Radeon VII on a mining-like bench
 

A few minutes ago, I have successfully installed Radeon VII on a handmade cardboard bench (see the picture in the next post). It is running nicely, only a tad slower than directly in the motherboard, with good temps because it's now on open air.

AliExpress can do wonders :smile:

Pictures

I was told I should post numbers, so here they are:

Exponent = 106M

In motherboard in case:
960 us/it
GEC 0.45 s
95 °C

On bench:
985 us/it
GEC 0.51 s
80 °C

[QUOTE=Viliam Furik;569118]I was told I should post numbers, so here they are:

Exponent = 106M
960 us/it[/QUOTE]

Time to install Linux. Your goal should be in the neighborhood of 700us

[QUOTE=Prime95;569123][B]Time to install Linux[/B]. Your goal should be in the neighborhood of 700us[/QUOTE]


Install Linux for a gain of 260 microseconds? Unless a person cut their OS teeth on Linux, there is not much to gain.

:two cents:

[QUOTE=Prime95;569123]Time to install Linux. Your goal should be in the neighborhood of 700us[/QUOTE]

If it's possible to run it through a VM, I might give it a try someday.

260/960 = 27% faster. I call that pretty significant.

I'm curious - does the same applies to PRP? Namely, are there any notable performance gains from using mprime instead of prime95 on LL/PRP tests?

[QUOTE=dcheuk;569132]I'm curious - does the same applies to PRP? Namely, are there any notable performance gains from using mprime instead of prime95 on LL/PRP tests?[/QUOTE]

I have never noticed much speed difference when using GWNUM programs under different operating systems.

There is a few percent to gain in throughput by running two instances of gpuOwl.

[QUOTE=paulunderwood;569133]I have never noticed much speed difference when using GWNUM programs under different operating systems.

There is a few percent to gain in throughput by running two instances of gpuOwl.[/QUOTE]

Thanks for letting us know. :smile:

[QUOTE=storm5510;569125]Install Linux for a gain of 260 microseconds? Unless a person cut their OS teeth on Linux, there is not much to gain.
[/QUOTE]

[QUOTE=Prime95;569129]260/960 = 27% faster. I call that pretty significant.[/QUOTE]

960/700 = 1.37 which means 37% faster (more thruput).

[QUOTE=axn;569140]960/700 = 1.37 which means 37% faster (more thruput).[/QUOTE]
:shock: How do you figure?

960/700 = 1.37, or 37% [U][B]slower[/B][/U]. ("he runs 37% slower than he could")

700/960 = 0.73, or 27% faster. ("he can run 27% faster than he does")
:razz:
If you trade forex and lose 33% of your account (i.e. you had $100 and you have now $67), then you need to make 50% profit, to come back to your initial sum ($100).

The confusion is because of the silly (IMO) choice of inverted figures used.

960 us/it ~= 1042 it/s
700 us/it ~= 1429 it/s

So from 1042 it to 1429 it (per s) is 1429/1042 ~= 1.371429...

Therefore ~37% faster

[QUOTE=LaurV;569145]:shock: How do you figure?[/QUOTE]

As retina correctly pointed out, you should use the thruput figure (iters/s) rather than time (us/iter) to do the calculation. Hence, it is "27% slower" / "37% faster".

[QUOTE=Viliam Furik;569128]If it's possible to run it through a VM, I might give it a try someday.[/QUOTE]

In a version of Windows 10 due to be released this year, you will be able to run Linux on WSL2, Windows Subsystem for Linux V2, which is what you want. It's available now, but only in the pre release "preview" version ( beta ) of Windows 10. It will probably be able to run CUDA, but none of that is completely certain. Whether it will be worth it or not remains to be seen. My impression from the previous posts here in several threads is that a lot of the Linux speed is derived from the special AMD Linux video driver "Rocm". The arrangement of device drivers on WSL2 is uncertain. If WSL2 just passes the CUDA requests thru to the windows video device driver, there may be little or no gain.

[QUOTE=storm5510;569125]Install Linux for a gain of 260 microseconds? Unless a person cut their OS teeth on Linux, there is not much to gain.

:two cents:[/QUOTE]
This comment hurts my soul. Three R7's on Linux are roughly equivalent to four R7's on Windows. Even an absolute noob can get Linux+ROCm+gpuowl running from scratch within a few hours by following a short list of instructions.

[QUOTE=M344587487;569177]This comment hurts my soul. Three R7's on Linux are roughly equivalent to four R7's on Windows. Even an absolute noob can get Linux+ROCm+gpuowl running from scratch within a few hours by following a short list of instructions.[/QUOTE]

No disparagement was intended. I have an [I]Ubuntu on Windows[/I] setup for 20.04 LTS on Windows 10 Pro v2004. Beyond running [I]mprime[/I], I have not been able to do much with it. I have been a M$ user since 1988. IBM XT's, no HD's, 5 1/4" floppies, and green monochrome to look at. Personally, I tend to go with what I know.

I tried he math in micro form. 0.00070 / 0.00096 which gave me 0.72916. Doing a reciprocal gives 1.37142. The fractional part indicates 37.142%. I [U]did not[/U] think this was correct. By looking at what is above, it is correct. I supposed I fooled myself with the decimal point and all the zero's. A 37% increase is considerable.

[QUOTE=axn;569148]As retina correctly pointed out, you should use the thruput figure (iters/s) rather than time (us/iter) to do the calculation. Hence, it is "27% slower" / "37% faster".[/QUOTE]
Hmmm... Makes sense[SUP](TM)[/SUP]. :whistle: