vLLM tops the Artificial Analysis ranking by delivering the highest throughput for DeepSeek V3.2, Qwen 3.5 397B, and MiniMax‑M2.5 on identical NVIDIA Blackwell Ultra hardware, thanks to extensive kernel‑fusion optimizations that remain in the main branch.
QuantClaw, an open‑source plug‑in for the OpenClaw AI agent framework, uses a systematic quantization study to dynamically route tasks to appropriate model precisions, achieving up to 21% cost reduction, 8‑15% latency improvement, and even higher task scores across diverse workloads.
A benchmark of Redis, Memcached, DragonflyDB, and Cashews using the same FastAPI workload reveals that Redis falls behind on latency, throughput, and memory efficiency, while DragonflyDB and Cashews offer superior performance and developer experience for Python caching.
A recent ICLR 2026 study from Zhejiang University, Ant Group, and Stony Brook introduces a problem‑oriented dataset and an anchor‑verification framework that enable large language models to not only generate correct code but also significantly improve its execution speed, achieving up to six‑fold acceleration while maintaining high correctness.
Taalas embeds the Llama 3.1 8B model directly into a 6nm ASIC, delivering 17,000 tokens per second—nearly ten times faster than top NVIDIA GPUs—while cutting system cost by over tenfold and power consumption by tenfold, albeit with limited flexibility and quantization trade‑offs.
ClusterAttn tackles the KV‑cache bottleneck of large language models by exploiting the natural clustering of attention scores, achieving up to 92% compression without accuracy loss, boosting throughput 2.6–4.8×, handling 128K‑token sequences on a single GPU, and outperforming existing training‑free compression methods.
This article introduces OpenAI's Triton language, compares its block‑wise programming model to traditional CUDA, walks through vector‑addition and fused‑softmax kernel implementations, and presents benchmark results that demonstrate significant speedups over native PyTorch operations.
This article explains the evolution of file system architecture from kernel‑level VFS to userspace via FUSE, reviews the historical role of NFS, details JuiceFS's implementation on top of FUSE, and presents benchmark results that demonstrate its high throughput and practical limitations.
The paper introduces DiC, a fully convolutional diffusion model that rethinks 3×3 convolutions, adds sparse skip connections, stage‑specific embeddings and conditional gating, and demonstrates superior FID/IS scores and faster inference compared to diffusion Transformers across multiple scales.
This article explains how OpenSearch's GPU‑accelerated vector search leverages parallel graph algorithms to achieve up to tenfold speed improvements over CPU solutions, detailing ANNS techniques, performance benchmarks, and practical GPU specifications for high‑QPS AI applications.
This article investigates why enabling line‑prefix regex for multiline logs slows Logtail down, explains the underlying regex matching mechanism, and demonstrates how switching from boost::regex_match to boost::regex_search with proper flags can dramatically improve collection throughput, achieving a seven‑fold speed increase.
This article presents a comprehensive benchmark comparing memcached and ScyllaDB across memory, disk, and read‑only workloads, analyzes architectural trade‑offs, and offers practical guidance on when to prefer a simple in‑memory cache versus a persistent wide‑column database.
With the growing complexity of LLM architectures like GQA, MLA, and MoE, runtime overhead has become a bottleneck; this article analyzes Python performance, communication costs, and synchronous execution in current inference frameworks, introduces the fully asynchronous TAG architecture, and demonstrates its superior throughput and latency through benchmarks.
This guide explains how to deploy open‑source large language models using vLLM, benchmark latency and throughput, and apply 8‑bit/4‑bit quantization techniques such as BitsandBytes and NF4 to achieve faster inference on limited‑GPU hardware.
This article examines the trade‑offs between centralized and distributed OLTP databases, presents industry usage statistics, performance benchmarks, practical questions for migration, and detailed guidance on sharding, SQL design, and operational considerations to help decide when a distributed solution is truly needed.
This article presents a detailed benchmark of four Transformer models of varying sizes trained on the high‑end Inspur NF5488A5 GPU server, compares its NVSwitch‑based interconnect with a PCIe‑based system, and analyzes the impact of model scale, tensor parallelism, and hardware bandwidth on training efficiency.
The article presents a detailed technical review of Kunlun Chip's XPU‑R AI accelerator, covering its evolution from early FPGA prototypes to the current 7nm, 256 TOPS chip, the architectural choices that address AI workload demands, performance advantages over CPUs/GPUs, and the product ecosystem supporting diverse AI scenarios.
The article explains CNN inference optimization by applying PyTorch quantization and module‑fusion techniques, compares model size and latency before and after quantization, shows code for building, quantizing, and fusing a simple CNN, and presents benchmark results on CPU, highlighting a four‑fold size reduction and up to 1.7× speed‑up.
The article introduces Volcano Engine's veGiantModel, a high‑performance large‑model training framework built on PyTorch, Megatron and DeepSpeed, details its distributed parallel strategies, hardware setups, benchmark results showing up to 6.9× speedup over Megatron and DeepSpeed, and provides open‑source links for further use.
In a series of local benchmarks on a 2.6 GHz six‑core Intel i7 machine, the author compares K6, Gatling, and FunTester under 10 k to 20 k QPS loads, detailing CPU, memory, and response‑time metrics, analyzing script languages, JVM settings, and offering optimization suggestions for FunTester.
This article benchmarks four load‑testing frameworks—JMeter, k6, Locust, and FunTester—across multiple concurrency levels, measuring CPU, memory, QPS and response time to reveal each tool’s strengths, weaknesses, and scalability limits.
Facing high memory usage and slow startup after containerizing its Java services, NetEase Media adopted Go in 2020, leveraging its fast compilation, low‑resource footprint and goroutine‑based concurrency to build the high‑performance Ngo framework, which outperforms Spring‑Boot in throughput while using far less memory.
This article explains how Tencent's TDSQL database tackles the combined challenges of transaction and distributed consistency by introducing a multi‑level strong consistency model that delivers several‑fold performance improvements over Spanner, CockroachDB, and native Greenplum while preserving ACID guarantees.
An in‑depth Java performance study explores LongAdder, compares it with AtomicLong and lock‑based counters using JMH, and walks through successive custom implementations (V0‑V5) that apply striping, modulo optimization, false‑sharing elimination, and advanced hash probing to approach or surpass LongAdder’s throughput.
This article outlines how 360’s data center built QuickSQL, a federated SQL engine that unifies queries across heterogeneous sources such as Hive, MySQL, and Elasticsearch, detailing the business challenges, architectural design, performance benchmarks, and future roadmap for multi‑source data analysis.
This article examines whether Pilosa, a bitmap index originally built for sparse high‑cardinality data, can efficiently handle dense relational datasets by benchmarking it against a billion‑row NYC taxi trip dataset and comparing query performance with other database systems.
The author rewrote a Laravel‑based Boxzilla application in Go, detailing migration steps, code‑size reduction, benchmark results, and testing advantages, showing how Go delivers faster response times, lower latency, and a more maintainable backend.
