DAG size in 2026, and which GPUs fall off next

Every few months, somewhere in the world, a rig that mined a coin perfectly well on Monday refuses to start on Tuesday. Nothing broke. The card simply ran out of memory for a file that grows on a schedule fixed in the protocol. That file is the DAG, and if you mine any Ethash-family or KAWPOW coin, its growth curve is a countdown clock attached to your hardware.

What the DAG is and why it grows

Memory-hard algorithms like Ethash, Etchash, Ubqhash and KAWPOW were designed to resist ASICs by making mining depend on fast random reads from a large dataset rather than raw computation. That dataset is the DAG, a pseudorandom file regenerated at fixed intervals called epochs. On Ethash-style chains an epoch lasts a set number of blocks, and each new epoch adds a fixed increment of size, roughly 8 MB per epoch, forever.

The whole DAG must sit in GPU memory. Not most of it, all of it, alongside a small buffer for the driver and framework overhead. The moment DAG size plus overhead exceeds usable VRAM, the card cannot allocate the dataset and hashrate goes to zero on that coin. There is no gradual slowdown at the cliff itself; participation is binary.

The history: 4GB, then 6GB

The 4GB story is the canonical example. Ethereum's DAG crossed the practical 4GB limit around epoch 373 in late 2020, and an enormous fleet of RX 470s, RX 480s and 4GB RX 570s dropped off the network within weeks. The same cards kept mining Ethereum Classic for years afterward, because ETC deliberately reset its DAG growth with the Etchash change in 2020, precisely to keep that hardware alive.

The lesson miners should have taken from 2020 is that the cliff is predictable to the day. Epoch length is protocol constant, size increment is protocol constant, so the epoch at which a DAG crosses any VRAM threshold is pure arithmetic. Nobody who lost their 4GB fleet was ambushed by mathematics; they were ambushed by not doing it.

By 2026 the same curve has come for larger cards. Ethereum Classic's post-reset DAG has been climbing for six years and now sits above the practical limit for many 4GB configurations again, with 6GB cards next in line on the horizon rather than in the distance. Ravencoin's KAWPOW DAG, which grows faster per unit time due to shorter effective epochs, crossed the 5GB mark and made 6GB cards the ones watching the calendar. The GTX 1060 6GB and RX 5500 XT class of hardware, long the workhorses of small rigs, are the cards to check this year, and 8GB cards inherit workhorse status with roughly a decade of margin on slower-growing chains.

The soft cliff before the hard cliff

VRAM capacity is the hard cutoff, but performance degrades before it. As the DAG approaches the memory ceiling, cards lose the headroom that drivers and mining software use for optimal memory access patterns. Some cards shed a few percent of hashrate in the final epochs before the cliff. Zombie modes, where miners page parts of the DAG over PCIe, can keep a technically dead card producing a fraction of its former hashrate, but on modern electricity prices a zombie card usually mines at a loss. Treat the soft cliff as the real deadline.

How to anticipate instead of react

This is exactly what the DAG size calculator exists for. It tracks the current epoch and DAG size for every supported coin and projects the date each VRAM class stops mining it. The workflow for a rig operator is simple and worth doing quarterly:

• Look up the current DAG size and epoch for every coin you mine.
• Note the projected cutoff date for your card's VRAM, minus a safety margin for driver overhead and the soft cliff.
• Decide in advance what the rig does after that date: switch to a coin with a smaller or reset DAG, switch algorithms entirely, or sell the hardware while it still has mining value.

That last point deserves emphasis. A GPU's resale value drops when it visibly falls off the largest coin it could mine. Selling six months before the cliff consistently beats selling six weeks after it.

Planning the next purchase

DAG growth also belongs in buying decisions. When comparing cards on /hardware, the VRAM number is not just a spec, it is a lifespan. A cheaper 8GB card and a pricier 12GB card can have identical hashrates today and very different useful lives on a growing chain. Run the profit numbers, then run the calendar.

The DAG does not negotiate and does not surprise. It publishes its entire future in the protocol. The only miners it hurts are the ones who never looked.