Upgrade to TVM0.42-alpha4 and use its new instruction sponge_absorb_mem to save 31 % of the op-stack table rows in the reference run of the STARK verifier for inner padded height $2^{20}$, expansion factor 4. This makes the processor table (clock_cycle_count below) the new bottle neck[^0].
The one failing test is due to an upstream problem in Triton VM which is fixed in this Triton VM PR.
[^0]: Adding the next instruction in our wishlist recurse_or_return should allow us to reduce clock cycle count by another 20 % which would bring it very close to the RAM table height.
Upgrade to TVM0.42-alpha4 and use its new instruction
sponge_absorb_mem
to save 31 % of the op-stack table rows in the reference run of the STARK verifier for inner padded height $2^{20}$, expansion factor 4. This makes the processor table (clock_cycle_count
below) the new bottle neck[^0].The one failing test is due to an upstream problem in Triton VM which is fixed in this Triton VM PR.
Before
After
[^0]: Adding the next instruction in our wishlist
recurse_or_return
should allow us to reduce clock cycle count by another 20 % which would bring it very close to the RAM table height.