Ethash hashrate scaling on RX 470/570/480/580 and how to optimise core clock

[rikuu]

Hello everyone,

Despite GPU Ethash mining being a topic somewhat old, I wanted to know how it scales with core and memory frequency on AMD RX 500 cards.

As everyone knows, mining Ethash is all about memory, therefore once you know how fast your card can get memory-wise (this includes both clock and timings), the next step is to find the best core clock for your hashrate, in order to minimise voltage and power consumption.

I tested using phoenixminer with RX470s, RX570s and RX580s, both 4GB and 8GB versions with almost all memory models, and results were almost identical. Interestingly, RX580 hashrate limit at a given core clock was the same as RX570, which was a surprise. I was expecting them to attain higher hashrates at the same core clock due to having more GPU cores, but that was not the case.

This table intends to highlight the highest hashrate you can achieve at a given core clock. Do not focus on hashrates at each core/memory configuration, as your results will be different based on your memory brand and timings. You may need more or less memory clock to reach the hashrate limit at a specific core clock, but the hashrate limit at a given core clock will the same.


       


A good rule of thumb would be to use the following core clocks:

1050 Mhz up to 29.0 MH/s
1100 MHz up to 30.5 MH/s
1150 MHz up to 32.0 MH/s
1200 MHz up to 33.5 MH/s
1250 MHz up to 35.0 MH/s

To be precise, each +50 MHz core clock raises the ceiling by 1.4 MH/s.

[WhyMe]

Interesting, but what about power ?

[MATHReX]

There is also the factor of memory vendor in the card.
All the RX 570s and 580s that I bought came with Hynix memory and BIOS modding was a bit of challenge and risk at first but it did produce good results.
The highest hash rate I was able to get from the best card was 31 MH/s which was also unstable at times.

[rikuu]

Interesting, but what about power ?

Power consumption is a topic all on its own. Unfortunately, that is outside the scope of this post.

To elaborate my response: 2 GPUs running at 1200Mhz may require 850mv and 950mv respectively, resulting in very different power consumption. To make things even worse, those 2 GPUs running at 1100 MHz may require the same voltage. For this reason, it is really hard to talk about power consumption in an objective way, and I don't want to just throw some numbers that won't be true for everyone.

To complicate things even further, people measures power consumption in different ways (most of them being wrong). Some of them just look at what the software says. Others just look at the power consumption at the wall. Neither of those represents the actual power consumption of the GPU. In order to know what the GPU power consumption is, you have to measure the voltage and amps going into the GPU, both from the PCIe power connectors and the PCIe slot. If I present those numbers here, most people are not going to know how that translates to their way of measuring power consumption, therefore it will just be meaningless or highly misleading. Each one has do their own power measurements to find the best configuration for their situation.

There is also the factor of memory vendor in the card.
All the RX 570s and 580s that I bought came with Hynix memory and BIOS modding was a bit of challenge and risk at first but it did produce good results.
The highest hash rate I was able to get from the best card was 31 MH/s which was also unstable at times.

This is about how to adjust core frequency once you know how fast your GPU can get. All GPUs can achieve almost identical hashrates at the same core clocks. Memory type doesn't change those limits.

[fenomenyaa]

Good job !

[ummasek]

Gr8 JOB!
+50C Core = extra 1,4Mhash for me x7GPUs = +50Wat on 7gpu`s

[adaseb]

With my RX 470 4GB with 1100mhz and 1950mhz memory it got around 27MH/s or so. Definately not 30MH/s. Only way I could get 30MH/s was on 1150Mhz and 2000Mhz memory which seems more inline with your calculations.

However I didn't really put too much work in optimizing the cards, just did the same bios strap mod for all of them. This was during the 2017-2018 days when you didn't really care about power consumption because it was very profitable to mine. Most of my GPUs were mostly the Hynix memory types. Rarely did I get Samsung or Elpida memory.

Thanks for posting these findings, very interesting indeed.