Expectations of Ryzen Threadripper 6000 Series

[JWNoctis]

Originally Posted by moebius

The Radeon RX Vega 7 Graphics of Ryzen 5 5600G is similar in speed to the GeForce GTX 1650 at PRP-Testing.
https://mersenneforum.org/showpost.p...postcount=2729

While neither are even half as quick as the CPU part of the SoC, and doing PRP-test with iGPU will worsen memory bandwidth contention and performance, if the CPU part, which is already memory bandwidth-bound, is doing the same.

Would be interesting to try and quantify, though.

[DrobinsonPE]

Originally Posted by JWNoctis

Would be interesting to try and quantify, though.

Not that this has anything to do with a Threadripper 6000 Series but I cannot resist....I tested the APU with different combinations of Prime 95, mfakto and gpuowl the first few days before tuning anything or adjusting any settings. You are correct that with Prime 95 and gpuowl both running they fought over the memory bandwidth and the combined output was lower than Prime95 alone. One of these days I am going to re-test with the CPU clocked lower and the memory clocked higher to see if it makes a difference.

Here is some of my initial data

Prime 95, AMD Ryzen 5 5600G, ASROCK B450-HDV R4.0, DDR4-3600 RAM, PBO
101 watt load at Kill A Watt, CPU Clock at 3840 MHz - Ryzen Master
----------------------------------------------------------------
[Work thread Aug 25 04:14] Starting primality test of M64069669 using FMA3 FFT length 3360K, Pass1=896, Pass2=3840, clm=1, 6 threads
[Work thread Aug 25 04:15] Iteration: 20000 / 64069669 [0.03%], ms/iter:  2.319, ETA: 41:15:31
[Work thread Aug 25 04:15] Iteration: 40000 / 64069669 [0.06%], ms/iter:  2.317, ETA: 41:12:30
[Work thread Aug 25 04:16] Iteration: 60000 / 64069669 [0.09%], ms/iter:  2.313, ETA: 41:07:47

GPUOWL, AMD Ryzen 5 5600G, ASROCK B450-HDV R4.0, DDR4-3600 RAM
64.4 watt load at Kill A Watt, APU Clock at 1900 MHz - Ryzen Master
----------------------------------------------------------------
C:\Users\User\GPUOWL\v611380>gpuowl-win
2021-08-24 21:46:12 gpuowl v6.11-380-g79ea0cc
2021-08-24 21:46:12 gfx90c-0 110004803 FFT: 6M 1K:12:256 (17.48 bpw)
2021-08-24 21:46:20 gfx90c-0 110004803 OK      800 loaded: blockSize 400, 49130cc65031159a
2021-08-24 21:46:20 gfx90c-0 Proof using power 8
2021-08-24 21:46:36 gfx90c-0 110004803 OK     1600   0.00%; 13104 us/it; ETA 16d 16:26; 3cb934adb14f7f96 (check 5.27s)
2021-08-24 21:48:31 gfx90c-0 110004803 OK    10000   0.01%; 13044 us/it; ETA 16d 14:33; 2cdce45e530be774 (check 5.27s)
2021-08-24 21:50:46 gfx90c-0 110004803 OK    20000   0.02%; 13045 us/it; ETA 16d 14:33; 109b7e06e48b8dba (check 5.27s)

Prime 95 and GPUOWL, AMD Ryzen 5 5600G, ASROCK B450-HDV R4.0, DDR4-3600 RAM, PBO
112 watt load at Kill A Watt, CPU clock at 4070MHz, APU Clock at 1900 MHz - Ryzen Master
-----------------------------------------------------------------
[Work thread Aug 25 06:56] Iteration: 4100000 / 64069669 [6.39%], ms/iter:  3.499, ETA: 58:17:15
[Work thread Aug 25 06:57] Iteration: 4120000 / 64069669 [6.43%], ms/iter:  3.492, ETA: 58:09:23
[Work thread Aug 25 06:58] Iteration: 4140000 / 64069669 [6.46%], ms/iter:  3.491, ETA: 58:07:13
[Work thread Aug 25 06:59] Iteration: 4160000 / 64069669 [6.49%], ms/iter:  3.507, ETA: 58:21:42
[Work thread Aug 25 07:01] Iteration: 4180000 / 64069669 [6.52%], ms/iter:  3.492, ETA: 58:05:55

C:\Users\User\GPUOWL\v611380>gpuowl-win
2021-08-25 06:54:22 gpuowl v6.11-380-g79ea0cc
2021-08-25 06:54:22 config: -log 10000
2021-08-25 06:55:01 gfx90c-0 110004803 OK    20800   0.02%; 19691 us/it; ETA 25d 01:35; cd21f0f8d5fdcd56 (check 8.09s)
2021-08-25 06:58:09 gfx90c-0 110004803 OK    30000   0.03%; 19563 us/it; ETA 24d 21:38; ced6b9f5dde22bde (check 7.92s)
2021-08-25 07:01:32 gfx90c-0 110004803 OK    40000   0.04%; 19550 us/it; ETA 24d 21:11; 31dcd5040e2e40b9 (check 7.92s)

[moebius]

Originally Posted by JWNoctis

While neither are even half as quick as the CPU part of the SoC, and doing PRP-test with iGPU will worsen memory bandwidth contention and performance, if the CPU part, which is already memory bandwidth-bound, is doing the same.

Would be interesting to try and quantify, though.

Sure the iGPU use shared system memory, it also depends on the latencys of your DRAM-Kit.