AMD vs Intel Processors In Linux

[mavere]

Weird how GCC gives the best or nearly so performance on all of those tests. Some instances by a wide margin. Called it.

"Lower is better"

 

[TerryMathews]

"Lower is better"

Huh. That first graph, that's a weird way to represent that normalized against ICC.

 

[jhu]

"Lower is better"

For the first graph, higher is better. I'm not quite sure about the second graph.

 

[TerryMathews]

For the first graph, higher is better. I'm not quite sure about the second graph.

That's what I thought...

 

[ShintaiDK]

OMG. This really is a devious benchmarketing lie....

It compares partly multi-threaded code against single threaded code
So what a wonder, multiple AMD cores are faster than a single AMD core..

The Open64 compiler generates on average 50% to 60% faster executables
as Intel's ICC 12 on the same code: Yes that really looks crippled!

Keep forgetting this?

Maybe we should see what you actually compare with:
http://www.spec.org/cpu2006/results/res2011q4/cpu2006-20111206-19147.html
http://www.spec.org/cpu2006/results/res2011q4/cpu2006-20111121-18939.html

FX8150 vs dual Opteron 4284.
8GB vs 32GB.
One running Windows 7, the other Red Hat 6.1.
Base pointers 32/64 vs 64.

You should learn to check your source material Hans.

 

[mavere]

For the first graph, higher is better. I'm not quite sure about the second graph.

I'm not sure why this is even a disagreement, but just to make sure everyone's on the same page, I directly linked to the source comparison. (Here it is again)

To save most the trouble, this is how NERSC (who is invested in maximizing their Opterons' efficiency) interprets their own results:

In general, the Cray and the Intel compilers outperform the others, so if performance is your prime consideration, it would repay you to test these compilers on your application with the recommended optimization options. The Gnu and PGI compilers generally produce code 5-10% slower, but on some benchmarks one or the other outperform all other compilers.

 

[MiddleOfTheRoad]

Those are some interesting numbers. As for the FX versus the 3770K -- there are actually several tasks (under Linux) where the i7 3770k trails an FX by a significant margin. Most things multi-threaded are usually neck and neck. An i7 is always faster single threaded. I've personally found that AMD FX CPU's don't perform well unless you're running DDR3 1866. I suspect many of the negative reviews of FX were running very slow DDR3 (in addition to crippling the FX by running Windows 7).

This is a pretty good comparison under Linux (where I got the below benchmark):
http://www.phoronix.com/scan.php?page=article&item=amd_fx8350_visherabdver2&num=1

For years, I've been scratching my head about how "inferior" the FX was compared to
Intel i5/i7. I really never saw a noticeable gap -- But I also spend about 80% of my time
in Linux. The only drawback IMO to an AMD CPU is really just power consumption....
However since the majority of their chips are usually unlocked, even that can be tuned
or undervolted.

 

[jhu]

Those are some interesting numbers. As for the FX versus the 3770K -- there are actually several tasks (under Linux) where the i7 3770k trails an FX by a significant margin. Most things multi-threaded are usually neck and neck. An i7 is always faster single threaded. I've personally found that AMD FX CPU's don't perform well unless you're running DDR3 1866. I suspect many of the negative reviews of FX were running very slow DDR3 (in addition to crippling the FX by running Windows 7).

This is a pretty good comparison under Linux (where I got the below benchmark):
http://www.phoronix.com/scan.php?page=article&item=amd_fx8350_visherabdver2&num=1

For years, I've been scratching my head about how "inferior" the FX was compared to
Intel i5/i7. I really never saw a noticeable gap -- But I also spend about 80% of my time
in Linux. The only drawback IMO to an AMD CPU is really just power consumption....
However since the majority of their chips are usually unlocked, even that can be tuned
or undervolted.

That's because most people here really just care about games, and games tend not to be as multithreaded.

 

[MiddleOfTheRoad]

That's because most people here really just care about games, and games tend not to be as multithreaded.

That is so very true.

It does make me shake my head when people on this forum state that an FX can't even keep with an i3 for pretty much everything. When the reality is far different for people that actually use both. There are a lot of people that use their PC's for more than just playing games.... (Although I do jump on Steam all the time).

 

[Enigmoid]

Those are some interesting numbers. As for the FX versus the 3770K -- there are actually several tasks (under Linux) where the i7 3770k trails an FX by a significant margin. Most things multi-threaded are usually neck and neck. An i7 is always faster single threaded. I've personally found that AMD FX CPU's don't perform well unless you're running DDR3 1866. I suspect many of the negative reviews of FX were running very slow DDR3 (in addition to crippling the FX by running Windows 7).

This is a pretty good comparison under Linux (where I got the below benchmark):
http://www.phoronix.com/scan.php?page=article&item=amd_fx8350_visherabdver2&num=1

For years, I've been scratching my head about how "inferior" the FX was compared to
Intel i5/i7. I really never saw a noticeable gap -- But I also spend about 80% of my time
in Linux. The only drawback IMO to an AMD CPU is really just power consumption....
However since the majority of their chips are usually unlocked, even that can be tuned
or undervolted.

