Search papers, labs, and topics across Lattice.
Affiliation:
6
0
5
2
Energy costs for computation are not just about operations; they include significant 'rent' and 'fare' that can dramatically increase with context length, reshaping our understanding of computational efficiency.
By abolishing traditional addressing, MADAR achieves a revolutionary reduction in energy consumption for data movement, making it a game-changer for AI computation.
Radar signal processing on Apple Silicon gets a 2.2x speed boost by cleverly managing the dynamic range of half-precision floats, proving that it's not about mantissa precision, but exponent range.
Compress transformers by 2x with no perplexity loss, or 3x with only a minor hit, simply by operating in the frequency domain.
Forget hand-tuning FFTs: a graph search that models cache effects finds a 5.2x faster implementation on Apple M1 by sandwiching radix-2 passes between radix-4 passes.
A simple tweak to FFT twiddle factor precomputation eliminates singularities and dramatically improves numerical precision, offering a free upgrade for anyone using FFTs.
