RDNA2 / Big Navi

[Viliam Furik]

Originally Posted by preda

The cache (L1/L2/L3) is used transparently for the *global* memory operations. It is managed automatically by the cache control (probably a variant of LRU), not explicitly by the software. So yes, GpuOwl should benefit from L3 without code changes.

Separate from the caches there is the *local* memory (LDS), which is managed explicitly by the software.

Thanks!

[moebius]

Sad but true, the "best buck for the bang at PRP" at my location is the ASUS Radeon™ RX Vega 56 ROG Strix OC 8GB for 272€ (334 US-$) at Saturn/Media Markt, since the big brother, the Vega 64 and the Radeon VII, are practically no longer available. I don't have any specific values ​​for the card, but they should be slightly below Vega 64. But that is also due to various people who should actually be shot dead, like this miner Four times the price for the current generation graphic cards with only twice the performance is a cheekiness to say the least.
https://pbs.twimg.com/card_img/13461...jpg&name=small

[M344587487]

The person that did the 6900XT benchmarks is looking to sell it or a reference 6900XT for MSRP (they just acquired a reference model and haven't decided which one to keep yet). They're from the US so are looking for a US buyer. If you're interested PM me and I'll get you in contact. I met them a month ago by answering a forum question about rocm, I can't vouch for them (beyond seeming genuine) any more than I can vouch for most of you, so it's at your own risk. I can supply all correspondence with them if you're interested and let you make up your own mind.


https://forum.level1techs.com/t/big-...port/166054/10

[Xyzzy]

RX 6800

2021-01-07 13:33:05 gfx1030-0 OpenCL compilation in 2.54 s
2021-01-07 13:33:05 gfx1030-0 77936867 OK        0 loaded: blockSize 400, 0000000000000003
2021-01-07 13:33:06 gfx1030-0 77936867 OK      800   0.00%;  871 us/it; ETA 0d 18:51; 1579c241dc63eca6 (check 0.39s)
2021-01-07 13:36:00 gfx1030-0 77936867 OK   200000   0.26%;  870 us/it; ETA 0d 18:47; f0b04b45b0855bd2 (check 0.39s)
2021-01-07 13:38:55 gfx1030-0 77936867 OK   400000   0.51%;  873 us/it; ETA 0d 18:48; c03f94396a5aa29e (check 0.39s)
2021-01-07 13:41:49 gfx1030-0 77936867 OK   600000   0.77%;  869 us/it; ETA 0d 18:40; b9decd65ca71b629 (check 0.38s)
2021-01-07 13:44:43 gfx1030-0 77936867 OK   800000   1.03%;  866 us/it; ETA 0d 18:33; 21ebf3636148f663 (check 0.41s)
2021-01-07 13:47:35 gfx1030-0 77936867 OK  1000000   1.28%;  862 us/it; ETA 0d 18:25; 9bf9d9e6bff4286e (check 0.39s)
2021-01-07 13:50:29 gfx1030-0 77936867 OK  1200000   1.54%;  868 us/it; ETA 0d 18:30; da42731132f00140 (check 0.38s)
2021-01-07 13:53:23 gfx1030-0 77936867 OK  1400000   1.80%;  864 us/it; ETA 0d 18:22; 56d6868b11382d6c (check 0.38s)

[M344587487]

Nice. Do you mind doing a benchmark spanning every FFT? It might give us an idea of how infinity cache scales. If you can run bash scripts with WSL then this might do the trick executed from the gpuowl directory:

#!/bin/bash
for W in 256 512 1024 4096; do
    for H in 256 512 1024; do
        for M in 1 3 4 5 6 7 8 9 10 11 12 13 14 15; do
            if [ $M -gt 1 ] || [ $((W*H)) -lt $((512*512)) ]; then
                E=$((W*M*H*2*6+1001))
                ./gpuowl -prp $E -fft $W:$M:$H -nospin -iters 10000 >>benchfft.log
                rm -r $E
            fi
        done
    done
done

Alternatively, attached is an unrolled batch file that calls gpuowl.exe and also needs to be executed from the gpuowl directory. It has the benefit of being sorted ascending by exponent (if you don't want to do the very large tests they can be deleted, no idea how long the very large FFT's take but at only 10k iterations everything else shouldn't take too long) but with the drawback of being completely untested (instead of just mostly untested like the bash script is) and it does no cleanup.

[Viliam Furik]

Two instance benchmarks would be great, too. Possibly for all FFTs.

[Xyzzy]

We will look into running the benchmark.

We have two of these cards that we bought. Both were (up until today) factory sealed.

If someone wants the second card drop us a PM.

[Viliam Furik]

Originally Posted by Xyzzy

We will look into running the benchmark.

Could you please run some mfakto, too? Possibly some 2M range exponent from 67 to 68 bits, or similar (I have observed peak performance for my 2080Ti in this range).

[Xyzzy]

Originally Posted by Viliam Furik

Could you please run some mfakto, too? Possibly some 2M range exponent from 67 to 68 bits, or similar (I have observed peak performance for my 2080Ti in this range).

Do you have a worktodo.txt file of the ranges to be tested?

[Viliam Furik]

Originally Posted by Xyzzy

Do you have a worktodo.txt file of the ranges to be tested?

I thought you'll just pick some exponent, but I can give you one, too.

Factor=2000003,67,68

One exponent is enough for me. Just make sure that the GPU is not doing any heavy work besides the mfakto.

[Xyzzy]

OpenCL device info
  name                      gfx1030 (Advanced Micro Devices, Inc.)
  device (driver) version   OpenCL 2.0 AMD-APP (3188.4) (3188.4 (PAL,LC))
  maximum threads per block 1024
  maximum threads per grid  1073741824
  number of multiprocessors 30 (1920 compute elements)
  clock rate                1815 MHz

got assignment: exp=2000003 bit_min=67 bit_max=68 (14.95 GHz-days)
Starting trial factoring M2000003 from 2^67 to 2^68 (14.95 GHz-days)
Using GPU kernel "cl_barrett15_69_gs_2"
Date    Time | class   Pct |   time     ETA | GHz-d/day    Sieve     Wait
Jan 09 19:52 | 4617 100.0% |  0.516   0m00s |   2606.77    81206    0.00%
no factor for M2000003 from 2^67 to 2^68 [mfakto 0.15pre7-MGW cl_barrett15_69_gs_2]
tf(): total time spent:  8m  5.565s (2659.35 GHz-days / day)