Yep, those tests generally occur when HT adds next to nothing.

http://openbenchmarking.org/index.php?c=result&i=1210149-SU-HYPERTHRE70

 

[MiddleOfTheRoad]

Funny... I have a 3770K and used to have an 8320 (gave it away and threw in a 4350) and in real world use (not benchmarking) the 8320 felt sluggish compared to the 3770K and I'm also a Linux user. Clock for clock, Intel's going to be ahead of AMD in a lot of ways.

Depends on the task. An FX 8350 is faster than a 3770K in 7 linux benchmarks on Phoronix.

 

[Zodiark1593]

Has anyone compared Blender performance (both Cycles and Internal) on both processors under Linux? I'm rather curious to see that as it should be FP heavy.

 

[jhu]

Has anyone compared Blender performance (both Cycles and Internal) on both processors under Linux? I'm rather curious to see that as it should be FP heavy.

There's always this. The take away message is: if you're using Blender, you might as well get the best NVidia card you can buy for rendering instead of messing around with CPU rendering. The massive decrease in rendering time is insane!

Anyway, based on the .blend file (cycles rendering) in the that thread here are my results:

FreeBSD 10, Blender 2.71, compiled with llvm 3.3
FX 8350 (stock, 4.1 GHz turbo): 2 minutes 4 seconds
1672258 samples/s, 407867 samples/s/GHz, 418064 samples/s/module, 101966 samples/s/module/GHz

Ubuntu 14.04, Blender 2.71, official Blender binary
Core i5 3317U (stock, 2.4 GHz turbo): 5 minutes, 20 seconds
648000 samples/s, 270000 samples/s/GHz, 324000 samples/s/core, 135000 samples/s/core/GHz

So Ivy Bridge has a higher IPC than Piledriver in Blender. But they all get blown away by NVidia.

 

[Zodiark1593]

There's always this. The take away message is: if you're using Blender, you might as well get the best NVidia card you can buy for rendering instead of messing around with CPU rendering. The massive decrease in rendering time is insane!

Anyway, based on the .blend file (cycles rendering) in the that thread here are my results:

FreeBSD 10, Blender 2.71, compiled with llvm 3.3
FX 8350 (stock, 4.1 GHz turbo): 2 minutes 4 seconds
1672258 samples/s, 407867 samples/s/GHz, 418064 samples/s/module, 101966 samples/s/module/GHz

Ubuntu 14.04, Blender 2.71, official Blender binary
Core i5 3317U (stock, 2.4 GHz turbo): 5 minutes, 20 seconds
648000 samples/s, 270000 samples/s/GHz, 324000 samples/s/core, 135000 samples/s/core/GHz

So Ivy Bridge has a higher IPC than Piledriver in Blender. But they all get blown away by NVidia.

There are certain use cases which bar the use of CUDA in Cycles rendering, custom shaders (OSL), memory limit, etc. While I'm not denying the superior speed of gpu rendering, it's a mistake to overlook the cpu as well.

 

[jhu]

Some more CPU results:

Core i5 3317U (stock, 2.4 GHz turbo)
Ubuntu 14.04, Blender 2.71, official Blender binary
Time: 5 minutes, 20 seconds
648000 samples/s, 270000 samples/s/GHz, 324000 samples/s/core, 135000 samples/s/core/GHz

Core i5 2400S, 2.5 GHz
Ubuntu 14.04, Blender 2.71, official Blender binary
Time: 3 minutes 15.67 seconds
1059743 samples/s, 423897 samples/s/GHz, 264936 samples/s/core, 105974 samples/s/core/GHz

FX 8350 (stock, 4.1 GHz turbo)
FreeBSD 10, Blender 2.71, compiled with llvm 3.3
Time: 2 minutes 4 seconds
1672258 samples/s, 407867 samples/s/GHz, 418064 samples/s/module, 101966 samples/s/module/GHz

Core 2 Duo, 2.2 GHz
Maco OS X 10.6.8, Blender 2.71, official Blender binary
Time: 10 minutes 38.54 seconds
324741 samples/s, 147609 samples/s/GHz, 162370 samples/s/core, 73805 samples/s/core/GHz

Phenom II x6 (stock, 3.2 GHz)
Ubuntu 14.04, Blender 2.71, official Blender binary
Time: 2 minutes 32.26 seconds
1361881 samples/s, 425587 samples/s/GHz, 226980 samples/s/core, 70931 samples/s/core/GHz

Not looking too good for Piledriver although it's possible that the llvm binary is slower than the official binary.

 

[Azuma Hazuki]

I'm mostly interested in this for the purpose of Gentoo compiles, and the integer/FP comparisons with the Ivybridge and Gescher series were really, really surprising.

