Skip to content
This repository has been archived by the owner on May 18, 2023. It is now read-only.

Speculation: Memory temperature issues in 3080/3090 are increased by physically contiguous DAG file loading in GDDR6X chips, distributing the DAG file would give thermal improvements #10

Open
CharlesHe16 opened this issue Feb 6, 2021 · 0 comments

Comments

@CharlesHe16
Copy link

I have a guess that the memory chip use of some Ethash mining implementations is uneven, constantly using the same subset of chips on the GPU.

For example, memory chips on one side of my 3080 GPU runs far hotter than the other side. (Note that I am not the first to notice this, but I cannot immediately find the other person's content).

Maybe this could be caused by Ethash implementations like T-Rex loading the DAG file (currently ~4 GB in size) into physically contiguous chips that are beside each other on the card. In the 3080 GPU, the DAG would fill about half the ten 1GB chips, which lines up with what I’ve seen.

If this speculation is true, the thermal problems caused by this unbalanced memory use could be large and spreading out the load would be beneficial.

For example, the 3090 can spread out loads and cooling over 5 times more space because it has 24 GB VRAM.

As you know, high memory temperatures are a big issue in the 3080/3090 cards. Thermal throttling due to memory temperature often limits hash rate, and there are concerns about wear and tear on the chips and on fans. To address this, people have spent a lot of money and time on various thermal pads and other cooling modifications.

Can T-Rex or another miner create an Ethash implementation that distributes the load across the chips?

Not a hardware engineer here, but for example, you could insert space to expand the DAG file across the memory chips, or just have two instances with two DAG files, and alternate work between them.

It seems that a miner that implements this would have a huge advantage over others, have higher market share, charge higher fees, etc.

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant