Loading…
Topic
812 posts about Performance. Every summary links to the original.
See the idea
Bare metal gives engineers direct control over GPU, network, and virtualization layers, but it shifts elasticity and lifecycle work onto the team. This trade-off map shows when that bargain pays off.
Bun packs a runtime, package manager, bundler, and test runner into one binary. Here is how those four parts fit together and where the seams still show.
A five-step how-to for deciding when an AMD Instinct GPU should stay dedicated and when SR-IOV virtual functions, VMs, or sidecar isolation can safely share it.
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Newest first
Parameter-efficient fine-tuning (PEFT) reduces the memory required to adapt open models, enabling fine-tuning with a fraction of the memory and, in some cases, quantized models. LoRA is dominant because it freezes base weights and trains a small set of added parameters, but its popularity may also reflect ecosystem momentum and uneven comparisons. The post presents shared-condition benchmarks for language-model math reasoning and image generation, using identical models, datasets, code, hardware, and measures including test performance, VRAM, forgetting or drift, runtime, and checkpoint size. The results show that LoRA performs well but is not consistently best: other techniques can outperform it on one or more axes, especially in the image-generation benchmark. The recommended choice is therefore to compare techniques against a use case’s needs and constraints, including quantization support, layer coverage, adapter merging, and downstream serving compatibility.
Final-answer checks can show whether an agent completed a task, but they miss the turns, tokens, seconds, errors, and command path required. The proposed tool-specific benchmark measures that work on transformers, comparing models, revisions, tasks, and three access tiers: bare installation, source clone, and a packaged Skill containing CLI documentation and examples. It uses deterministic tasks with exact matching, launches one Hugging Face Job per model–revision–task combination on identical hardware, and records traces in a Hugging Face Bucket. Reports include match rate, median new, cached, and generated tokens, time, error rate, and marker adoption, while traces expose the agent's commands. On transformers, the CLI plus Skill helped large open models but caused Qwen3-14B to misinterpret documentation as an executable tool, dropping classify-sentiment match from 100% to 0% in the cited comparison.
ClickHouse’s June 2026 newsletter marks the project’s ten years as open source, noting more than 2,000 contributors and community projects alongside engineering updates. It profiles Vasily Chekalkin’s 26.5 improvements to WebAssembly user-defined functions, including the DETERMINISTIC keyword and visibility in system.functions, and covers native Jaeger storage, the ClickCannon benchmark tool, and Open House announcements. The 26.5 release reports Top-N optimizations that make ORDER BY … LIMIT through JOIN 20.4× faster with 175× less memory on TPC-H SF100, while GROUP BY … LIMIT is 11.9× faster with 185× less peak memory. Other featured systems include a 15-million-rows-per-hour KPI pipeline, join-engineering work yielding a 26× improvement on join-heavy TPC-H workloads, and community guidance on inserts, columnar storage, ArrowStream, and RAG ingestion.
Open Electricity uses ClickHouse to serve live Australian energy data through a public platform and API spanning roughly one billion records. Migrating from Postgres and TimescaleDB addressed queries that could take more than five minutes, reducing a benchmark to just over three seconds and to 60 milliseconds with a materialized view. The new setup keeps heavy relational workloads in Postgres while ClickHouse handles analytics, using ReplacingMergeTree base tables and more than 20 materialized views for different aggregation patterns. It also uses substantially less hardware than the previous configuration: four cores, four GB RAM, and 200 GB disk versus 16 cores, 32 GB RAM, and 1.6 TB, while serving data added daily and API users.
GLM-5.2 is presented as a flagship model for long-horizon tasks, extending GLM-5.1’s capability to a solid 1M-token context for sustained engineering work. It was trained extensively on 1M-context coding-agent scenarios involving large-scale implementation, automated research, performance optimization, and complex debugging, and adds effort-level controls to balance capability, latency, and computational cost. Its architecture uses IndexShare, sharing one lightweight indexer across every four transformer layers and reducing per-token FLOPs by 2.9× at 1M context; MTP changes for speculative decoding increase acceptance length by up to 20%. On reported long-horizon and standard coding benchmarks, GLM-5.2 is described as the highest-ranked open-source model: it trails Opus 4.8 by 1% on FrontierSWE, ranks second only to Opus 4.8 on PostTrainBench, and scores 81.0 on Terminal-Bench 2.1 versus GLM-5.1’s 63.5.
Production financial intelligence systems frequently rely on hand-engineered tabular features and rule sets that are brittle, expensive to maintain, and blind to historical sequential patterns. NVIDIA demonstrates an accelerated reference pipeline that replaces standard BPE tokenization with a GPU-based domain tokenizer, converting raw transactions into semantic tokens with an 8,192-token context window. Using the NeMo AutoModel library, a compact 29-million parameter decoder-only transformer is pretrained from scratch on unlabeled transaction sequences using causal language modeling. Learned sequence embeddings are extracted, compressed using PCA, and concatenated with raw tabular features to train a downstream GPU-accelerated XGBoost fraud detection model. On the IBM TabFormer benchmark dataset, this combined approach achieves a 41.76% lift in Average Precision over the baseline model relying solely on raw tabular features.
