Dear all,
Are there plans for developing AMD based miners?
I’m working on OpenCL miner for linux/windows/amd/nvidia, it already works on both amd/nvidia, but it is not fully optimized yet and it does not have the ability to connect to grin yet. So you need to wait a bit longer.
1 Like
Nice work man, i’d like to help out optimizing it. Or do you think you have it under control? also what timeline do you have in mind?
It will be open source, but it is a testing mess that even I won’t be able to understand tomorrow so better wait for the code that will go to github.
Too much ALU / memory ratio for AMD with present Siphash-2-4, nvidia handles it just fine. I’d have it out in a week, but I need to work and do other stuff, so timeline is: soon ™.
2 Likes
Hi, any updates on your amd Miner?
My miner is for cuckoo cycle, grin pow may not be cuckoo cycle.
Hello Photon!
Did you happen to finish this OpenCL miner for Cuckoo Cycle protocol? If you did, where can I find it?
Thanks.
I stopped all the work in August when Equihash replacement was announced. Now it looks like cuckaroo29 will hit main-net and there will be OpenCL miner, but probably in early January. Even grin-miner should have basic OpenCL support soon (for cuckatoo31) - thanks to @hashmap .
Thank you.
Yes, well, I was interested on mining AE which uses cuckoo cycle and mainnet was released yesterday. Thing is there is no OpenCL implementation, and someone on their forum pointed up to this forum saying there was someone developing one for this protocol.
That’s why I was reading a little bit here looking for this.
Thank you for your answer.
AE is business. They care about their bottom line not you mining on AMD. They gambled that somebody would write a (fee supported) pool for their project and somebody else would have a proper (fee supported) miner that connects to it. They were wrong.
You can only mine locally on the node by playing the lottery. No external miner can connect to the node even if it wanted to. Even if I had OpenCL code, there is nothing I can do.
anyone has tips on how to mine on amazon ec2? or perhaps buying a dedicated miner (hardware) like nvidia?
EC2 is crazy expensive. It only makes sense if you need GPUs for a short time frame, like training a deep learning network once. IIRC, something like 45 days of EC2 time is enough to have bought the GPU outright.
I’m writing a miner also, starting with Nvidia cards. I’ll port to AMD, but the miner is really complicated and the AMD version may not be ready for genesis.
2 Likes
I have tried, out of curiosity, EC2 (just the free test machine t2.micro).
However, after successfully compiling the grin miner, the command
grin-miner help
returns the following error:
thread ‘main’ panicked at ‘called Option::unwrap() on a None value’, libcore/option.rs:355:21
note: Run with RUST_BACKTRACE=1 for a backtrace.
Remark: the t2.micro has only 1Gb of RAM, so it was really just a test.
I think it would be interesting to make grin-miner work in Amazon p2.xlarge machines with NVIDIA K80 GPU with 12 Gb of ram.
Any ideas?
Anyone @photon and @makkiato interested or have done an AMD implementation that can be tested or interested to work on one before the livenet 15th Jan date ?
Why not ready for genesis?
I didn’t start until they revealed what the 90% PoW was going to be. I begged them to give developers more time, but it wasn’t announced until October I think. I’ve been working night and day since then, but I’m not sure I’ll even be ready with Nvidia by Genesis. My solver is considerably more complex than Tromp’s. Do not believe the claim that Cuckoo Cycle is simple. Far from it. It is easily more complex than any other PoW I’ve worked on, and the relatively simple public miner is not optimal.
1 Like
We don’t expect it to be.
How far from optimal were Equihash miners at time of launch?
We do hope that further Cuckaroo29 optimizations are less than 2x…
PR for Cuckatoo31 lean has been sent https://github.com/mimblewimble/grin-miner/pull/86 don’t expect too much from lean. I’m wrapping @photon OpenCL cuckaroo29 trimmer into Rust, there is a chance that mean cuckaroo29 will be ready before the launch
1 Like
You mean that the algorithm is not optimal and you found another one significantly better or that the generic implementation does not optimally run on specific HW? From profiling it would appear there are some options at the cost of very high complexity, but I would be surprised if 1080 Ti goes under 150ms.