Poster Type: Research Posters
Author: Tamino Steinert (Friedrich Schiller University Jena), Alex Breuer (Friedrich Schiller University Jena)
Abstract: AMD's XDNA2 NPU packs 32 six-issue VLIW cores where the schedule must resolve nearly all pipeline hazards. Each core achieves 1,843 GFLOPS by always issuing a VMAC.F operation, an 8×8×8 block-floating-point (BFP16) matrix multiply-accumulate. This corresponds to an array-wide performance of 59 TFLOPS. One obstacle is register pressure: a standard double-buffered 2×2 output block needs eight accumulator registers, whereas a core has only five.
We present a hand-scheduled VLIW BFP16 tensor-contraction kernel that time-multiplexes the output block's four accumulation chains onto two registers and issues a VMAC.F in every steady-state cycle.
On one core with L1-resident data, our kernel achieves 96.4% of theoretical peak (5.0× faster than the compiled mm_bfp.cc kernel). With the operands streamed from DRAM, a pre-tiled M3840×N3840×K8192 matrix multiplication reaches 21.4 TFLOPS on the whole NPU.
At the poster, we will walk visitors through the VLIW scheduling diagrams and the scannable artifact reproducing every measurement.
Best Poster Finalist (BP): no
Poster: PDF
Poster Summary: PDF