Navi (RX 5700, RX 5700XT)

Originally Posted by ET_

Waiting for the tests

almost ready


https://www.phoronix.com/scan.php?pa...RM-Brings-Navi

[M344587487]

Originally Posted by hansl

Interesting thermal performance improvement over stock graphite pad, maybe simlarly applicable to Radeon VII's:

I've removed the pad, applied decent thermal paste and used washers on a Radeon VII. The difference was a minimal improvement, low single digit centigrade, which is not to say that someone else doing it won't get better results but for our purposes there's not much point. IMO everyone prime hunting on Radeon VII should underclock the core with "rocm-smi --setsclk 4" and overclock the memory to 1100 (if overclocking memory manually set fan speed to 120, higher if ambient is high) and be done with it.

[retina]

Originally Posted by M344587487

IMO everyone prime hunting on Radeon VII should underclock the core with "rocm-smi --setsclk 4"...

Yeah. I agree. This also applies to CPUs. There seems little point in overclocking a CPU when it can outpace the memory by a handy margin.

Originally Posted by M344587487

... and overclock the memory ...

Here I am sceptical. You increase the risk of bad data corrupting the computation. Requiring more restarts and perhaps decreasing overall throughput. And if you spend a week characterising it with tests then that is just a whole week of lost computation time that you now have to catch up over the next year.

[Prime95]

Originally Posted by retina

Yeah. I agree. This also applies to CPUs. There seems little point in overclocking a CPU when it can outpace the memory by a handy margin.Here I am sceptical. You increase the risk of bad data corrupting the computation. Requiring more restarts and perhaps decreasing overall throughput. And if you spend a week characterising it with tests then that is just a whole week of lost computation time that you now have to catch up over the next year.

I'm overclocking memory to 1170 and 1200. I get no errors at all. I will check it regularly and if errors begin (they wont corrupt the calculation) I will dial back.

[mackerel]

Originally Posted by retina

You increase the risk of bad data corrupting the computation. Requiring more restarts and perhaps decreasing overall throughput. And if you spend a week characterising it with tests then that is just a whole week of lost computation time that you now have to catch up over the next year.

Dialling in a rough overclock is hours at most. Find the point where it is obviously unstable, then back off a fair amount. The verification process of course can be much longer, but if you get it right you win from the start. If it errors you waste a little time but ideally you progressively test from shorter to longer units so more obvious faults will be detected faster and at less cost. You can also look at what other people have used to aid working out your running point. It takes a little digging around but some ram chips are known to be better or worse clockers than others.

One thing I both like and dislike about PrimeGrid is they double check all their work as they go along. So I use that as my soak test if I make any hardware changes. Try to pick tasks that are more of the magnitude of minutes or hours to start with, and work up. I don't overclock as such these days since power efficiency has gone way up my list of priorities, outside of competitive reasons anyway. Still, I've been thinking about undervolting if I ever get around to that, where similar applies. Still needs testing.

[lavalamp]

Originally Posted by mackerel

I don't overclock as such these days since power efficiency has gone way up my list of priorities

If it's power efficiency you're after, see my handy guide:

1) Don't buy CPU/GPU, wait until next generation
2) Go to step 1

Originally Posted by ET_

Waiting for the tests

http://lkml.iu.edu/hypermail/linux/kernel/1907.2/04769.html

[lavalamp]

Couldn't pop it in a windows machine for a quick benchy benchy?

Originally Posted by lavalamp

Couldn't pop it in a windows machine for a quick benchy benchy?

that needs an available Windows machine.


after summer vacations I think I can test


http://lkml.iu.edu/hypermail/linux/k...7.2/04769.html

[lavalamp]

So until then you have a GPU sitting in a box that you can't use?

Originally Posted by lavalamp

So until then you have a GPU sitting in a box that you can't use?

so what ?


I could use it with amdgpu-pro, after installing a machine, but it is vacation time, thus I am procrastinating.