1/ We've been working on Exactly that, for some time now. While the general idea is nice and correct...
Leveraging existing GPU mining infra to do ZK is in fact FAR from straight forward.
Time for a thread!🧵👇
Leveraging existing GPU mining infra to do ZK is in fact FAR from straight forward.
Time for a thread!🧵👇
https://twitter.com/pseudotheos/status/1587849391586430976
2/ GPU mining was dominated by #Ethereum mining which is now, post merge, no longer possible.
So we came up with the idea of helping the community repurpose their GPUs for ZK compute a few weeks ago...
So we came up with the idea of helping the community repurpose their GPUs for ZK compute a few weeks ago...
https://twitter.com/Ingo_zk/status/1571798998171656192
3/ ...and we got CRAZY demand.
Which took us completely by surprise! (surely, we thought, miners with tens of thousands of GPUs had a plan B for after the merge, right? well no.. but that's another story to tell)
Which took us completely by surprise! (surely, we thought, miners with tens of thousands of GPUs had a plan B for after the merge, right? well no.. but that's another story to tell)
4/ This motivated us to develop a #GPU based solution for running ZK.
We teamed up with @zksync, currently with the best ZK GPU implementation in the market
We teamed up with @zksync, currently with the best ZK GPU implementation in the market
https://twitter.com/Ingo_zk/status/1576145806440419328
6/ We quickly realized that miner setup, while optimal for mining, is FAR from optimal for ZK compute.
Basically, the only option for significant speed-up is...
Basically, the only option for significant speed-up is...
7/ ...moving as much as possible of the computation from running on #CPU to running on #GPU.
Otherwise, data transfers AND/OR a weak CPU simply kill the performance of the GPU.
Otherwise, data transfers AND/OR a weak CPU simply kill the performance of the GPU.
8/ Miners are using different kinds of GPUs. A popular chip is the @nvidia CMP series, which is very good for mining, but...
nvidia.com/en-us/cmp/
nvidia.com/en-us/cmp/
9/ A) CMPs are VERY far from the latest in heavy duty gaming chips, like the 4090 / 3090 series...
...or other advanced GPU chips by @nvidia that we demonstrated can handle the #ZKP workload extremely well
...or other advanced GPU chips by @nvidia that we demonstrated can handle the #ZKP workload extremely well
https://twitter.com/Ingo_zk/status/1562740123669651457
10/ B) @nvidia actively made the CMP chips sub-optimal for usage that isn't mining 🤷♂️
11/ We focused on @AleoHQ prover for Testnet3.
Main reason: it's the most simple yet practical ZK computation that exist today.
See this image for a breakdown of the different computational steps 👇
Main reason: it's the most simple yet practical ZK computation that exist today.
See this image for a breakdown of the different computational steps 👇
12/ Even for this simple design, which is far more basic than @Scroll_ZKP, or other zkRUs...
We identified challenges in how memory is being used, communication with the host, new primitives to be implemented like Blake Hash, and in tasks management for multiple GPUs
We identified challenges in how memory is being used, communication with the host, new primitives to be implemented like Blake Hash, and in tasks management for multiple GPUs
13/ Another challenge:
We worked on both @nvidia and @AMD GPUs.
The GPU miners are fragmented between these manufacturers, and there's a requirement of writing and maintaining new code for each
We worked on both @nvidia and @AMD GPUs.
The GPU miners are fragmented between these manufacturers, and there's a requirement of writing and maintaining new code for each
14/ Finally, if you are a miner with access to GPUs, or if you outside of #crypto but have access to a data center with GPUs...
We would love to help onboard you to the wonderful world of ZK compute.
DMs are open 💌
We would love to help onboard you to the wonderful world of ZK compute.
DMs are open 💌
• • •
Missing some Tweet in this thread? You can try to
force a refresh
