ZmnSCPxj [ARCHIVE] on Nostr: 📅 Original date posted:2021-05-18 📝 Original message:Good morning Michael, > ...
📅 Original date posted:2021-05-18
📝 Original message:Good morning Michael,
> That’s a fair point about patents. However, note that we were careful about this. oPoW only uses SHA3 (can be replaced with SHA256 in principle as well) and low precision linear matrix multiplication. A whole industry is trying to accelerate 8-bit linear matrix mults for AI so there is already a massive incentive (and has been for decades).
>
> See companies like Mythic, Groq, Tesla (FSD computer), google TPU and so on for electronic versions of this. Several of the optical ones are mentioned in the BIP (e.g. Lightmatter)
Please note that ASICBOOST for SHA256d is based on a layer-crossing violation: SHA256 processes in blocks, and the Bitcoin block header is slightly larger than one SHA256 block.
Adding more to a direct SHA3 (which, as a "sponge" construction, avoids blocks, but other layer-crossing violations may still exist) still risks layer violations that might introduce hidden optimizations.
Or more succinctly;
* Just because the components have (with high probability) no more possible optimizations, does not mean that the construction *as a whole* has no hidden optimizations.
Thus, even if linear matrix multiplication and SHA3 have no hidden optimizations, their combination, together with the Bitcoin block header format, *may* have hidden optimizations.
And there are no *current* incentives to find such optimizations until Bitcoin moves to this, at which point we are already committed and it would be highly infeasible to revert to SHA256d --- i.e. too late.
This is why changes to PoW are highly discouraged.
Remember, ASICBOOST was *not* an optimization of SHA256 *or* SHA256d, it was an optimizations of SHA256d-on-a-Bitcoin-block-header.
ASICBOOST cannot speed up general SHA256 or even general SHA256d, it only applies specifically to SHA256d-on-a-Bitcoin-block-header.
Regards,
ZmnSCPxj
📝 Original message:Good morning Michael,
> That’s a fair point about patents. However, note that we were careful about this. oPoW only uses SHA3 (can be replaced with SHA256 in principle as well) and low precision linear matrix multiplication. A whole industry is trying to accelerate 8-bit linear matrix mults for AI so there is already a massive incentive (and has been for decades).
>
> See companies like Mythic, Groq, Tesla (FSD computer), google TPU and so on for electronic versions of this. Several of the optical ones are mentioned in the BIP (e.g. Lightmatter)
Please note that ASICBOOST for SHA256d is based on a layer-crossing violation: SHA256 processes in blocks, and the Bitcoin block header is slightly larger than one SHA256 block.
Adding more to a direct SHA3 (which, as a "sponge" construction, avoids blocks, but other layer-crossing violations may still exist) still risks layer violations that might introduce hidden optimizations.
Or more succinctly;
* Just because the components have (with high probability) no more possible optimizations, does not mean that the construction *as a whole* has no hidden optimizations.
Thus, even if linear matrix multiplication and SHA3 have no hidden optimizations, their combination, together with the Bitcoin block header format, *may* have hidden optimizations.
And there are no *current* incentives to find such optimizations until Bitcoin moves to this, at which point we are already committed and it would be highly infeasible to revert to SHA256d --- i.e. too late.
This is why changes to PoW are highly discouraged.
Remember, ASICBOOST was *not* an optimization of SHA256 *or* SHA256d, it was an optimizations of SHA256d-on-a-Bitcoin-block-header.
ASICBOOST cannot speed up general SHA256 or even general SHA256d, it only applies specifically to SHA256d-on-a-Bitcoin-block-header.
Regards,
ZmnSCPxj