GPU Owl 3080ti Benchmarks

[Magellan3s]

GPU is an EVGA 3080ti FTW3
OS is Linux 20.04.4

Slight GPU Overclock @
+200 MHZ +1000 Mhz Memory

jesus@Magellan:~/gpuowl-6$ ./gpuowl -prp 57885161 -iters 30000
2022-04-19 10:51:02 gpuowl 
2022-04-19 10:51:02 config: -user Magallan3s -cpu Magellan -maxAlloc 10500M -yield
2022-04-19 10:51:02 config: -prp 57885161 -iters 30000 
2022-04-19 10:51:02 device 0, unique id ''
2022-04-19 10:51:02 Magellan 57885161 FFT: 3M 1K:6:256 (18.40 bpw)
2022-04-19 10:51:02 Magellan Expected maximum carry32: 42500000
2022-04-19 10:51:02 Magellan OpenCL args "-DEXP=57885161u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=6u -DPM1=0 -DMAX_ACCURACY=1 -DWEIGHT_STEP_MINUS_1=0x1.07673850f37p-1 -DIWEIGHT_STEP_MINUS_1=-0x1.5bd9e39e14a3dp-2  -cl-unsafe-math-optimizations -cl-std=CL2.0 -cl-finite-math-only "
2022-04-19 10:51:02 Magellan 

2022-04-19 10:51:02 Magellan OpenCL compilation in 0.00 s
2022-04-19 10:51:03 Magellan 57885161 OK    30000 loaded: blockSize 400, fe1565094c7f7b47
2022-04-19 10:51:03 Magellan validating proof residues for power 8
2022-04-19 10:51:03 Magellan Proof using power 8
2022-04-19 10:51:04 Magellan 57885161 OK    30800   0.05%; 1194 us/it; ETA 0d 19:11; 4f153add2832ca8a (check 0.50s)
2022-04-19 10:51:38 Magellan Stopping, please wait..
2022-04-19 10:51:38 Magellan 57885161 OK    60000   0.10%; 1148 us/it; ETA 0d 18:27; 175901ec29adfa87 (check 0.47s)
2022-04-19 10:51:39 Magellan Exiting because "stop requested"
2022-04-19 10:51:39 Magellan Bye

For Wavefront Exponent

jesus@Magellan:~/gpuowl-6$ ./gpuowl -prp 113613007 -iters 30000
2022-04-19 10:55:04 gpuowl 
2022-04-19 10:55:04 config: -user Magallan3s -cpu Magellan -maxAlloc 10500M -yield
2022-04-19 10:55:04 config: -prp 113613007 -iters 30000 
2022-04-19 10:55:04 device 0, unique id ''
2022-04-19 10:55:04 Magellan 113613007 FFT: 6M 1K:12:256 (18.06 bpw)
2022-04-19 10:55:04 Magellan Expected maximum carry32: 4CFA0000
2022-04-19 10:55:05 Magellan OpenCL args "-DEXP=113613007u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=12u -DPM1=0 -DMM2_CHAIN=1u -DMAX_ACCURACY=1 -DWEIGHT_STEP_MINUS_1=0x1.d7719ff404155p-1 -DIWEIGHT_STEP_MINUS_1=-0x1.eae2bbc5c8218p-2  -cl-unsafe-math-optimizations -cl-std=CL2.0 -cl-finite-math-only "
2022-04-19 10:55:05 Magellan 

2022-04-19 10:55:05 Magellan OpenCL compilation in 0.83 s
2022-04-19 10:55:06 Magellan 113613007 OK        0 loaded: blockSize 400, 0000000000000003
2022-04-19 10:55:06 Magellan validating proof residues for power 8
2022-04-19 10:55:06 Magellan Proof using power 8
2022-04-19 10:55:09 Magellan 113613007 OK      800   0.00%; 2132 us/it; ETA 2d 19:18; 420cf6918603e7e1 (check 0.91s)
2022-04-19 10:56:12 Magellan Stopping, please wait..
2022-04-19 10:56:13 Magellan 113613007 OK    30000   0.03%; 2158 us/it; ETA 2d 20:06; 32d4895e2a4b9a36 (check 0.91s)
2022-04-19 10:56:13 Magellan Exiting because "stop requested"
2022-04-19 10:56:13 Magellan Bye

