AMD Zen Features Double the Per-core Number Crunching Machinery to Predecessor

[bononos]

The compiler crippling had a pronounced effect even long into the past, the result of which remains with us in terms of AMD's weakened position.

http://www.agner.org/optimize/blog/read.php?i=49
http://www.yeppp.info/home/yeppp-performance-numbers/

You misread what I posted earlier. I was saying AMD cpus still as fast or fast using Intel's compiler in crippled mode compared to other alternative compilers. That fact made it difficult for AMD to prove damage from Intel's anticompetitive compiler policy.

Heck even AMD's compiler cripples its own cpus and needs a hack to run properly.
"This issue is getting more and more absurd the more I dig into it. AMD makes a function library called AMD Core Math Library (ACML) to match Intel's Math Kernel Library (MKL). I have tested a Windows version of ACML and found that some of the functions run faster when the CPU vendor ID is artificially changed to "GenuineIntel""

AMD should come out with its optimized math library together with Zen.

 

[Phynaz]

The compiler crippling had a pronounced effect even long into the past, the result of which remains with us in terms of AMD's weakened position.

http://www.agner.org/optimize/blog/read.php?i=49

http://www.yeppp.info/home/yeppp-performance-numbers/

We haven't heard of this before, thanks for bringing it to our attention!

 

[ShintaiDK]

You misread what I posted earlier. I was saying AMD cpus still as fast or fast using Intel's compiler in crippled mode compared to other alternative compilers. That fact made it difficult for AMD to prove damage from Intel's anticompetitive compiler policy.

Heck even AMD's compiler cripples its own cpus and needs a hack to run properly.
"This issue is getting more and more absurd the more I dig into it. AMD makes a function library called AMD Core Math Library (ACML) to match Intel's Math Kernel Library (MKL). I have tested a Windows version of ACML and found that some of the functions run faster when the CPU vendor ID is artificially changed to "GenuineIntel""

AMD should come out with its optimized math library together with Zen.

Haha.

If they cant even fix their own stuff....

 

[itsmydamnation]

Haha.

If they cant even fix their own stuff....

would help if you actually read the post... but will all know where you sit on the partisan fence..........

Maybe this is not so surprising after all, since this version of ACML is compiled with Intel's Fortran compiler.

I found that drandlognormal calls a logarithm function in Intel's Short Vector Math Library (SVML). This logarithm function is dispatched into three branches for the SSE2/generic, SSE3, and the future AVX instruction set, respectively. It uses the standard Intel CPU dispatcher, which gives the generic branch to all non-Intel processors. The SVML library supports only SSE2 and above, so the generic branch uses SSE2. When my VIA processor fakes to be an Intel, it gets the SSE3 branch, which is better optimized.

There is another version of ACML for Windows built with the PGI compiler, but I couldn't make it work because some library files were missing.

Fortran

Fortran

Fortran

Fortran

 

[ShintaiDK]

And because its Fortran, AMD cant optimize it for its own CPUs? Right....

Lets be honest here, AMDs own compiler dedication is pretty much nonexistent. And this is one of the things we always see with AMD. Trying to make someone else do the job for them.

The best compiler for AMD in almost all situations funny enough tends to be the Intel compiler.

 

[TheELF]

Yeah well this happens if you're stuck in 2010...
This case was resolved back then and both companies came to terms and agreed that there was no foul play involved.
It's only relevant to software that was made around that time and did not get patched/updated since ever since its known it has been worked on.

Anyway it's completely off topic to the conversation.

So,zen has supposedly double the processing machinery.
Now it remains to see if software will be able to use all of it running a single thread or if software will have to use smt to fully use it.

Also double the stuff on even less space = not much room for high frequencies...

 

[Dresdenboy]

So,zen has supposedly double the processing machinery.
Now it remains to see if software will be able to use all of it running a single thread or if software will have to use smt to fully use it.

Also double the stuff on even less space = not much room for high frequencies...

4 symmetric ALU's, 2 FMUL and 2 FADD don't need a lot of special treatment. But HSW/SKL optimized code could run worse than generic but still optimized code.

SMT helps in 90% or more cases. To increase throughput. It's not a "fix".

And smaller area? With going 28 nm -> 14 nm and CMT -> SMT this is likely.
2 -> 4 ALUs might add ~5%. Replacing FMAC by FMUL+ FADD adds maybe also ~5%.

 

[Dresdenboy]

And because its Fortran, AMD cant optimize it for its own CPUs? Right....

Lets be honest here, AMDs own compiler dedication is pretty much nonexistent. And this is one of the things we always see with AMD. Trying to make someone else do the job for them.

The best compiler for AMD in almost all situations funny enough tends to be the Intel compiler.

I see a pattern in your behaviour.

You mean, AMD should use some of their huge lumps of money to build their own compiler team additionally to the SW work already being done? (GCC, Open64 in the past, PGI partnership, ACML, OpenCL, HSA, etc.)

 

[ShintaiDK]

I see a pattern in your behaviour.

You mean, AMD should use some of their huge lumps of money to build their own compiler team additionally to the SW work already being done? (GCC, Open64 in the past, PGI partnership, ACML, OpenCL, HSA, etc.)

That's why you get stuff like this?

Outright humiliating.

You forgot to mention AMD is good at developing something and then dump it on others to carry the software part. And then you wonder why it dont get any traction.

 

[Dresdenboy]

That's why you get stuff like this?

Outright humiliating.

You forgot to mention AMD is good at developing something and then dump it on others to carry the software part. And then you wonder why it dont get any traction.

I see I need to be clear: AMD doesn't have the time and money to catch up with Intel's well funded (SPEC is marketing relevant) compiler team. Even MS and PGI are unable to do that. That's reality.

 

[Soulkeeper]

gcc produces pretty fast code for AMD. Being a community driven project it benefits from AMD, intel, and countless other companies who contribute. There are a lot of factors that can result in one binary running faster than another, even with the same compiler. Many compilers have different default optimization levels or features that are enabled. Also a coder can target features specific to one compiler or cpu when writing their programs. Typically the coder that writes a particular program will be more familiar with optimizing for one compiler or cpu than another. I guess the point i'm trying to make is a lot of these compiler vs. compiler benchmarks can be misleading.

 

[ThatBuzzkiller]

And because its Fortran, AMD cant optimize it for its own CPUs? Right....

Lets be honest here, AMDs own compiler dedication is pretty much nonexistent. And this is one of the things we always see with AMD. Trying to make someone else do the job for them.

The best compiler for AMD in almost all situations funny enough tends to be the Intel compiler.

Doesn't matter ...

Intel is obligated to provide an agnostic path for their CPU dispatchers in their compilers as per the settlement ...

 

[TheELF]

Doesn't matter ...

Intel is obligated to provide an agnostic path for their CPU dispatchers in their compilers as per the settlement ...

Yes they do,but it does not need to be a optimized path,as long as it doesn't artificially cripple the performance.