Fast and efficient plotting - creating plots using the GPU

For some time now, the Chia Network community has been experimenting with plotting using graphics cards. It turns out that creating plots in this way is not only fast but also (contrary to appearances) energy-efficient. Top graphics cards based on the PCIe 4.0 x16 bus can generate a single plot in 80 - 90 seconds. This is only half the time of the fastest recorded plot creation time using a processor. These are exciting news because graphics cards are more accessible and significantly cheaper than top-of-the-line server equipment.

This achievement lowers the entry threshold for Chia farmers as plotting is one of the most challenging tasks in the whole process. Creating plots using graphics cards allows reducing power consumption by up to three times. The temporary memory requirement for a k=32 plot remains unchanged. For the first phase of generating uncompressed plots, about 256 GB will be needed.

Most GPU-based plotters work similarly to those based on processors. The table is divided into smaller parts, which then go into separate buckets. Each bucket is then passed to the GPU, which performs phase-specific sorting, matching, and compression operations. The current implementation of this process is highly efficient for modern graphics cards. The limitations may only be GPU performance and PCIe or memory bandwidth. If the miner has a limited amount of RAM, additional temporary files need to be created. This process generates more disk operations, thereby slowing down plotting up to three times. Even considering this limitation, generating plots using GPUs will be significantly faster than using CPUs.

GPU Performance

Graphics cards gain performance with each successive generation. This is due to the increasing number of GPU computing units and silicon energy efficiency. The latest Nvidia graphics cards (RTX 4090) have almost twice the computational power of their predecessors (RTX 3090). The cumulative annual growth rate of Nvidia cards' performance from 2016 to 2022 is 25%. The following chart shows the growth of CUDA core performance over the years.

Source: https://bit.ly/3jAm3QI

PCI Express (PCIe) is the dominant general-purpose interface on the market. This interface allows connecting fast components to the computer's motherboard, such as computing units, data storage, or network cards. In the past, SSD drives transitioned to the PCIe bus. Currently, the same can be observed with the CXL standard. It allows connecting memory expansions, accelerators, AI controllers, and 400 Gbit/s network cards. The PCI specification community estimates that its bandwidth will double every three years. Currently, PCIe 5.0 devices are appearing on the market, offering a two-way transmission of 64 GB/s for an x16 connector. The following chart presents the throughput of successive PCI Express versions.

Source: https://bit.ly/3jAm3QI

Cheaper and More Efficient Plotting

Thanks to graphics cards, plotting is almost three times more efficient than using processors. This means that farmers can plot their drives faster while reducing the amount of consumed energy. The efficiency of plot generation is measured in kilowatts per plotted terabyte. The following chart shows the power consumption of the graphics card during plotting.

Source: https://bit.ly/3jAm3QI

Summary

Plotting with GPUs is significantly faster and more efficient than using CPUs. This approach reduces power consumption by nearly three times and significantly shortens plot generation time. This, in turn, lowers the entry threshold into the Chia ecosystem and positively impacts decentralization.

The new plotting technology is also good news for miners who had to give up mining based on the Proof of Work method. Unused, problematic-to-sell graphics cards can now be reused for the needs of the Chia blockchain!

back ››