Haswell-E Prelim. Benchmark

sdbardwick · 2014-09-02, 01:43

i7-5820k @ stock, 4x4GB RAM @ 2133 (XMP not enabled yet), HT disabled. Absolutely non-optimized setup; for all I know right now it might be 20% thermal throttled.

ewmayer · 2014-09-02, 01:49

Differential diagnosis, please. (I.e. compare - preferably in "overall XY% faster per cycle" summary mode - to older Haswell running same benchmarks).

We seek insight, not data dumps. :)

Comparing the CPU temperature range during the self-tests to an older Haswell doing same might shed light on the likelihood of thermal throttling. Again, differential diagnosis is your friend.

sdbardwick · 2014-09-02, 02:35

Quote:

Originally Posted by ewmayer

Differential diagnosis, please. (I.e. compare - preferably in "overall XY% faster per cycle" summary mode - to older Haswell running same benchmarks).

We seek insight, not data dumps. :)

Comparing the CPU temperature range during the self-tests to an older Haswell doing same might shed light on the likelihood of thermal throttling. Again, differential diagnosis is your friend.

Tough room.

Data acquisition before analysis.

Attached are results with XMP enabled, so RAM @ 2800. The motherboard utility also boosted CPU clock by approx. 300MHz for some reason.

ewmayer · 2014-09-02, 05:06

You left out "...and analysis before presentation." :)

Batalov · 2014-09-02, 07:43

Quote:

Originally Posted by ewmayer

Differential diagnosis, please.

It's not lupus.

axn · 2014-09-02, 08:28

It would seem that running 1 test with 6 threads is better than running 6 threads with 1 worker each.

Code:

FFT   6t timing 6t thruput 6w thruput
1024K    0.852   1173.71     1043.83
1280K    1.072    932.84      822.41
1536K    1.284    778.82      679.03
1792K    1.553    643.92      567.52
2048K    1.802    554.94      479.4
2560K    2.409    415.11      385.5
3072K    2.968    336.93      299.52
3584K    3.591    278.47      254.08
4096K    4.194    238.46      ???.??
5120K    5.309    188.36      ???.??
6144K    6.401    156.23      156.16
7168K    7.583    131.87      132.41
8192K    8.681    115.19      115.17

Code:

FFT   6t timing 6t thruput 6w thruput
1024K    0.747   1338.69    1155.58
1280K    0.943   1060.45     915.91
1536K    1.139    877.96     755.36
1792K    1.37     729.93     634.42
2048K    1.575    634.92     542.75
2560K    2.126    470.37     428.64
3072K    2.651    377.22     362.52
3584K    3.203    312.21     305.53
4096K    3.762	  265.82     264.48
5120K    4.763    209.95     211.43
6144K    5.729    174.55     175.0
7168K    6.828    146.46     147.47
8192K    7.776    128.6      129.21

ewmayer · 2014-09-02, 21:14

Quote:

Originally Posted by axn

It would seem that running 1 test with 6 threads is better than running 6 threads with 1 worker each.

How about the following combos?

o 1 x 4-thread, 1 x 2-thread

o 3 x 2-thread

o 2 x 2-thread (i.e. idling 2 cores)

o 1 x 4-thread (again idling 2 cores)

Not sure if Prime95 allows such combos, if not you'll have to run several instances and manually derive the throughput, or "effective per-iteration time" -- assuming all jobs running same FFT length one just takes the various per-iter times t1, 2, ... and computes

t* = 1/(1/t1 + 1/t2 + ...)

Things get trickier if the various jobs are running different FFT lengths.

Prime95 · 2014-09-02, 21:20

Quote:

Originally Posted by ewmayer

How about the following combos?

There is a way to do this with the prime95 benchmark:

Code:

In Addition To The Benchmarking Options Above, The Following Options Are
Available In The Multiple Workers Benchmark
	Benchmultipleworkers=0 Or 1	(Default Is 1)
	Benchtime=N			(Default Is 10)
	Benchhyperthreads=0 Or 1	(Default Is 1)
	Benchmultithreads=0 Or 1	(Default Is 0)
	Benchoddmultithreads=0 Or 1	(Default Is 0)
Benchmultipleworkers Can Be Used To Disable Benchmarking Multiple Workers.
Benchtime Can Be Used To Change The Number Of Seconds To Run Each Benchmark.
Benchhyperthreads Controls Can Be Used To Turn Off The Hyperthreaded Benchmarks.
Benchmultithreads Can Be Used To Also Benchmark Multi-Threaded Ffts (E.G. 2 Workers On 4 Cpus).
Benchoddmultithreads Can Be Used To Also Benchmark Asymetric Multi-Threaded Ffts
Combinations Such As 2 Workers On 3 Cpus.

