd 0x82032c0,%xmm0
80638f1: f2 0f 59 15 d8 32 20 mulsd 0x82032d8,%xmm2
8063908: f2 0f 59 c2 mulsd %xmm2,%xmm0
8063b56: f2 0f 59 05 e0 32 20 mulsd 0x82032e0,%xmm0
8063b5e: f2 0f 59 46 4c mulsd 0x4c(%esi),%xmm0
8063cc6: f2 0f 59 05 e0 32 20 mulsd 0x82032e0,%xmm0
[Geekbench-3.3.2-Linux]$ objdump -d geekbench_x86_32 | grep mul | grep -c '%x'
907NEON SIMD isn't IEEE-compliant and only supports single precision. So its use is limited. ARM 64-bit fixed that.BTW, they also do not use Neon SIMD instruction for ARM either. So neither are highly optimised for their target environment.
Are you sure? I don't see any x87 code in the 32-bit Linux while I do see SSE instructions.
Yes, I was replying to the x87 part only, not to the AVX one...You see scalar SSE instructions.
Yes, I was replying to the x87 part only, not to the AVX one...
Yes, sorry for thinking people who comment and read know what's being discussed :ninja:You were rather imprecise, since most people would think the term SSE/AVX indicate SIMD operations and most offen it does. But not here - it's scalar.
NEON SIMD isn't IEEE-compliant and only supports single precision. So its use is limited. ARM 64-bit fixed that.
Yes, sorry for thinking people who comment and read know what's being discussed :ninja:
No it isn't. As an example there's a single rounding mode.NEON SIMD single precision arithmetic is fully IEEE 754 compliant. What makes you think it is not?
And then morphed into what it is today; to make Apple cores look good.
And if you feel 3DMark physics is broken, then certainly so is Geekbench and all of the javascript benchmarks, essentially all of the same benchmarks that were broken in 2013.
SPEC 2006 requires about 10 times as long to run as SPEC 2000, might be a good throttling test too :biggrin:Let's wait for SPEC2000 score for A9, but I hope anandtech gives us SPEC06, LOL
In Metal bench, iPhone 6s is 3X as fast as iPhone 6, and 1.5X as fast as iPad Air 2.
http://thenextweb.com/insider/2015/09/25/metal-is-3x-faster-on-iphone-6s-compared-to-iphone-6/
You do have a point. The 3dmark Physic test is seriously broken, at least on Apple cores. Cannot be taken seriously when an extra core with double the cache and double the memory bandwidth does not improve the score.
You're the one talking about "graphics benchmark". The OP talks about Metal benchmark and this definitely looks like a GPU benchmark.Since when does "Particle Lab" count as a graphics benchmark ??
Some recent ARM cores have hardware prefetchers, some even can cross pages, a feature Intel only recently added. And if accesses are really random, there's nothing a prefetcher can do, even on Intel CPU. The best you can do is trying to lower latency.3dmark physics stresses the random access performance of the memory controller as well as FP performance. A lot of ARM cores are focused on sequential memory IO for graphics and so punch lower than other benchmarks would indicate.
You're the one talking about "graphics benchmark". The OP talks about Metal benchmark and this definitely looks like a GPU benchmark.