madMAx43v3r/chia-gigahorse v2.2.1.giga34

2.2.1.giga34 Gigahorse 3.0 Node / Farmer

on GitHub

New in this release:

• 4000 Series (Ada): Improved C29 farming compute by 20%, C30 by 10% and C31 by 5%
• 3000 Series (Ampere): Improved C29 farming compute by 7% and C30 by 5% (compared to giga33)
• 2000 Series (Turing): Improved C29 and C31 farming compute by 5% (compared to giga33)
• Older GPUs should also see improved performance for C29 to C31
• Potential fix for error GPU recompute failed with: an illegal memory access was encountered
• Chia GUI is now blocked from shutting down farmer on exit (need to exit twice to close GUI now)
• Because of that change: Need to chia stop all -d first before updating.

New since giga33:

• Chia version 2.2.1 update
• Fixed performance regression on 4000 series GPUs (as seen in giga32)
• This release does not implement CHIP-22 yet, however the upcoming GH Fast Farmer supports it.
• Both Node & Farmer & Harvester need to be updated at the same time to 2.2.1.giga33 (because the protocol changed).

New since giga32:

• Improved C29 and C30 farming compute on Turing (CC 7.5) and Ampere (CC 8.6) cards by up to 10% (depends on GPU)
• HTTP/1.1 based remote compute support (using port 11988 with chia_recompute_proxy_http, or any port XX988)
• Changed remote compute re-try interval from a fixed 100 sec to 30 sec per failed request. So for example if 10 requests were in-flight and the server crashes / the connection fails or they all timeout, it will wait 300 sec before trying that server again. This way the impact of a bad server is minimized.

Notes on C29:

• Plotting C29 is supported with latest cuda_plot_k32_v3 (3.1.0, see here)
• Farming C29 is supported since giga30

If you used chia_recompute_proxy before, you are encouraged to use chia_recompute_proxy_http as it's a more robust solution. It also allows servers to connect to the proxy by themselves via chia_recompute_server -u <proxy>.

chia_recompute_proxy_http is designed to work over the public internet, all connections are long lived. However you still need to implement access control by yourself, either via your own front-end proxy or some kind of WAF.

New since giga30:

• Gigahorse 3.0 plot format, C29 to C33
• Fixed an issue with remote compute where fail-over to another server did result in a lost lookup once every 100 sec. (ERROR: connect() failed with: Connection refused)

Limitations for GH 3.0 plots:

• Plotting is only supported with the CUDA plotter (cuda_plot_k32_v3, 16G RAM / 4G VRAM minimum)
• Farming is only supported with Nvidia GPUs (starting with 900 series)
• K32 only

Notes:

• Need to use cuda_plot_k32_v3 binary for plotting the new format (See here)
• Partial difficulty is now even more important for maximum farm size. Minimum recommended partial difficulty is 5000 for the new 3.0 format.
• Using a too low partial difficulty will increase GPU load by a factor of 2, in the worst case.
• Higher -S can help with phase 3 plotting times, like -S 8 if you have enough VRAM.
• Use ProofOfSpace farm --diff 800000 to benchmark for solo farming.
• chia_recompute_server is backwards compatible to old clients, but needs to be updated for new clients running giga30.
• Both Node & Farmer & Harvester need to be updated at the same time to support 3.0 format.

This release provides Chia Blockchain binaries to farm compressed plots created with the new plotters provided in this repository.

The compressed plot harvester and farmer are not compatible with the official Chia node, it only works together with the Gigahorse node. However it's possible to use a wallet from the official Chia repository, instead of the Gigahorse binary wallet.

Both NFT and OG plots are supported, as well as solo and pool farming (via the official pool protocol). Regular uncompressed plots are supported as well, so you can use the Gigahorse version while re-plotting your farm.

The dev fee is as follows:

• 3.125 % when using GPU(s) to farm compressed plots
• 1.562 % when using CPU(s) to farm compressed plots
• 0 % for regular uncompressed plots

The following packages are needed to run on Linux: sudo apt install ocl-icd-libopencl1

Make sure to choose the correct version for your system, see uname -m, usually it's x86_64. aarch64 is the same as arm64.

On Windows:

• You might have to install latest Microsoft Visual C++ Redistributable: https://aka.ms/vs/17/release/vc_redist.x64.exe

Note: See limit-gpu-usage on how to select an OpenCL platform if you have multiple platforms (excluding Nvidia).