[astrodust]

Using a lot of memory complicates any ASIC design but it doesn't eliminate the possibility of such a thing being made.

If someone really knows their hardware engineering they could interface with GDDR5 memory on a custom board and mine as fast or faster than a GPU.

It's just a matter of how much money it will take to develop such a chip. I'd guess at least $2-5 million given how you'd need to fabricate it at 14nm or better to get sufficient performance, and that process can be a serious headache for smaller firms.

[yuhong]

Thinking about it, I wonder how much QDR SRAM bandwidth compare with GDDR5. This would be useful for Monero for example. Nobody is going to care about spending a few hundred for FPGAs with large amount of memory if a 5x-10x boost over GPUs is possible, but...

[astrodust]

There's a price trap here. The good FPGAs still cost $10K or more each and if you're intending to ship that hardware, which is how the Bitcoin and Litecoin ASIC platforms bootstrapped themselves, you're going to be in trouble if your FPGA solution can't beat the equivalent spend on GPUs. Bitcoin lends itself to FPGA acceleration extremely well, SHA256 is trivial to implement in an absurdly parallel way. Scrypt is a bit more messy but not impossibly hard. The Ethereum one is a beast by design, so it'll be a true challenge for any implementor.

Ten R580 cards can really crank out hashes for Ethereum. For a GPU to keep up it's going to have serious memory bandwidth issues as that's one of the limiting factors in this brand of mining.

What you might see is someone getting a license to make "OEM video cards" and then produce a line of mining-optimized cards. Given the constrained supply across the board on any AMD GPU this would be an easy win for a company like MSI or Gigabyte that's already making GPUs. Strip off the useless ports, tune the memory bandwidth, and make them work over USB-C instead of PCI-e so you can really pack a system full of these things.

Making an external GPU that comes in a nice housing with USB-C interconnect would also sell well in the machine learning market where you wouldn't have to worry so much about shoe-horning GPUs in your case. You could just stuff them in a rack.

[yuhong]

I am thinking of Monero and CryptoNight not Ethereum for FPGAs for that reason. I am thinking of ideally something like Virtex-7 or Vertex-8 with 32Mbit or more of memory.