NVIDIA submitted results across all MLPerf Training v6.0 benchmarks, setting performance records on workloads including the DeepSeek-V3 and GPT-OSS-20B Mixture of Experts models. To scale training up to 8,192 Blackwell GPUs, NVIDIA combined hardware cluster designs with scale-out networking via Spectrum-X Ethernet and Quantum InfiniBand. The engineering team eliminated CPU-GPU synchronization bottlenecks in token-dropless MoEs by implementing full-iteration CUDA graphs with synchronization-free operators and paged stashing. Further software optimizations included CuTe DSL kernel fusions, an MXFP8 attention block in the Transformer Engine, and transitioning router elementwise math to FP32. Across these benchmarks, the GB300 NVL72 platform achieved the fastest time to train at scale and the highest normalized per-accelerator performance.
Game developers integrating on-device artificial intelligence companions face challenges such as conversational latency, state synchronization, and loop prevention. NVIDIA announced the open-source ACE Game Agent SDK alongside a suite of Unreal Engine 5 plugins to streamline native, hardware-accelerated character pipelines. The SDK offers Agent, Chat, and RAG APIs, enabling characters to execute multi-step tool-assisted reasoning and query game databases. Unreal Engine 5 plugins introduce local runtime models including nemo-conformer-ctc-120m for speech recognition, Qwen 3.5 4B for language generation, and Chatterbox Turbo 350M for speech synthesis. Additional tooling expands to motion generation via Animotive Kimodo and rendering updates with the DLSS 4.5 Unreal Engine plugin.
Accelerating transformer training requires understanding how low-precision formats such as FP8 and NVFP4 impact specific general matrix multiplication (GEMM) workloads. Transformer configurations do not explicitly reveal these shapes, making microbenchmarks necessary before committing to full training runs. NVIDIA Transformer Engine enables quantization and kernel dispatch across precisions, which can be evaluated in realistic autocast mode or kernel-only prequantized mode. Profiling the ESM2-15B model on NVIDIA B300 GPUs demonstrated that NVFP4 achieved a 1.79x blended forward propagation speedup over MXFP8 and up to 4.01x over BF16 in prequantized execution. Although large GEMM dimensions successfully overcome quantization overheads, dynamic scaling, Hadamard transforms, and kernel selection asymmetries moderate real-world gains.
Databricks introduces Lakehouse//RT, a real-time data warehouse designed for operational analytics, BI, app serving, and observability workloads. It is powered by Reyden and is intended to deliver millisecond performance directly on lakehouse data without copying it into a separate serving layer. Preview participants saw up to 16x better performance, with response times as low as 10ms on smaller datasets, sub-100ms on larger ones, and sub-100ms latency at 12,000 queries per second on standard analytical benchmarks. Tests covering concurrency, dataset scale, and complex TPCDS queries report low latency where alternatives slowed or failed. Lakehouse//RT is in Beta for select read-only workloads, with incremental autoscaling and automatic baseline compute sizing.
ClickHouse marks ten years since becoming open source under the Apache 2.0 license, crediting more than 2,600 contributors and a decade of monthly releases for its growth into one of the world’s most widely deployed analytical databases. Recent engineering work spans joins, updates and deletes, JSON, full-text search, vector search, query acceleration, integrations, and developer tooling. Join-focused changes made the TPC-H SF100 workload 26× faster between versions 22.4 and 26.4, while statistics-based reordering reduced one six-table query from 3,903.7 seconds to 2.7 seconds. The project also publishes open benchmarks, including ClickBench, JSONBench, CostBench, and PostgresBench, with scripts, configurations, data, and results intended for independent validation. Its next phase continues to emphasize community contributions and broader ways to use ClickHouse.
ClickHouse’s tenth-anniversary history describes its path from an internal web-analytics experiment to an open-source analytical database with more than 2,000 contributors. The original problem was processing rapidly growing, real-time logs while answering flexible reports quickly; tested MySQL extensions and standalone analytical databases did not meet the stated scale of 100 billion records per day and 500 columns. The project built column-oriented storage, lightweight compression, custom aggregation structures, background merges, and query support for grouping, aggregation, filtering, and sorting. Released publicly on June 15, 2016, ClickHouse is presented as a level-three open-source project with public contribution guidelines, code review, testing, CI, documentation, and a roadmap, while contributors receive credit even when implementations are substantially rewritten.
Cisco Talos operates a threat-intelligence reputation service that classifies file hashes from Cisco product telemetry, enabling researchers to make security decisions about benign and potentially malicious artifacts. Previously, a custom pipeline moved data from Databricks Delta and S3 through EKS, where staging and aggregated tables supported an API; the self-managed design retained about a petabyte of compressed, replicated data on 32 high-compute nodes. The team rebuilt the system in four phases, using Terraform, SSO/SAML, sharded ClickPipes for a 12-month backfill, shadow traffic, and a validated cutover to ClickHouse Cloud on AWS. The resulting architecture uses EventBridge, SQS, ClickPipes, SharedMergeTree, object storage, and AWS PrivateLink, delivering less than a minute of queue lag, zero data loss, a 100% response match across more than 1,000 hashes, zero downtime, roughly 90% lower storage costs, and about 75% lower TCO while reclaiming over 300 engineer hours annually.
