Quote from: kerney666 on March 30, 2019, 01:17:10 PM

Quote from: professorkuusamo on March 30, 2019, 12:44:23 PM

Quote from: kerney666 on March 30, 2019, 12:08:18 PM

Quote from: professorkuusamo on March 30, 2019, 09:50:09 AM

Quote from: seefatlow on March 30, 2019, 09:19:07 AM

OK. I have been trying over several days to replicate the results from pbfarmer on cnv8_trtl and I can say that I am unable to get the efficiency reported by him. Any attempt to reach close to his reported voltage levels of 805mv to 840mV resulted in random DEAD GPUs(stuck in enqueue) after 20-30 minutes hashing or significant hashrate drop per GPU

Even at 870mV, I am still unable to get a stable system running. Dead GPUs and hashrate drop to 15kH/s still occurs after several tens of minutes of mining. And the weird thing is hashrate drops happens to Hynix mem GPUs only. Huh

Somehow CN-TRTL algo is more taxing that CN_R? I am able to run CN_R without failures for a week and with lower voltage settings (850mV - 870mV). I can't seem to do this for CN_TRTL

My settings are as below. But the hashrates aren't sustainable

GPUs : Ref Vega 64 and Vega 56 reference bios
V64 cclk/memclk 1220/1100 @ 870mV L28+28 (Samsung mem) - 19.5kH/s
V56 cclk/memclk 1220/940 @ 870mV L24+24 (Samsung mem) - 19.3kH/s (Hynix mem) - 18.7 kH/s
ATW power draw 190W per GPU
Adrenalin Driver 18.6.1

Kerney/Todd,

Any ideas what could possibly be wrong? Unoptimized CN_TRTL code?

i have same situation.unstable only on trtl.try different drivers nothing change.

Hey guys!

Well, unoptimized code isn't the issue, it's rather that it's optimized so much that it's pounding the gpu more than anything we've produced before, especially the memory subsystem.

I'm not chasing efficiency to the same degree that you professional tuners are since my rig(s) are a combination of test and more serious mining. However, on my 8 x Vega 56 ref cards flashed to V64, win10, 18.6.1, clocks at 1408@900, 1100@900, I have zero stability issues, it's been mining for 14h straight now in the current run, and much longer before that as well. I've just stopped it to reconfigure things a few times.

However, on my Vega 64 Liquid Cooling in my dev workstation, CN-trtl is the first algo I've ever seen that kills that specific card but not my blower Vegas. It dies after 1-2h, also running at 1408@900, 1100@900. Effective clocks+voltages in hwinfo64 look very similar to the V56s.

Right now, I'm doing tests with the single-threaded config support that we also added in 0.4.3. This means I'm running --cn_config=L56+0 instead of --cn_config=L28+28. On my V64 LC, hashrate drops from 19.6 kh/s to 18.8 kh/s, same efficiency. The hashrate is expected to drop a little, but this should be more lean on the gpu, it won't be going full throttle on all parts of the hardware at the same time. Theory vs practice is always a bitch though, I'll report back in a few hours on the results.

Meanwhile, you're of course free to try the same trick: switch your problematic Vegas to L56+0 and see if it helps, I'd love to get some more data here.

56 on samsung are stable, works since release 0.4.3.crashed only in rigs with hynix.

Got it, although Ive gotten a nr of reports from liquid V64s having issues as well. Regardless, does anything change for you if you try L56+0 on those Hynix V56s?

hash drop to 14 and power to 150w.unstable.