AI Explained
Plain explanations of trending AI concepts, with live visualizations.
ALTO MXFP4 training — Weight de-oscillation — What does it mean?
ALTO stops 4-bit weights flickering across bin edges, cutting MXFP4 training overhead from 20% to 5%.
AMD Primus fuses MoE all-to-all into GEMMs — Tile-level communication overlap — What does it mean?
MegaMoE fuses each MoE all-to-all into the GEMM that uses it, so compute overlaps communication per tile.
Co-design FP–INT GEMM hardware for 2–3-bit LLM weights — Why low-bit weights save memory but not compute — What does it mean?
Why 2–3-bit weights shrink memory but not GPU compute, and how a radix-4 FP–INT datapath fixes it.
Racer finds 82 GPU kernel bugs — symbolic equivalence checking — What does it mean?
Racer proves a GPU kernel matches its reference for every input at a checked launch, finding 82 bugs sampled tests missed.
Token-Energy Law — Active bytes and decode time — What does it mean?
Energy per token from shared weight bytes, private KV bytes and time per token.
Meta on AMD MI350X — Origami analytical kernel selection vs autotuning — What does it mean?
Origami picks GEMM tile shapes from a roofline-style cost model in microseconds, replacing per-shape autotuning tables.
BitSparse — Lossless sparse activation cache — What does it mean?
A bitmask plus nonzero values caches ReLU activations exactly, cutting VRAM per token about 30%.
LCT paper — Lossless Huffman coding of BF16 training tensors — What does it mean?
Trained BF16 tensors waste bits on predictable exponents; Huffman-coding only the exponent saves about 24% memory, bit-exact.
Apple's Minions GPU cloud — Pull-based work claiming — What does it mean?
Minions turns idle Macs into a GPU pool by letting workers pull and atomically claim tasks, with heartbeats to requeue lost work.
TACO fine-tunes 32B models on one H100 — One-sparse optimizer state — What does it mean?
TACO remembers 16 gradient entries per column instead of two numbers per weight, so a 32B model fine-tunes on one H100.
ThunderEP on PCIe GPUs — Ring relay vs single-step host broadcast — What does it mean?
Without GPU-to-GPU links, NCCL's ring relays MoE tokens through CPU memory many times; ThunderEP posts each shard once and reads it directly.
AMDKernelVault trains an 8B model to write AMD GPU kernels — Hierarchical execution reward — What does it mean?
Binary pass/fail leaves most GRPO rounds with zero advantage. AMDKernelVault grades a kernel 0.4, 1.0 and up to 1.5 instead.