| # Honest limits — read before you rely on DaisyChain |
|
|
| DaisyChain is a real tool for a **specific** job: chaining spare/old machines to |
| train **small** models faster by pooling their compute. It is not magic, and it |
| is easy to misapply. Here's the honest envelope. |
|
|
| ## What it does |
| - **Data-parallel** training: every node holds a full copy of the model and |
| trains on its slice of the data; gradients are averaged across the network. |
| - **Capacity-weighted**: each node measures its own speed (CPU or GPU) and takes |
| a proportional batch, so a faster machine does more. GPU nodes auto-join and |
| carry more load; CPU-only nodes still participate. |
|
|
| ## What it does NOT do (expect these) |
| - ❌ **Not GPU-class training.** N pooled old CPUs/GPUs are still that class of |
| hardware. A single modern GPU will beat the whole cluster for real models. |
| - ❌ **Pools compute, not memory.** Every node needs the *whole* model in RAM |
| (or VRAM). You **cannot** train a model bigger than a single node can hold. |
| Chaining 5 laptops does not give you one big machine. |
| - ❌ **Sublinear scaling.** Gradient sync over WiFi is bandwidth-bound, and the |
| **slowest node paces every step** (synchronous). Each added node helps less |
| than the last; a slow link can erase the benefit. |
| - ❌ **Not for huge/modern models.** Transformers/large CNNs at CPU (or old-GPU) |
| speed are impractically slow. |
|
|
| ## The sweet spot |
| Small models (small MLPs, tabular/classifier tasks, small fine-tunes), |
| a handful of machines, model fits on each, you want more throughput and don't |
| have a GPU handy. Education, research, retro/old-hardware hobby clusters. |
|
|
| ## Hardware / OS |
| - **Multi-node reliably runs on Linux/macOS.** On **Windows**, gloo tensor |
| collectives are unstable — use **Docker** or **WSL**. Single-machine use is |
| fine on Windows. |
| - **Old GPUs** must be new enough for current PyTorch (compute capability ≳ 5.0, |
| roughly GTX 900 / Maxwell and up). Kepler/Fermi-era cards won't work with |
| modern torch. |
| - Put all nodes on the same **interface** (Tailscale `tailscale0`, or a real LAN |
| NIC) via `GLOO_SOCKET_IFNAME` — not a VPN/Docker/WSL virtual NIC. |
|
|
| ## The one-line version |
| You can chain old machines to train a **small** model together, faster by |
| throughput, with honest sublinear scaling — and DaisyChain won't pretend to be |
| more than that. |
|
|