ClickHouse has achieved the AWS Retail Competency in the Advanced Data Insights category, a designation recognizing validated technical expertise and demonstrated customer success in retail real-time analytics. Retail workloads described here combine high-volume, high-cardinality, high-concurrency data with hard freshness requirements, including inventory, pricing, personalization, and peak seasonal demand. The post attributes the platform's fit to compressed columnar storage, parallel queries, continuous Kafka and Amazon S3 ingestion through ClickPipes, compute-compute separation, and AWS support for private networking, multi-region deployment, and Marketplace billing. Examples include Mercado Libre's 90-TB-per-day observability ingestion and trace lookups reduced from 15–20 minutes to about two seconds, while other customers report cost, speed, and analyst-efficiency gains.
Sentry upgraded its AI-driven issue grouping system to better prevent duplicate issues without merging distinct application errors. The team trained lightonai/modernbert-embed-large using Matryoshka Representation Learning on hundreds of thousands of stacktrace pairs labeled by Claude Sonnet 4.5. To optimize the high-throughput ingestion pipeline, embeddings were truncated from 768 to 64 dimensions, combined with bfloat16 precision, PyTorch SDPA, and CUDA graph compilation. A phased live rollout used threshold-gated index backfilling and automatic fallback to prevent issue spikes during migration. In production, the v2 model cuts the overgrouping rate from 8% to 4%, increases prevented duplicate issues to 70%, and delivers 6x faster inference.
AI development teams face constant shifts in model providers and agent frameworks, demanding data platforms that provide scalable, real-time context management. Agentic workloads require blending unstructured enterprise data, short-term session state, and persistent long-term memory. MongoDB addresses these requirements through its native JSON document model and integrated retrieval capabilities, combining full-text search, vector search, and hybrid search directly over operational data. Customer implementations such as DevRev, ElevenLabs, and Adobe rely on Atlas to achieve sub-100 millisecond hybrid retrieval and handle billions of requests. Additionally, MongoDB is collaborating with LangChain and ecosystem partners to establish open reference architectures and shared interfaces for portable agent memory across frameworks.
Postgres-to-Postgres migrations between providers can require manual schema handling, data movement, CDC, and cutover coordination, with risks around database objects and long-running loads. Postgres to Postgres ClickPipes, now in public beta, provides a fully managed migration service in ClickHouse Cloud designed to reduce operational overhead and support minimal downtime. It automates schema dump and restore, including secure connections, role conflicts, version differences, transactional application, retries, and UI or notification-based error reporting, while PeerDB supplies parallel initial loads and CDC capabilities. Independent pull and push processes use retries and alerting for transient errors, and the roadmap includes features to assist with cutover into Postgres by ClickHouse.
FlashAttention-4, originally designed for pre-training, required modifications to efficiently handle LLM inference workloads dominated by the memory bandwidth-limited decode phase. Modal’s contributions include adjusting parallelism strategies from query parallelism to key/value parallelism, implementing split KV to parallelize across KV tiles, and adding a heuristic for optimal splits. They added support for FP8 inputs, arbitrary KV page sizes via cp.async loads, and reduced wasted work for small query sizes by using a single query tile. The kernel also gained extended GQA packing for irregular head ratios, using cp.async for query tiles to bypass TMA alignment restrictions. These changes, measured in throughput improvements up to 3x for single-token decode and up to 2.4x for small page sizes, demonstrate the leverage of flexible parallelism and tile-based programming for inference optimization.
Part 2 of the series examines how PyTorch profiler traces change from a single bias-enabled nn.Linear to a three-layer MLP with an intervening activation, using scripts run on an NVIDIA A100-SXM4-80GB GPU. It explains that nn.Linear represents y = x @ w.T + b, dispatches aten::addmm, and folds bias addition into the cuBLAS GEMM epilogue, while aten::t only rewrites weight strides and launches no GPU kernel. For the MLP, eager execution uses three GEMMs plus separate activation and multiplication kernels, whereas torch.compile fuses the latter operations and keeps the intermediate out of HBM, while leaving the GEMMs unchanged. A hand-written Liger Triton kernel provides the same fusion without Dynamo guards or compile latency, trading Inductor’s shape-specialized speed for generic execution across changing shapes.
ClickCannon was created to answer how much CPU, memory, and disk are needed for ClickStack observability workloads, while accounting for schema, ingestion rate, hardware, and concurrent queries. It replays about 2 TiB of real OpenTelemetry trace data, converted to ClickHouse Native format with ZSTD compression, rather than relying only on synthetic datasets. A custom multithreaded Go reader/inserter provides control over uncompressed throughput and exposes stream metrics such as compressed and uncompressed bytes, row counts, and block counts, while the framework simulates users and query workloads during ingestion. The project evolved into an open-source, generic ClickHouse benchmarking tool for custom insert and query workloads, with fine-grained control over throughput, concurrency, and user behavior; the post reports that it can push substantial throughput from a single machine.
The best engineering reads, once a week.