Same wavefront exponent tested on Intel 12900k (AVX512 clock @ 4.8 mhz) using (Corsair DDR5 ram @ 5400 mhz)

jesus@Magellan:~/Prime95 (copy)$ ./mprime -m 
[Main thread Apr 19 11:02] Mersenne number primality test program version 30.7
[Main thread Apr 19 11:02] Optimizing for CPU architecture: Core i3/i5/i7, L2 cache size: 8x1280 KB, L3 cache size: 30 MB
	    
Your choice: [Main thread Apr 19 11:13] Starting worker.
[Work thread Apr 19 11:13] Worker starting
[Work thread Apr 19 11:13] Setting affinity to run worker on CPU core #1
[Work thread Apr 19 11:13] Setting affinity to run helper thread 1 on CPU core #2
[Work thread Apr 19 11:13] Setting affinity to run helper thread 2 on CPU core #3
[Work thread Apr 19 11:13] Setting affinity to run helper thread 6 on CPU core #7
[Work thread Apr 19 11:13] Setting affinity to run helper thread 7 on CPU core #8
[Work thread Apr 19 11:13] Setting affinity to run helper thread 5 on CPU core #6
[Work thread Apr 19 11:13] Setting affinity to run helper thread 3 on CPU core #4
[Work thread Apr 19 11:13] Setting affinity to run helper thread 4 on CPU core #5
[Work thread Apr 19 11:13] Starting Gerbicz error-checking PRP test of M113613007 using AVX-512 FFT length 6048K, Pass1=1152, Pass2=5376, clm=1, 8 threads
[Work thread Apr 19 11:13] Preallocating disk space for the proof interim residues file p113613007.residues
[Work thread Apr 19 11:13] PRP proof using power=10 and 64-bit hash size.
[Work thread Apr 19 11:13] Proof requires 14.5GB of temporary disk space and uploading a 156MB proof file.
[Work thread Apr 19 11:13] Iteration: 10000 / 113613007 [0.00%], ms/iter:  2.032, ETA: 64:08:00
[Work thread Apr 19 11:13] Iteration: 20000 / 113613007 [0.01%], ms/iter:  1.970, ETA: 62:09:09
[Work thread Apr 19 11:14] Iteration: 30000 / 113613007 [0.02%], ms/iter:  1.962, ETA: 61:54:44
[Work thread Apr 19 11:14] Iteration: 40000 / 113613007 [0.03%], ms/iter:  1.961, ETA: 61:52:20
[Main thread Apr 19 11:14] Stopping all worker windows.
[Work thread Apr 19 11:14] Stopping PRP test of M113613007 at iteration 40747 [0.03%]
[Work thread Apr 19 11:14] Worker stopped.
[Main thread Apr 19 11:14] Execution halted.

[kriesel]

That gpu model family is well suited to TF, and not well suited to gpuowl or other DP work.
A Radeon VII GPU that costs less is more than double the speed in DP. (6M fft, under 890. usec/iter).

Also, please do not proliferate threads when an existing thread will serve. There are existing benchmark threads, for example.

[Prime95]

Originally Posted by Magellan3s

GPU is an EVGA 3080ti FTW3
OS is Linux 20.04.4

Slight GPU Overclock @
+200 MHZ +1000 Mhz Memory

Thanks for the info!

[Magellan3s]

Originally Posted by Prime95

Thanks for the info!

You are very welcome George!

Here are the same exponents tested with an NVIDIA A100 (top of the line $12,000 Nvidia compute card).

The A100 NVIDIA compute accelerator card is performing worse than a $~1,000 Radeon VII.