The multithread observation is odd. I don't know of any other CPUs where multithreaded is better than multiworker.

sdbardwick · 2014-09-02, 21:35

Quote:

Originally Posted by ewmayer

How about the following combos?

o 1 x 4-thread, 1 x 2-thread

o 3 x 2-thread

o 2 x 2-thread (i.e. idling 2 cores)

o 1 x 4-thread (again idling 2 cores)

Not sure if Prime95 allows such combos, if not you'll have to run several instances and manually derive the throughput, or "effective per-iteration time" -- assuming all jobs running same FFT length one just takes the various per-iter times t1, 2, ... and computes

t* = 1/(1/t1 + 1/t2 + ...)

Things get trickier if the various jobs are running different FFT lengths.

Yep, I think I did something very similar with my AMD 1090T hex-core back in the day.

Quote:

Originally Posted by Prime95

There is a way to do this with the prime95 benchmark:

Code:

In Addition To The Benchmarking Options Above, The Following Options Are
Available In The Multiple Workers Benchmark
	Benchmultipleworkers=0 Or 1	(Default Is 1)
	Benchtime=N			(Default Is 10)
	Benchhyperthreads=0 Or 1	(Default Is 1)
	Benchmultithreads=0 Or 1	(Default Is 0)
	Benchoddmultithreads=0 Or 1	(Default Is 0)
Benchmultipleworkers Can Be Used To Disable Benchmarking Multiple Workers.
Benchtime Can Be Used To Change The Number Of Seconds To Run Each Benchmark.
Benchhyperthreads Controls Can Be Used To Turn Off The Hyperthreaded Benchmarks.
Benchmultithreads Can Be Used To Also Benchmark Multi-Threaded Ffts (E.G. 2 Workers On 4 Cpus).
Benchoddmultithreads Can Be Used To Also Benchmark Asymetric Multi-Threaded Ffts
Combinations Such As 2 Workers On 3 Cpus.

The multithread observation is odd. I don't know of any other CPUs where multithreaded is better than multiworker.

Cool! Thanks for the heads up (and new benchmark code - I was surprised with the addition (guess I haven't been diligent in tracking progress lately))!

Might not get to it until tomorrow - I've been unexpectedly busy this holiday weekend. The parts for the -E box were purchased Saturday, and they sat (taunting me) until Monday evening (they started mocking me, so I had to do something).

henryzz · 2014-09-02, 22:31

Does doing one test with more than one thread use more memory than just one thread? If so how much?

Single thread per test calculations http://mersenneforum.org/showpost.ph...&postcount=485

LaurV · 2014-09-03, 03:24

Quote:

Originally Posted by Prime95

The multithread observation is odd. I don't know of any other CPUs where multithreaded is better than multiworker.

It looks like a memory bottleneck to me. Single worker gets all its data in time, multiworker (who need more) does not.

I think we should wait for the "real", "big chip" DDR4.

The actual "products" are still in their infancy, and I don't know so many boards taking full advantage of the new channeled architecture...

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Haswell Preview Benchmark	kracker	Hardware	543	2015-10-05 05:28
Prime95 and Haswell	Pleco	Information & Answers	22	2014-07-13 16:03
Haswell Rig	Mini-Geek	Hardware	64	2014-05-27 13:22
LLR benchmarks for Haswell	Unregistered	Information & Answers	0	2013-06-16 21:48
Haswell New Instructions / AVX2	ixfd64	Hardware	72	2013-03-20 00:00

2014-09-02, 01:49	#2
ewmayer ∂²ω=0 Sep 2002 República de California 103·113 Posts	Differential diagnosis, please. (I.e. compare - preferably in "overall XY% faster per cycle" summary mode - to older Haswell running same benchmarks). We seek insight, not data dumps. :) Comparing the CPU temperature range during the self-tests to an older Haswell doing same might shed light on the likelihood of thermal throttling. Again, differential diagnosis is your friend.

2014-09-02, 05:06	#4
ewmayer ∂²ω=0 Sep 2002 República de California 103·113 Posts	You left out "...and analysis before presentation." :)

2014-09-02, 22:31	#10
henryzz Just call me Henry "David" Sep 2007 Cambridge (GMT/BST) 2³×3×5×7² Posts	Does doing one test with more than one thread use more memory than just one thread? If so how much? Single thread per test calculations http://mersenneforum.org/showpost.ph...&postcount=485