How the Directed Acyclic Graph (DAG) Determines GPU Compatibility

Memory-hard Proof-of-Work algorithms—including Etchash (Ethereum Classic), KawPow (Ravencoin), FiroPow (Firo), Autolykos v2 (Ergo), and Octopus (Conflux)—are engineered to prevent cheap fixed-function ASICs from dominating consensus. They achieve this by requiring every graphics card to generate and hold a multi-gigabyte cryptographic lookup table called a Directed Acyclic Graph (DAG) directly inside high-speed GDDR6/GDDR6X/GDDR7 video memory (VRAM).

During every single hash attempt, the GPU core executes dozens of random 64-byte or 128-byte memory fetches across the entire DAG buffer. If a graphics card does not have enough free physical VRAM to store the current epoch's DAG file in full, it cannot mine that coin.

1. How DAG Epochs Grow Over Time

Unlike fixed-scratchpad algorithms, DAG-based blockchains systematically increase the size of the DAG file every epoch (a fixed number of blocks):

  • KawPow (Ravencoin) & FiroPow: Advance to a new epoch every 7,500 or 1,300 blocks, adding approximately 8 MB per epoch so the DAG grows steadily by roughly ~500 MB per year.
  • Etchash (Ethereum Classic): Following the Thanos upgrade (ECIP-1099), Etchash doubled its epoch length from 30,000 blocks to 60,000 blocks, slowing DAG growth so 4 GB and 6 GB GPUs remain usable for a longer horizon.
  • Octopus (Conflux): Uses a faster DAG growth schedule that requires 8 GB+ GPUs for sustained operation.
GPU Physical VRAM Tier Usable DAG Ceiling (Linux OS) Usable DAG Ceiling (Windows 10/11) Compatible Proof-of-Work Algorithms
4 GB VRAM (e.g. RX 570 4GB, GTX 1650) ~3.65 GB ~3.30 GB Etchash (ETC), Autolykos v2 (ERG), Blake3, kHeavyHash, Equihash 144,5
6 GB VRAM (e.g. RTX 2060, GTX 1660 Super) ~5.65 GB ~5.25 GB KawPow (RVN), FiroPow (FIRO), ProgPowZ, Etchash, Autolykos v2
8 GB VRAM (e.g. RTX 3070, RTX 4060 Ti, RX 6600 XT) ~7.65 GB ~7.20 GB Octopus (CFX), KawPow, FiroPow, Etchash, BeamHashIII, Cortex
10 GB – 12 GB VRAM (e.g. RTX 3080, RTX 4070 Super) 9.6 GB – 11.6 GB 9.1 GB – 11.1 GB All current DAG algorithms with multi-year epoch headroom
16 GB – 32 GB VRAM (e.g. RTX 4090, RTX 5090, RX 7900 XTX) 15.6 GB – 31.5 GB 15.0 GB – 31.0 GB All active GPU PoW algorithms, dual-mining & high-memory workloads

Why Operating System Overhead Reduces Usable VRAM

2. Linux Mining OS vs. Windows WDDM Display Reservation

A common frustration for beginners is seeing an out of memory or clCreateBuffer (-61) error when trying to mine a 3.5 GB DAG coin on a 4 GB card or a 5.4 GB DAG coin on a 6 GB card in Windows:

  • Windows 10/11 (WDDM Driver Model): Automatically pre-allocates 500 MB to 900 MB of VRAM on every GPU for desktop composition and DirectX buffers—leaving only ~7.20 GB usable on an 8 GB card. Connecting your monitor to the motherboard's integrated graphics (iGPU) frees 150–250 MB on GPU #0.
  • Linux Headless / Dedicated Mining OS: Reserves only 150 MB to 350 MB for CUDA/OpenCL runtime initialization, extending the productive lifespan of 6 GB and 8 GB GPUs by 6 to 12 months.
Automated GPU VRAM Compatibility Filtering on ZU Mining

When you select any graphics card in the ZU Mining GPU Profitability Calculator or inspect benchmarks in our Mining Hardware & Smart Rig Builder, our engine automatically compares the card's physical VRAM against live DAG requirements across all 92 coins in the Mineable Coins Directory. For tuning memory clocks once your DAG loads, see our GPU Undervolting & Memory Tuning Guide, download verified clients from the Mining Software Directory, and connect to low-latency servers in our Mining Pools Directory.