Originally Posted by kriesel

That gpu model family is well suited to TF, and not well suited to gpuowl or other DP work.
A Radeon VII GPU that costs less is more than double the speed in DP. (6M fft, under 890. usec/iter).

Apparently not even the A100 is suited to GPUOWL or other DP work becuase your Radeon VII is performing 6M FFT at a faster rate than the 1050 usec I got with an A100....

[ATH]

Something must be wrong then because when I get an A100 rarely on Google Colab Pro+ with gpuowl I'm getting 389µs on a ~109.8M exponent 6M FFT (and 600µs on a V100).

[Magellan3s]

Originally Posted by Prime95

Thanks for the info!

Originally Posted by ATH

Something must be wrong then because when I get an A100 rarely on Google Colab Pro+ with gpuowl I'm getting 389µs on a ~109.8M exponent 6M FFT (and 600µs on a V100).

This was on a ubuntu 20.04 virtual machine!

[storm5510]

Originally Posted by Magellan3s

GPU is an EVGA 3080ti FTW3
OS is Linux 20.04.4

Slight GPU Overclock @
+200 MHZ +1000 Mhz Memory

I read this as +200 MHz on the GPU core clock and +1,000 MHz on the memory clock. Is this an addition beyond the default settings?

[kriesel]

Originally Posted by ATH

Something must be wrong then

Perhaps the benchmark instance was not the only one running on his Colab session's A100? One more reason to run GPU and CPU GIMPS apps in background and top -d 180 in foreground of Colab; to possibly spot such errors.
Also I wonder what version of gpuowl was used for each; there's a null version output on his pix; no version stated with your timings.

Originally Posted by storm5510

I read this as +200 MHz on the GPU core clock and +1,000 MHz on the memory clock. Is this an addition beyond the default settings?

+1GHz seems a rather ambitious amount of GPU ram overclock. On Radeon VII, AMD's tools only allow up to +200Mhz atop the stock 1GHz. Haven't messed with NVIDIA clocks on RTXxxxx myself though.
GPU specs here.

[M344587487]

Originally Posted by kriesel

...
+1GHz seems a rather ambitious amount of GPU ram overclock. On Radeon VII, AMD's tools only allow up to +200Mhz atop the stock 1GHz. Haven't messed with NVIDIA clocks on RTXxxxx myself though.
GPU specs here.

HBM2 and GDDR6X are very different, +1GHz does seem ambitious (I have no practical experience with it like you) but it may be possible that the nomenclature is just different. If the speed rating of GDDR6X is anything like DDR a +1GHz memory OC may translate to an actual OC of the modules of +250MHz (GDDR6X uses DDR and also PAM4 encoding to double the rate again), whereas an HBM OC of +200MHz will (I think) have a 1:1 correspondence of +200MHz module OC. I could be entirely wrong.

[kriesel]

Originally Posted by M344587487

HBM2 and GDDR6X are very different

Yes. The spec sheet linked earlier gave 1188MHz base clock, 1800 boost, both for the memory on the EVGA FTW3 GPU. Going +1000 MHz from base would be 2188 MHz. Boost tends to be rated more for transient use than steady state, so 2188 for long GIMPS runs would be quite ambitious. It would probably requiring extra cooling or voltage curve reduction or something, or an insensitivity to errors (either not detecting them, in applications where an occasional error has little consequence, such as high refresh rate graphics display, or detecting and correcting in error-intolerant demanding applications like GIMPS). At some point rising power consumption and clock may cause damage.

[tdulcet]

Originally Posted by ATH

Something must be wrong then because when I get an A100 rarely on Google Colab Pro+ with gpuowl I'm getting 389µs on a ~109.8M exponent 6M FFT (and 600µs on a V100).

Originally Posted by Magellan3s

This was on a ubuntu 20.04 virtual machine!

Colab currently use Ubuntu 18.04, but that should not matter. I would try removing the -yield argument from your config file, as that will slow things down and should not be needed for headless GPUs like the A100.