Dear Software Devs: Possible 4GB DAG file workaround

[TheBottomTurtle]

Actually pulling this off is far beyond my technical capabilities, but I was thinking it might be possible to use two 4GB cards working in tandem to simulate an 8GB card.  If it's reasonable to split the DAG into two parts, each half could be loaded into a separate card.  Both cards could begin hashing together, passing control of the function back and forth between them as elements are selected from the portion of the DAG loaded in the other card.  As I said, the technical details of how Ethash works are outside my understanding, but as current mining software seems to lean heavily on pagefiles and does use system RAM could a pool of pointers be shared between two cards allowing them to work together like this?

I understand this would half the overall hashrate, perhaps even worse due to increased overhead, but I think it's safe to say those of us with a lot of 4GB cards would prefer that to having a bunch of paperweights.

If my suggestion is completely absurd feel free to flame and carry on with your life.  My only request is that your insults are creative or clever  

-TBT

[buraz11]

Its so quiet that i think you actually made some devs think, i have the same problem here. I thought about reaching the guys doing the teamredminer but i dont know as im just a newbie here...time will tell

[anamichii]

good idea, but maybe need another software / or another hardware to fusion 2 card

[DrG]

The time to access the GPU VRAM vs time to talk to the other card across the BUS alone are hugely different. The only purpose of 4GB cards will be to find some other coin or to play games. Dump them now while prices are high due to shortages. Covid was like a free gift to anybody holding 4GB cards.

[rikuu]

I agree with DrG. Unfortunately, if GPUs have to talk to each other, mining speed would be incredibly slow. it is not viable.

[JayDDee]

I'll play along.

If the file can be split the excess could be stored in main mem. Only accesses to the excess
part of the file would be penalized.

[TheBottomTurtle]

I see the point being made.  I thought it might be possible to avoid the cards needing to talk to each other by sharing access to pointers.  The hashing could be done on two cards in parallel using the same pointers to working memory.  I'm not familiar with what is loaded into the GPUs themselves though and what the system RAM is used for.

If the file can be split the excess could be stored in main mem. Only accesses to the excess
part of the file would be penalized.

That's an interesting idea I didn't consider though.  I'm not sure what sort of numbers we'd be talking about but if you could load 90% of the DAG into the VRAM and only be penalized when accessing the missing 10% in the system RAM you might end up with something viable.

[mnhoster]

I tried that over and over but it exceeds me. I am not even sure if the result will satisfy us if someone manages to do this.

If it was possible to transfer little part of the DAG file to ram 

GPU    4 - 4 - 4 - 4
RAM    2 - 2 - 2 - 2
TOTAL 6 - 6 - 6 - 6

[buraz11]

If im not wrong the teamredminer devs already did some kind of split dag thing because some drivers dont support full ram allocation or something

[astraleureka]

Its so quiet that i think you actually made some devs think, i have the same problem here. I thought about reaching the guys doing the teamredminer but i dont know as im just a newbie here...time will tell

it's quiet because it's illogical and detached from reality

[buraz11]

you could be right, maybe something happens, i think there is a slight possibility one of the programming geniuses makes a solution if not nobody is going to die life goes on

[adaseb]

It's a clever idea but if it was actually possible then something would of been developed about 5 years back so the 1gb GPUs would be able to mine. And there actually was a type of beta software miner 5 years ago which would of made the 1gb mine but it was scrapped because it achieved very slow speeds.

Two GPUs communicating with each other will be too slow because the bus speed would slow them down. GDDR is very fast ram but only on the GPU itself and nowhere else.