mfakto: an OpenCL program for Mersenne prefactoring

[kracker]

My guess would be drivers. (mfakto on my i5-4670k also uses a full cpu core, while on a i3-8100 it uses almost no resources)

[James Heinrich]

Originally Posted by kracker

mfakto on my ... i3-8100

Thanks for posting that -- my "other" computer has an i3-8100 and I didn't know it could run mfakto.
Free extra 21Ghd/d, thanks!

[Viliam Furik]

It was probably discussed earlier, but can anyone tell me, why can't I squeeze more than 1300 GHz-D/D (2.6 TFLOPS) from Radeon VII, despite it promising 13.8 TFLOPS (about 10 % more than RTX 2080Ti) of SP throughput?

Also in gpuOwl, only about 0.75 TFLOPS despite AMD promising 3.46 TFLOPS.

It seems to always be a fifth of what it should be according to the official AMD page.

[James Heinrich]

Originally Posted by Viliam Furik

It was probably discussed earlier, but can anyone tell me, why can't I squeeze more than 1300 GHz-D/D (2.6 TFLOPS) from Radeon VII, despite it promising 13.8 TFLOPS (about 10 % more than RTX 2080Ti) of SP throughput?

I'm not sure how you're calculating FLOPS performance in mfakto/gpuowl?

The mfakto performance you quote seems to be in line with expected mfakto performance (noting that my chart shows stock-clock performance). Also note AMD's somewhat deceptive practice of quoting "peak performance" numbers, which mean performance at 1750 boost instead of 1400 stock. The numbers on my pages are all stock-clock values.

[kriesel]

Originally Posted by Viliam Furik

It was probably discussed earlier, but can anyone tell me, why can't I squeeze more than 1300 GHz-D/D (2.6 TFLOPS) from Radeon VII, despite it promising 13.8 TFLOPS (about 10 % more than RTX 2080Ti) of SP throughput?

Also in gpuOwl, only about 0.75 TFLOPS despite AMD promising 3.46 TFLOPS.

It seems to always be a fifth of what it should be according to the official AMD page.

If you're using mfakto as the measure of sp performance, note that mfakto & mfaktc use int32, not fp32.

Also the AMD spec sheet says "peak". Gpuowl is memory bandwidth constrained per Mihai. Which would make sustained rate difficult to maintain at peak rate.

https://www.amd.com/en/products/graphics/amd-radeon-vii

https://www.techpowerup.com/gpu-specs/radeon-vii.c3358

[LaurV]

Originally Posted by kriesel

If you're using mfakto as the measure of sp performance, note that mfakto & mfaktc use int32, not fp32

This!
Radeon VII has a poor integer performance.

[axn]

Originally Posted by James Heinrich

I'm not sure how you're calculating FLOPS performance in mfakto/gpuowl?

He's using standard Primenet conversion 1Gd/d = 2GFLOPS

[kriesel]

A small utility to list the OpenCL platforms and devices on them and a bit of description.
See https://www.mersenneforum.org/showpo...74&postcount=6

Could help the occasional mfakto installation startup.

[DrobinsonPE]

After switching to Windows 10, I got mfakto to run on my ASRock Deskmini A300W, AMD A8-9600, 16GB DDR-4, SSD.

Selftest statistics
  number of tests           34026
  successful tests          34026

selftest PASSED!

mfakto 0.15pre7-MGW (64bit build)
OpenCL device info
  name                      Bristol Ridge (Advanced Micro Devices, Inc.)
  device (driver) version   OpenCL 2.0 AMD-APP (2841.19) (2841.19)
  maximum threads per block 1024
  maximum threads per grid  1073741824
  number of multiprocessors 6 (384 compute elements)
  clock rate                900 MHz

Automatic parameters
  threads per grid          0
  optimizing kernels for    GCN

Loading binary kernel file mfakto_Kernels.elf
Compiling kernels.
  GPUSievePrimes (adjusted) 81206
  GPUsieve minimum exponent 1037054
Started a simple selftest ...
Selftest statistics
  number of tests           30
  successful tests          30

selftest PASSED!

got assignment: exp=212335483 bit_min=73 bit_max=74 (9.01 GHz-days)
Starting trial factoring M212335483 from 2^73 to 2^74 (9.01 GHz-days)
Using GPU kernel "cl_barrett15_74_gs_2"
Date    Time | class   Pct |   time     ETA | GHz-d/day    Sieve     Wait
Dec 05 23:13 | 4617 100.0% | 11.685   0m00s |     69.39    81206    0.00%
no factor for M212335483 from 2^73 to 2^74 [mfakto 0.15pre7-MGW cl_barrett15_74_gs_2]
tf(): total time spent:  3h  6m 46.416s (69.46 GHz-days / day)

Next up is a little tuning to see if it can get above 70 GHz-days / day.

[DrobinsonPE]

Here is the results of mfakto tuning on the ASRock Deskmini A300W, AMD A8-9600, Radeon R7 IGPU, 16GB DDR-4, SSD, Windows 10, mfakto 0.15pre7, Prime95 v30.3 b6 system.

mfakto tuning.
AMD A8-9600, Radeon R7 IGPU, 16GB ram, SSD, Windows 10, mfakto 0.15pre7
exp=103122301 bit 73 to 74 
Initial settings and speed. 
GPUSieveProcessSize=24, GPUSieveSize=96, GPUSievePrimes=81157, 67.97GHz-day
Final settings and speed.
GPUSieveProcessSize=32, GPUSieveSize=128, GPUSievePrimes=179766, 70.39GHz-day
Speed with prime95 also running a P-1 stage 2, 70.32GHz-day

Step 1, vary GPUSieveProcessSize
Possible values: 8, 16, 24, 32
# Also must divide GPUSieveSize * 1024
# Default: GPUSieveProcessSize=24

GPUSieveProcessSize=8	64.93GHz-day
GPUSieveProcessSize=16	67.10GHz-day
GPUSieveProcessSize=24	67.97GHz-day
GPUSieveProcessSize=32	68.68GHz-day	*

Step 2: vary GPUSieveSize with GPUSieveProcessSize=32
# Minimum: GPUSieveSize=4
# Maximum: GPUSieveSize=128
# Default: GPUSieveSize=96
GPUSieveSize=32		68.36GHz-day
GPUSieveSize=64		68.54GHz-day
GPUSieveSize=96		68.68GHz-day
GPUSieveSize=128	68.70GHz-day	*

Step 3: vary GPUSievePrimes with GPUSieveSize=128, GPUSieveProcessSize=32
# Minimum: GPUSievePrimes=54
# Maximum: GPUSievePrimes=1075766
# Default: GPUSievePrimes=81157
GPUSievePrimes=21814	62.88GHz-day
GPUSievePrimes=67894	68.35GHz-day	
GPUSievePrimes=81157	68.70GHz-day	
GPUSievePrimes=99894	69.10GHz-day
GPUSievePrimes=120374     69.52GHz-day
GPUSievePrimes=139830     69.84GHz-day
GPUSievePrimes=160310     70.14GHz-day
GPUSievePrimes=179766     70.39GHz-day*
GPUSievePrimes=200246     70.22GHz-day

[Dylan14]

I have created a PKGBUILD for this software for use on Arch. You can find it here: https://aur.archlinux.org/packages/mfakto/