I always was told Orochi and its derivatives had very weak FP but fairly competent integer performance, as they are server chips, and that compilation is almost entirely integer. Yet, earlier in the thread someone showed a larger gap between the two for integer than FP.

Also, regarding compilation, is all compiling integer tasks, even in code that deals with floating-point math? I don't understand...

 

[DrMrLordX]

Some more CPU results:

Not looking too good for Piledriver although it's possible that the llvm binary is slower than the official binary.

Looks great for Thuban, though. Something like Magny-Cours would have been amazing for Blender, no?

 

[jhu]

Looks great for Thuban, though. Something like Magny-Cours would have been amazing for Blender, no?

Compared with even Sandy Bridge, not really. Maybe back when Sandy Bridge didn't exist perhaps.

But GPU rendering is even faster. Dual Nvidia GTX 580 renders the benchmark in 16s! That's 12,960,000 samples/s and nearly 10x the performance of the CPUs I've tested.

 

[MiddleOfTheRoad]

Compared with even Sandy Bridge, not really. Maybe back when Sandy Bridge didn't exist perhaps.

But GPU rendering is even faster. Dual Nvidia GTX 580 renders the benchmark in 16s! That's 12,960,000 samples/s and nearly 10x the performance of the CPUs I've tested.

That's insanely impressive -- I knew Nvidia GPU's were awesome under Linux, but even this goes beyond what I thought was possible.

 

[Abwx]

That's 12,960,000 samples/s and nearly 10x the performance of the CPUs I've tested.

At 5x the TDP also, with 5 8350 you would get 25s and about the same TDP for the processing parts, so the GPU scores are not that impressive considering that the CPUs are general purpose calculators.

 

[Enigmoid]

Depends,

This is not under linux.

 

[Cerb]

I'm mostly interested in this for the purpose of Gentoo compiles, and the integer/FP comparisons with the Ivybridge and Gescher series were really, really surprising.

I always was told Orochi and its derivatives had very weak FP but fairly competent integer performance, as they are server chips, and that compilation is almost entirely integer. Yet, earlier in the thread someone showed a larger gap between the two for integer than FP.

Also, regarding compilation, is all compiling integer tasks, even in code that deals with floating-point math? I don't understand...

There's scalar and there's vector.
Then, there's integer and FP math.
Then, there's memory operations, including non-instruction work like branch prediction and prefetching.

Lots of uses for floating point work well for SIMD vector processing. That code will show very high FP ops/sec v. integer, because it's counting every op on every lane. Scalar FP tends to be far slower, and is a major weakness of x86 (it's also not used too much, and has been a focus of Itanium, Fujitsu's SPARC line, and various IBM POWER CPUs, for sciency folks with lots of R&D money to burn ).

Compilation is going to be mostly address/memory work, with some integer math, and negligible FP. It's going to be almost entirely scalar, and low-IPC, so even the FP that may get used is not going to affect performance. Whether the code being compiled is FP or not makes no difference. An FP instruction is a number of bytes in a dictionary, no different than any other instruction, as far as the compiler is concerned.

Ivy or Haswell with Hyperthreading should be the fastest, by some degree, but the dedicated int units on the AMDs should make the FX-63x0 and FX-83x0 reasonable, if your electricity costs are low.

Best I could find, with Openbenchmarking's search currently off:

http://openbenchmarking.org/result/1409038-SO-QQWE3829779
http://openbenchmarking.org/result/1409055-PL-QHQM0022984 (near the bottom)
http://openbenchmarking.org/result/1408317-SO-TESTBUILD46
http://openbenchmarking.org/result/1408319-LI-XXXLINUXC39

The edge in performance and power consumption go to Intel. A new Core i5 should be about as fast as an FX-8350, while using less power, and having better IGP support overall. HT only gives about 20-30%, it looks like (typical, but actually good for the AMD CPUs), but the performance per core of Intel's is just plain higher.

 

[DrMrLordX]

Compared with even Sandy Bridge, not really. Maybe back when Sandy Bridge didn't exist perhaps.

No, but it would beat the pants off Piledriver if those results are any indicator. Granted, the clockspeeds would be more than a bit lower . . .

 

[Zodiark1593]

That's insanely impressive -- I knew Nvidia GPU's were awesome under Linux, but even this goes beyond what I thought was possible.

The 10x mark isn't performance of Linux vs Windows, rather the AMD cpu vs CUDA on the gpu side. It's well knows gpus are vastly quicker than cpus at parallel tasks like rendering.

 

[MiddleOfTheRoad]

The 10x mark isn't performance of Linux vs Windows, rather the AMD cpu vs CUDA on the gpu side. It's well knows gpus are vastly quicker than cpus at parallel tasks like rendering.

I knew the GPU's were quicker -- I didn't realize the margin was that big, though.

Hotter! Faster! Better!
http://www.youtube.com/watch?v=EIPggCgYK38