Chia Plotting on (13) 3090 GPU VRAM Raid 0 Stripped Set

MinerFortyNine · May 2, 2021

First of all, I am unaware if this is even possible.

One has to acquire (13) 3090s, an H110 Pro BTC Board, (13) Risers.

There is software that exists to access the VRAM from a single card; however, unaware of any for accessing it across multiple cards.

Example of Single Card

An NVME has an advantage in reading (and a real ramdisk blows reading away), a VRAM GPU drive has an advantage of writing (and pretty sure a real ramdrive would be better as well).

So the real question is can you make an array from GPU VRAM disks? No need to go to (13), as plotting is just a little bigger than 200GB (need to pay closer attention next time).

It would be kind of twisted for Chia to be most profitable from having GPUs.

One advantage of VRAM and RAM is it doesn’t degrade over time like NVME.

WereCatf · May 2, 2021

4 minutes ago, MinerFortyNine said:

One has to acquire (13) 3090s, an H110 Pro BTC Board, (13) Risers.

As far as I know, for Chia capacity is king, not read/write speed. Secondly, you'd get far more capacity for far less money if you just bought an Epyc and hundreds of gigabytes of RAM. You'd have to pay ten times the money if you went with a bunch of 3090's.

MinerFortyNine · May 2, 2021

22 minutes ago, WereCatf said:

As far as I know, for Chia capacity is king, not read/write speed. Secondly, you'd get far more capacity for far less money if you just bought an Epyc and hundreds of gigabytes of RAM. You'd have to pay ten times the money if you went with a bunch of 3090's.

Even if you have capacity, that capacity has to be filled with plots before you earn anything. A large array does not fill itself, one or more computers have to process data and make plots. Once its filled with 101GB files, then it has a chance to start earning. First 101 GB file is once in 9 years (at present network capacity), a second drops it to 5 years... a 72 TB array filled is about 5 days. That is why Chia is about space and time (takes time to fill it).

Loki0111 · May 2, 2021

If you are going to do something like this why not just load up on RAM instead and use a RAMdisk? Running it through GPU memory seems like an insanely expensive way to get minimal efficiency gains if any at all. In fact my guess is it will be significantly slower then a RAMdisk with high speed memory given you are piping it all through a riser constrained GPU PCIe interface.

Additionally you only need IO speed for plotting. Once the plot is done it goes to your farm (usually a massive array of spinning rust) where IO speed is almost completely irrelevant.

porina · May 2, 2021

5 hours ago, MinerFortyNine said:

First of all, I am unaware if this is even possible.

One has to acquire (13) 3090s, an H110 Pro BTC Board, (13) Risers.

There is software that exists to access the VRAM from a single card; however, unaware of any for accessing it across multiple cards.

Possible, maybe. Practical, nope.

VRAM has very high bandwidths, but only locally on the card. The second you go off card, you're limited by the much slower PCIe bus. If you're using 1x risers on a mobo like this, you're down to roughly 1GB/s per card. 13 cards, 13 GB/s if you raid 0 them, if we assume all that bandwidth can be used at once which is not a given depending on how that particular mobo lays out its lanes between CPU and chipset ones. Single channel system ram is much FASTER.

If you were to make the ultimate plotting machine, just fill it with fast ram instead.