[bitcoin-dev] About hardware accelerators advantages for full-node (not mining)

[bitcoin-dev] About hardware accelerators advantages for full-node (not mining)

[Alex Barcelo]

[-- Attachment #1: Type: text/plain, Size: 1080 bytes --]

I am searching for guidance and opinion in the subject's matter. I will begin with my use case, too see whether my ideas makes sense or not.

I have a Jetson TK1[1], which is a GPU (CUDA) powered development board. I thought that it may be a power-efficient device (in bitcoin environment), and thought about having it as a full-node. Either as a public full-node, if it makes sense, or a local full-node, to allow my PCs to perform relay onto the bitcoind in the Jetson. My idea is to run a bitcoind daemon on the Jetson as a node with high performance-per-watt (also cheap and repurposable). A pure-CPU implementation of bitcoind will clog the CPU

I assume that there are a bunch of heavy-compute highly-parallel functions which could be "outsourced" to a GPU. I may want to fork and/or contribute on that. However, maybe I am speaking nonsense. I have more background on parallel programming than my knowledge on bitcoin protocol. So, before coding a complete mess, I wanted to hear some opinions on the idea/configuration.

[1] https://developer.nvidia.com/jetson-tk1

[-- Attachment #2: Type: text/html, Size: 1320 bytes --]

Re: [bitcoin-dev] About hardware accelerators advantages for full-node (not mining)

[Tier Nolan]

[-- Attachment #1: Type: text/plain, Size: 795 bytes --]

The main processor load is for signature verification.

This requires a hash function call, some large number maths and an elliptic
curve operation.  The elliptic curve stuff is the longest step.

This takes around 1ms per signature on normal hardware, but optimized code
is faster.

The main task is to prove that

R = u1 * G + u2 * Q

G is a constant, and the rest are different per signature.

Some of the core team have created a fast CPU implementation.

 https://github.com/bitcoin/secp256k1.git

There was talk of batch verification of signatures.  The process might take
16 signatures and compute them together.

I think a lot of the benefit of the GPU would be lost due to communication
bandwidth.  GPU miners benefit from needing very little information to be
sent to the GPU routine.

[-- Attachment #2: Type: text/html, Size: 1056 bytes --]