@mattpocockuk If the spec is deleted right after being generated, how does anyone, including yourself, six months from now, verify the code still does what was decided during ‘grilling’?
@mohanp_ai @akshay_pachaar Not quite textbook.. Ring-reduce is implemented in practice, but the implementation will chose an appropriate algorithm to perform the reduction, depending on the network topology, message size, number of ranks. The ring algorithm is very effective up to a certain node count
1/9
A new class of AI accelerators is gaining traction—based on analog computing.
They offer blazing-fast, ultra-efficient matrix-vector computations. But there’s a twist... 🧵
@BenjDicken Transfer speed is also a function of physical distance of the memory unit from compute logic, clock frequency, bus width, number of outstanding requests in flight, number of pipeline stages within the interconnect logic.
The viz shows one view of a much more complex system.