Radeon VII (2nd gen consumer Vega GPU)

[M344587487]

The price of the PowerColor version on ebuyer has dropped by £10 to £640. First time I've seen the price drop below £650 which is interesting: https://www.ebuyer.com/875303-powerc...i-16gbhbm2-3dh

[preda]

I now have a Radeon VII, I'd like to share my setup. I overclocked the RAM from 1000 to 1100, and undervolted a bit.
Initial pp_od_clk_voltage:
OD_SCLK:
0: 808Mhz
1: 1801Mhz
OD_MCLK:
1: 1000Mhz
OD_VDDC_CURVE:
0: 808Mhz 695mV
1: 1304Mhz 791mV
2: 1801Mhz 1089mV

Modified to:
D_SCLK:
0: 808Mhz
1: 1801Mhz
OD_MCLK:
1: 1100Mhz
OD_VDDC_CURVE:
0: 808Mhz 695mV
1: 1304Mhz 780mV
2: 1801Mhz 1070mV

I run it with --setsclk 4 (1547Mhz). I get 0.92ms/it at wavefront (FFT 4608K), and the GPU uses 160W. Fan on auto, temperature reported by sensors is 105C, but I suspect this value is with 20C over the real temperature (because the limit is reported at 118C).

Originally Posted by preda

I now have a Radeon VII, I'd like to share my setup. I overclocked the RAM from 1000 to 1100, and undervolted a bit.
Initial pp_od_clk_voltage:
OD_SCLK:
0: 808Mhz
1: 1801Mhz
OD_MCLK:
1: 1000Mhz
OD_VDDC_CURVE:
0: 808Mhz 695mV
1: 1304Mhz 791mV
2: 1801Mhz 1089mV

Modified to:
D_SCLK:
0: 808Mhz
1: 1801Mhz
OD_MCLK:
1: 1100Mhz
OD_VDDC_CURVE:
0: 808Mhz 695mV
1: 1304Mhz 780mV
2: 1801Mhz 1070mV

I run it with --setsclk 4 (1547Mhz). I get 0.92ms/it at wavefront (FFT 4608K), and the GPU uses 160W. Fan on auto, temperature reported by sensors is 105C, but I suspect this value is with 20C over the real temperature (because the limit is reported at 118C).

I have ambient temp 22~24 C and the gpu is at 107 C with good cooling.

[preda]

With its 16GB of RAM, R7 is also quite good at P-1.
I'm doing P-1(B1=300K,B2=9M) on 91M exponents in about 17min/test.
(and I found this 118-bit factor: 420168247365933163207630527781851871 )

Originally Posted by preda

With its 16GB of RAM, R7 is also quite good at P-1.
I'm doing P-1(B1=300K,B2=9M) on 91M exponents in about 17min/test.
(and I found this 118-bit factor: 420168247365933163207630527781851871 )

I am waiting to deplenish my current worktodo.txt before starting P-1 :-)

[M344587487]

Originally Posted by preda

With its 16GB of RAM, R7 is also quite good at P-1.
I'm doing P-1(B1=300K,B2=9M) on 91M exponents in about 17min/test.
(and I found this 118-bit factor: 420168247365933163207630527781851871 )

Do you have to manually set anything to do P-1 with gpuowl? I'm not familiar with P-1 and these lines generated from this calculator were ignored: https://www.mersenne.ca/prob.php

Pminus1=1,2,344587487,-1,1645000,32900000,82
 Pfactor=1,2,344587487,-1,82,2

edit: I figured it out, the calculator is older than an identifying hash being added to the line.

[preda]

Originally Posted by M344587487

Do you have to manually set anything to do P-1 with gpuowl? I'm not familiar with P-1 and these lines generated from this calculator were ignored: https://www.mersenne.ca/prob.php

Pminus1=1,2,344587487,-1,1645000,32900000,82
 Pfactor=1,2,344587487,-1,82,2

edit: I figured it out, the calculator is older than an identifying hash being added to the line.

Yes, assignments like
PFactor=AID,1,2,91157513,-1,77,2

these are easily generated/submitted by e.g.:
gpuowl/primenet.py -u user -p passwd --dirs workdir -w PM1 --tasks 40

and for B1/B2 I add to openowl e.g.:
./openowl -B1 300000
./openowl -B1 1000000 -rB2 25

Bounds (B1,B2) can also be specified per-exponent, by prefixing the worktodo line with:
B1=x;line
B1=x,B2=y;line

[M344587487]

Thanks. Could it make sense to do a P-1 test and a PRP test simultaneously? Two PRP tests improves throughput, two P-1 cannot be done simultaneously as they both want to max out RAM. One of each could make the most of the hardware.

[preda]

Originally Posted by M344587487

Thanks. Could it make sense to do a P-1 test and a PRP test simultaneously? Two PRP tests improves throughput, two P-1 cannot be done simultaneously as they both want to max out RAM. One of each could make the most of the hardware.

Yes I think that would work. I'm not doing it myself, because the benefit is too small IMO, but it may be worth if you're patient. You should watch the memory use reported by the P-1 at start of test (e.g. "P-1 GPU RAM fits 388 stage2 buffers @ 40.0 MB each, using 360"), to make sure the PRP has enough space -- in this example there is plenty of buffers between 360 and 388 for the PRP.

[kriesel]

Originally Posted by preda

With its 16GB of RAM, R7 is also quite good at P-1.
I'm doing P-1(B1=300K,B2=9M) on 91M exponents in about 17min/test.
(and I found this 118-bit factor: 420168247365933163207630527781851871 )

For such an exponent, B1=300K, or B2=9M, look low to me.
See for example https://www.mersenne.org/report_expo...1140901&full=1
B1= 745 000
B2= 17 135 000
run in CUDAPm1 v0.20 on a GTX1070 (8GB)

Or M90888739, B1=870000, B2=18052500, e=12, in CUDAPm1 v0.22 on a GTX1080Ti (11GB)

[M344587487]

Here's a rough test exploring PRP and P-1 simultaneously. Timings in various configs of a 90M P-1 (B1=500000, B2=15000000) and 88M PRP, auto test settings, 1546 SCLK 1200 MCLK:

[PRP ms/it]  [PRP ms/it] [P-1 Stage 1 ms/it] [P-1 Stage 2 ms/it] [Power W] [P-1 Stage Completion minutes est.]
 1.02         0           0                   0                   170
 1.89         1.89        0                   0                   180
 0            0           1.02                0                   170       ~12
 0            0           0                   1.2                 160       ~16
 1.89         0           1.89                0                   180       ~22
 2.06         0           0                   2.09                175       ~28

Which I make out to be:

• ~7.9% increase in throughput from 1 PRP to 2 simultaneous PRP at ~6% increase in power. Definitely worth it
• P-1 running at ~56% speed relative to solo and PRP running at ~51.5% speed relative to solo at ~6% increase in power. Still worth it for throughput but the power numbers make efficiency less certain.