Ethereum + OpenCL Benchmarks With The Latest AMDGPU-PRO Mining & NVIDIA Linux Drivers

[starshipeleven]

Depends on your definition of "fantastic". It's just hardware being used as it should be (e.g. everybody knows AMD consumer hardware is better at number crunching than Nvidia, but their driver sucked till now). If you're an AMD owner, that can look fantastic, but if you're in the green camp, it's just "things works as they should have from the beginning".

(said with an aloof voice, while sipping some expensive wine)

[GreatEmerald]

Huh, why no pre-post comparison? Does this beta driver help any or not, and which cards? And I wonder why F@H results are like that, maybe the driver is limiting the performance in this case... Or maybe it's optimised for CUDA?

(said with an aloof voice, while sipping some expensive wine)

Good show!

[bridgman]

On Vega, I believe it takes performance from ~5.5MH/s to ~30 MH/s.

[coder]

On Vega, I believe it takes performance from ~5.5MH/s to ~30 MH/s.

Congrats! This is awesome (not specifically the Hash rate, but in general). I'll probably jump on a Vega 56, once the aftermarket coolers start to appear.

I just hope some flavor of Vega 11 is available in a half-height, half-length card. My employer sells cost-constrained 2U systems that don't contain a riser, limiting us to this form factor. We use GPUs for compute, so something like the RX 550 isn't very attractive.

[illwieckz]

While reading the article I was hoping for some darktable benchmarks. Well, in fact I don't need nvidia vs amd benchmark since buying nvidia is not among the available options, what is interesting to know is how this compute-aimed driver performs with some various OpenCL tasks compared to generic driver. Basically, darktable is my daily opencl usage. The luxmark performance amazed me though, but I'm not using it every day.

So Michael if one day you have the idea to do a before/after amd benchmark with this driver, it would be very good for readers if you include the most possible opencl benchmark so people know if this driver is good for their own task. Also I would be curious to know how bad this driver performs in games. We all know this driver is not for games, but depending of the average loss, people doing both gaming and working on the same rig can decide if the loss in games is acceptable or not compared to the improvement in opencl tasks.

For example on my own usage, I can play all my games at full resolution and quality even when my GPU profile is set to low battery (dividing shader frequency by 4 and memory frequency by 10), I only set it to high performance for compute tasks, so even losing 70% of the gaming performance will not prevent my games to running smoother (I would just use the performance profile for gaming instead of the severely underclocked one), and I'm OK with that if it can gives at least 20% more on compute power for example. And we know some compute tasks got a 600% boost…

[Michael]

While reading the article I was hoping for some darktable benchmarks. Well, in fact I don't need nvidia vs amd benchmark since buying nvidia is not among the available options, what is interesting to know is how this compute-aimed driver performs with some various OpenCL tasks compared to generic driver. Basically, darktable is my daily opencl usage. The luxmark performance amazed me though, but I'm not using it every day.

So Michael if one day you have the idea to do a before/after amd benchmark with this driver, it would be very good for readers if you include the most possible opencl benchmark so people know if this driver is good for their own task. Also I would be curious to know how bad this driver performs in games. We all know this driver is not for games, but depending of the average loss, people doing both gaming and working on the same rig can decide if the loss in games is acceptable or not compared to the improvement in opencl tasks.

For example on my own usage, I can play all my games at full resolution and quality even when my GPU profile is set to low battery (dividing shader frequency by 4 and memory frequency by 10), I only set it to high performance for compute tasks, so even losing 70% of the gaming performance will not prevent my games to running smoother (I would just use the performance profile for gaming instead of the severely underclocked one), and I'm OK with that if it can gives at least 20% more on compute power for example. And we know some compute tasks got a 600% boost…

I normally include Darktable but it was failing to run with AMDGPU-PRO this time around and didn't have time to investigate.

[coder]

i think testing 1 card does not reflect the mining usecase because miners always use many cards and the performance goes down because of a lot of heat and multible cards block each others cooling air.

I think most serious miners are using open cases. Some even use PCIe extension boards/cables to increase GPU spacing.

[coder]

coder open cases have wirst airflow controll and no pressur-in from external Fans in the airflow...

Open cases result in lower GPU temps. There's plenty of experimental evidence available, if you care to search for it.

[USER="92240"]So your open-case just turns the case Problem in a airflow controll Problem and room Problem

Assuming your closed case exhausts into the room, the only way that open cases can worsen room heating is by allowing GPUs to run faster, due to cooling them better.

Again, you can find plenty of pics where miners are using open cases. Unlike datacenters, these guys prioritize low hardware cost over high reliability. So, they go for the cheapest cooling solution that gets the job done.

[coder]

server-farms/datacenters use the air-pipe control Principe with strict distinction between the cool air as income and hot air as output.
and they do it because of higher performance with air pipe control principe... not because "cheapest cooling solution"

Either you're changing the subject because you know I'm right, or you misread my earlier statement. I was taking about miners using cheap solutions - NOT datacenters. Datacenters care about reliability and density to the point that they'll spend more money on custom rooms and expensive chassis than most miners will.

Miners know that the gig could end with the next big crash, so they have less incentive to invest in fancy long-term solutions than datacenters.

if you try to put 100 vega-64 cards in a normal lifing room with a normal 100cm*100cm window you have NO chance whatsoever to cool the cards.

Miners with that many cards are not doing this in such rooms. They tend to rent out cheap industrial space. I doubt you could get away with using so much power, in a residential setting.

[en97]

Is there a guide on how you guys were able to run AMD and NVIDIA cards at the same time?