X12 Port PCI-2.0 GPU Expasion (24 to 28 GPUs Per Motherboard)

Well, one of my other project is going nowhere for now (working on a consumer friendly Bitcoin miner); so, I decided to take a break and do something else. After a lot of searching and nagging emails, I finally scored a PCI Express ASIC chip for a descent price. This chip has 16 ports and can be configured in different modes. The initial board will be configured in x4 (Upstream/PC) and 12 x1 Downstream to 12 USB3.0 ports for compatibility with the popular GPU risers everyone is familiar with.

Would like to gauge interest. It's an ambitious project; it really needs a market that can absorb a 1000 of these cards in order for it to be much more economical than the alternative (more motherboards, ram, CPU, etc.).

@1000, the current "High" price estimate for the end consumer would be around $138.00. Either way, I'm still going to engineer it, but will be checking back before putting in a bulk order. If there is anyone that can help with driver support, that would be very nice.

Thanks for stopping by!

1 Like

If I follow you correctly, you'd like to build one of these but much much larger?
I think we'd all be interested in something like that, the question is whether the ASIC chip is doing most of the lane management (acting as a switch) or are the 24 to 28 GPUs something the CPU+Chipset has to have enough resources for? Either way, I'm interested, keep us in the loop.

moarpower,

Yes, like that; however, there will only be one PCB that slides into a 168 pin PCI-Express socket and has all the USB 3.0 Type A Female Sockets. It will be much more friendly to making nice rigs.

The PCI-Express 2.0 Switch ASIC Chip will handle pretty much everything. Other than initialization, the PC will not need to do much.

I'm interested in the photo you posted, Who makes that one? Do you have a link?

Sintech.cn makes that one: http://www.sintech.cn/riser%20card/ST-PCIE4PCIE%20PCI-E%20X1%20to%204%20PCIe%20x1%20riser%20cable.html

Here's another you may find interesting: http://amfeltec.com/products/flexible-x4-pci-express-4-way-splitter/

And one more, different product but same maker as above. Lets you connect up to 16 GPUs per available PCI-e slot: http://amfeltec.com/products/gpu-oriented-cluster (be sure to watch their video).

Considering there are no OS drivers that will handle more than 8 GPUs, I'm curious what the goal is here?

Hi Dlehenmy,

Thanks for the comment.

We are looking for engineering expertise to adress this dilemna. Do you
think you could help?

Thanks,

Well, personally, I can’t see any reason to put large numbers of GPUs on one system. It creates a much more costly single point of failure. MB, memory, cpu, PSU, or storage failure takes down 2 dozen GPUs? One GPU has hw issues and it takes down the other 23? No thanks. After hash rate, THE most important attribute of a mining rig is stability/uptime. A rig that large is highly unlikely to exhibit that vital quality. In my opinion, you’re trying to create a solution to a problem that doesn’t exist. The system you’re looking to build will become very complicated, as you try to configure multiple PSUs to power all the cards, etc. Also, even if you had a driver that could handle 24 GPUs, you’d have to have a multi-core/multiple thread CPU to handle the driver concurrency necessary to efficiently drive that many. I’m sure it’s an interesting problem to think about, but I don’t see it as the least bit practical or desirable in a real-world production mining operation. Of course, that’s just my take on it. I have a medium sized farm, and I wouldn’t touch a rig like that. I wouldn’t even build an 8 GPU rig. The only 8 GPU rig I’ve seen that I’d consider is the PandaMiner, since it’s a integrated design that was intended to run 8 GPUs reliably from the get-go.