Section V
APEX
Parallel Execution Engine
APEX runs transactions in parallel. Every transaction that does not touch the same state executes at the same time, on its own core, with no coordinator deciding the order. Add cores, add throughput.
what it does
What APEX Does
Almost every chain runs transactions in a line. One finishes, the next begins. It is the simplest thing that is obviously correct, and it wastes almost the entire machine.
But most transactions have nothing to do with each other. Two people trading different assets. Two applications writing different state. There is no reason on earth for one to wait on the other. Sequential execution was never a safety requirement — it is a default that nobody went back to revisit.
APEX runs everything that does not collide, at the same time.
composition not collision
Composition, Not Collision
Two transactions that never touch the same state produce the same result in any order. So the order does not have to be decided. It has to be shown not to matter.
APEX proves independence and schedules from that: parallel where the transition permits it, ordered where the transition requires it. No global lock, no optimistic guess, and no replay when the guess turns out wrong.
Other parallel engines run first and check afterward, then pay for it whenever they collide. APEX classifies before it runs.
exact conflict detection
Exact Conflict Detection
Exact means exact. When two transactions really do touch the same state, they get ordered. When they do not, they do not. And there is no third case where the scheduler guesses wrong and holds something back just in case.
Most conflict detectors use probabilistic filters. They are fast and they lie in one direction: they report a collision that is not there. Every false alarm is throughput thrown away for nothing, and at high load the false alarms are what set your ceiling.
APEX uses exact addresses. Zero false positives, by construction rather than by tuning.
linear scaling
Add a Core, Add Capacity
Throughput rises with the number of cores, not with the clock speed of any one of them.
This is the same turn the processor industry made twenty years ago and blockchains never did. You do not need a faster machine. You need a wider one, and wider is something you can buy.
More than 1.1M transactions per second on a 96-core node.
zero proving overhead
Zero Proving Overhead
Parallel execution normally costs something at proving time: scheduling metadata, ordering evidence, a record of who waited for whom. You win in execution and hand part of it back in the proof.
Here it costs nothing. The prover sees the same transition whether the work ran in parallel or in a line, because the consistency of the memory it touched is established once for the whole block rather than operation by operation.
The speedup is free. That is unusual enough to be worth saying twice.