Build Advice Help with choosing components for a multi GPU build ?

[ElmoPA]

I currently have more than twenty RTX3070 GPUs in my house due to cryptomining. I want to repurpose them as much as possible to test Apache Spark and Kafka. However, to use many GPUs from my understanding you need a threadripper. However, I found the price of the recent motherboards and threadrippers to be extremely expensive, so I would like to ask if there are older models you would recommend. My current idea is to get something like M12SWA-TF and AMD Ryzen™ Threadripper™ PRO 5955WX Processor. I think server components also seems to work but I'm totally new to this space so I have no idea on what to choose. Also I think 128GB of RAM would be great.

 

[Eximo]

I guess the main question is what are you going to do with this multi-GPU build?

Aside from rendering, possibly deep learning and AI, not much you can do with multiple GPUs. And depending on the type of work, they don't necessarily need high PCIe lane counts, just like crypto mining. You might just try out your existing rig if it is still viable.

Threadripper and other more enthusiast class boards support PCIe lane bifurcation better than consumer boards. Threadripper is basically 'server' chips for the consumer, so probably your most cost effective choice. With the latest threadripper cheapest chip being like $1400, not very economical, also not quite launched yet.

Intel doesn't really have an equivalent, but you can look at the old 10th Gen X class chips and LGA2066 boards. Might be able to find a setup fairly cheap.

 

[ElmoPA]

I am planning to use it for deep learning. Learning how to distribute models using Apache Spark and Kafka that is why I specifically want to utilize as many GPU as possible. That the way to learn how to utilize Apache Spark effectively.

Recently I found that EPYC 7642 and EPYC 7402 seems to be viable options. They cost 1100 and 1700 dollars respectively which includes RAMs and motherboard. What are your thoughts on this? I am just not sure if they are compatible.

 

[Lucky_SLS]

For reference:

Getting-Started

Spark RAPIDS plugin - accelerate Apache Spark with GPUs

nvidia.github.io

and it seems Kafka is mainly storage throughput dependednt?

The problem with multiple 3070 is that the bandwidth between the cards is not that great. Which would be the main bottleneck here. Unlike professional quadro RTX cards which use latest NVlinks for 2tb/s bandwidth, the 3070 would be a quarter of that. So multiple GPUs might not even scale lineraly. Do some digging and decide for yourself.

And like Eximo said, your mining rig can be good starting point. try it out before you spend some big cash on costly parts.

 

[ElmoPA]

I see that makes very much sense. I will try to research some more to check if the card inherent bandwidth will result in a bottleneck. I think I have to study how the workload is distributed in much more detail. Thank you very much for your advice!