Tagged articles
kernel optimization
16 articles · Page 1 of 1
Fudan University and Huawei's XiaoYi team propose SAGE, an AI agent system that combines knowledge cold-start, tool-enhanced diagnosis, and experience learning to generate high-performance kernels for Ascend NPUs, achieving 95.5% success rate on NPUKernelBench with up to 10.86x speedup over vendor libraries.
AI AgentAscend NPUExperience Learning
0 likes · 39 min read

vLLM's PR #63199 introduces a dedicated FP8 W8A8 Blockscale GEMM kernel for AMD's gfx1151 (RDNA 3.5) architecture, accelerating specific small-batch matrix multiplications by 1.70–2.54x and improving end-to-end generation throughput by ~6.8% on Ryzen AI MAX+ 395, though the PR remains in draft and covers only a subset of FP8 operations.
AMDFP8GEMM
0 likes · 8 min read

Tencent shares four kernel-level Elasticsearch optimizations: memory-based smooth rate limiting to prevent cluster avalanches, Rollup pre-aggregation cutting storage costs 10x, moving FST off-heap to manage 50TB per node, and metadata scalability improvements supporting million shards with sub-5-second index creation.
ElasticsearchRollupTencent
0 likes · 7 min read

MongoDB’s physical backup can cause hidden nodes to balloon in disk usage, but by analyzing WiredTiger’s space‑reclamation process and introducing a checkpoint‑release API, the team reduced backup‑induced disk expansion from up to 200% down to under 10%, saving hundreds of gigabytes and cutting backup time by over 40%.
Checkpoint ReleaseDisk BloatMongoDB
0 likes · 19 min read

The article breaks down DeepSeek’s massive hiring push, outlines the core traits DeepMind’s Vlad Feinberg says are essential, and provides a step‑by‑step roadmap—including kernel work, scaling‑law studies, and open‑source contributions—to help aspiring engineers break into top AI research labs.
AI labsDeepSeekScaling Laws
0 likes · 10 min read

The article presents RTPurbo, a lightweight two‑stage training method that converts full‑attention LLMs into highly sparse models with over 97% sparsity, achieving up to 9.36× prefill and 2.01× decode speedups while preserving near‑lossless accuracy across long‑context benchmarks up to 512K tokens.
Dynamic Token SelectionLLM inferenceLow-Dimensional Projection
0 likes · 17 min read


SuanNiMay 22, 2026 · Artificial Intelligence
Qwen3.7-Max demonstrates product‑level long‑task autonomy with 35 hours of uninterrupted operation, 1,158 tool calls, and kernel‑level optimizations, while outperforming Gemini 3.5‑Flash, Claude Opus, and GPT‑5.5 across a wide range of benchmarks, cost‑effectiveness, and real‑world agent scenarios.
AIAgentEnvironment Scaling
0 likes · 11 min read

The article walks through a step‑by‑step optimization of a simple elementwise addition kernel (C = A + B) on HyperAI's RTX 5090 cloud instance, covering FP32 baseline, vectorized FP32, several FP16 variants, benchmark methodology, performance results, and the reasoning behind thread‑block sizing.
CUDAElementwiseFP16
0 likes · 30 min read

The article details Xavier Xia’s iterative patches to contpte_ptep_get(), showing how early‑exit logic and subsequent refinements consistently improve performance across all tested scenarios without regressions, backed by benchmark data and community discussion.
Xvisorbenchmarkcontpte_ptep_get
0 likes · 4 min read

Collaborating across TencentOS and database kernel teams, the article details how architecture upgrades—moving to TKE HouseKeeper, switching to AMD CPUs, and adding a portable 16 KB atomic‑write feature—combined with kernel optimizations like huge‑page support, NUMA‑aware qspinlocks, speculative page‑fault handling, and ORC unwinding to deliver up to 30 % mixed workload and over 100 % write‑only performance gains while reducing memory usage.
NUMAORC unwinderatomic write
0 likes · 16 min read

ByteDance won the 2024 Open Compute Best Practice Award for its groundbreaking work in cloud firmware, OpenBMC, VDUSE virtualization, and kernel memory optimizations, illustrating how open‑source collaboration drives more efficient, scalable data‑center infrastructure in the AI era.
Open ComputeOpenBMCkernel optimization
0 likes · 10 min read

The article examines Linux’s Completely Fair Scheduler in Android, detailing weight‑based fairness, nice‑value mapping, group scheduling via CONFIG_FAIR_GROUP_SCHED, and the resulting challenges where vruntime compensation and priority adjustments become ineffective, leading to uneven time‑slice allocation among critical and background threads.
Android SchedulingCFS SchedulerGroup Scheduling
0 likes · 19 min read

veLinux, an open‑source Linux distribution from Volcano Engine, delivers a stable, high‑performance, and secure operating system for cloud workloads on both x86 and ARM, offering multiple editions, deep kernel optimizations, fast boot, and extensive security features tailored for modern cloud‑native environments.
Securitykernel optimizationperformance
0 likes · 17 min read


ITPUBJan 12, 2021 · Fundamentals
A seven‑day series of experiments measures the memory footprint of TCP sockets in various states—including ESTABLISH, data transfer, and TIME_WAIT—showing how the Linux kernel optimizes socket memory usage and what factors influence the actual consumption.
LinuxTCPkernel optimization
0 likes · 4 min read

The article presents a seven‑day series of experiments measuring how much memory a TCP connection occupies in various states—including ESTABLISH, data transfer, and TIME_WAIT—revealing kernel optimizations for socket buffer reclamation and overall memory overhead.
LinuxTCPkernel optimization
0 likes · 3 min read

In this interview, JD.com’s infrastructure director Bao Yongcheng discusses his work on large‑scale container platforms, kernel‑level scheduling improvements, power‑saving experiments, and the future of cloud‑native data centers, highlighting practical lessons from the 11.11 shopping festival preparation.
KubernetesOpenStackcontainers
0 likes · 9 min read
