Design study
The 12 nm study: why we went from 16 nm straight to 7 nm
Before committing a mask set, we studied whether 12 nm pays off for full‑custom hashing, with custom cells and ultra-low-voltage libraries. It did not, so we took our programs from 16 nm straight to 7 nm.
The study
One question, before the mask set
The question
Does a new node pay for itself in a full‑custom SHA-256 design, once the custom cells and the ultra-low-voltage libraries are counted?
The baseline
Our 16 nm design, where our full‑custom methods were first built.
What we studied
The node for our methods: custom cells, ultra-low-voltage libraries and full‑custom hashing cores.
The finding
12 nm does not pay off for full‑custom hashing with those methods.
The decision
Go from 16 nm straight to 7 nm, where our design has 300 hashing cores per die.
How we weigh a node
What a node study counts
The methods, not the node alone
A node pays only if our custom cells and libraries carry to it, so the study counts that work, not the node's headline figures.
Libraries at low voltage
Near-threshold hashing needs libraries characterised at low voltage: at 16 nm we re-characterised the foundry's libraries for near-threshold voltage.
Cost before commitment
The study comes before the mask set, when a change of course costs least.
Questions about node choice
Is 12 nm a good node for a Bitcoin ASIC?
Not for full‑custom hashing with custom cells and ultra-low-voltage libraries, in our study. We went from 16 nm straight to 7 nm instead.
Your Scrypt miners run on 12 nm chips. Does that contradict the study?
No. The study weighed one 12 nm FinFET process for full‑custom SHA-256 hashing, where custom cells and ultra-low-voltage libraries decide the result. The Scrypt ASICs in our miners were made on a different 12 nm process, 12LP, and a Scrypt chip is bound by its memory, not by those cells.
Can you run a node study for our chip?
Yes. A feasibility study weighs the algorithm, the node and the cost before you commit to a mask set.
Related pages
Choose the node before the mask set
Tell us the algorithm, your efficiency target and your schedule. We will study the nodes that could pay, and tell you which.
Our figures are our own chip-level measurements, or the specifications of our designs.