arXiv:2608. 10103v1 Announce Type: cross Abstract: High-performance Tensor Core kernels rely on a low-level PTX pipeline built from asynchronous data movement with cp.
arXiv:2606. 13392v1 Announce Type: new Abstract: Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale.
arXiv:2607. 14568v1 Announce Type: cross Abstract: A companion study ran a 35B mixture-of-experts model on a 2011 NVIDIA Tesla C2075 (Fermi, sm_20, 6GB) as a GPU-prefill/CPU-decode hybrid, because the 4-bit model did not fit in device memory (arXiv:2606.
FlyWheel introduces optimized TPU kernels for softmax and linear attention, reducing vector work and improving matrix operand reuse. By overlapping independent computations, FlyWheel achieves up to 4.14× speedups over the RPA kernel and 1.80× over Splash Attention, reaching 81% of peak BF16 throughput on TPU v6e. It also boosts Gated DeltaNet prefill by up to 3.61× and enhances vLLM serving of Qwen3.8-27B prefill throughput by up to 1.56× while preserving accuracy.
arXiv:2610. 23796v1 Announce Type: new Abstract: Transformer prefill reuses weights across tokens and is dominated by matrix multiplication and activation data movement; single-token decode has limited reuse and is dominated by weight and KV-cache reads.
FlashBoB introduces an I/O‑efficient algorithm for exact backward‑over‑backward (BoB) in softmax attention, enabling precise second‑order differentiation without large intermediate tensors. By exploiting a hierarchical affine structure, the method confines computation to on‑chip tiles and limits off‑chip memory traffic, achieving θ(N² d²/M) HBM usage. Experiments show FlashBoB scales to sequence lengths of 262K on a single A100 GPU, outperforming prior exact baselines and FlashBack by up to 6.3×.
arXiv:2611.05419v1 Announce Type: cross Abstract: The RTX 2080 Ti (TU102) ships full second-generation Tensor Cores in hardware, yet the vendor throttles FP16 matrix multiplication at the driver leve...
EFQ-Softmax is a low‑bit probability‑generation technique that replaces the traditional exp‑then‑quantize path in Transformer attention. It maps shifted attention scores directly to block‑scaled E2M1 operands using an exponent‑only scale and a single affine rule, allowing the same low‑bit representation to be used for both numerator and denominator updates. Experiments on Qwen3‑8B, Qwen3‑VL‑8B‑Instruct, and WAN2.2‑TI2V‑5B show that EFQ‑Softmax maintains or improves model quality while reducing vector‑stage latency by about 40% on the A5 vector unit.
The paper identifies a missing cost term in the Tensor‑Memory Equilibrium (TME) model for emulating fp64 on GPUs: the per‑input deconstruction cost required before matrix multiplication can proceed. By adding this term and calibrating it against cuBLAS, the authors derive a closed‑form operational‑intensity threshold (OI*) below which emulation cannot match native fp64 performance. On an NVIDIA B300 GPU, this threshold is about 0.56 FLOP/B, limiting the speed‑ups of memory‑bound kernels such as GEMV, SpMV, and low‑batch GEMV to 0.3–0.9× native, and reducing the 7‑point stencil speed‑up from 3.1× to 1.8×, while dense GEMM remains unaffected.
The paper investigates using short polynomial approximations to accelerate special‑function operations in large language models on NVIDIA Blackwell GPUs. By replacing native sigmoid, tanh, and SiLU with degree‑3 or degree‑4 bfloat16 programs, the authors achieve up to 2.19× speed‑ups in isolated FP16 benchmarks and modest training‑step throughput gains (2.7–8.0%) across four integration tasks. The study also evaluates model behavior, finding negligible training‑loss differences within 100 billion tokens.
The article argues that on AI‑optimised NVIDIA B300 GPUs and newer, the FP8 tensor‑core matrix operation—implemented via the CRT‑based Ozaki Scheme II—can become the primary substrate for matrix‑heavy FP64 kernels while maintaining FP64‑grade accuracy. It introduces the Tensor‑Memory Equilibrium (TME) model, a Roofline extension with four parameters, to show that FP8 can match native FP64 performance under certain intensity thresholds and tile‑fusion conditions. The study identifies two notable exceptions—large dense‑square DGEMM and 3‑D FFT—where additional hardware or software adjustments are required to reach the memory roof. whyItMatters":"The paper demonstrates that FP8, with appropriate reconstruction and deconstruction strategies, can replace native FP64 for high‑performance computing workloads on modern GPUs, potentially reducing hardware complexity and energy consumption while preserving accuracy."
The paper introduces Group-Query Latent Attention (GQLA), a variant of Multi-head Latent Attention that shares group‑sharded up‑projections to enable a joint latent representation and a shardable expanded cache. GQLA supports both MQA‑absorb and GQA execution using the same weights and existing kernels, and roofline analysis shows how absorption trades traffic for computation while GQA shards KV groups. Experiments on DeepSeek‑V3.1‑Base and other models demonstrate that GQLA achieves comparable training losses to MLA, maintains high accuracy after checkpoint conversion, and improves output throughput by up to 35% on eight H20‑3e GPUs, while also allowing larger batch sizes due to a smaller per‑rank cache.