---
title: "Performance"
description: "817 posts about Performance, summarised, each linking to the original."
---

# Performance
> 817 posts about Performance, summarised, each linking to the original.

## Articles

### [Replacing the HDB: ClickHouse for historical ticker data](https://yomu.fyi/post/replacing-the-hdb-clickhouse-for-historical-ticker-data.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Published: Jul 15, 2026
- Type: Benchmark

Historical market-data systems separate a durable archive from the real-time trading path, but the archive still must support research, backtesting, transaction-cost analysis, and regulatory reconstruction. The worked example uses Binance’s public archive to load trades and quotes for five USD-M perpetual-futures symbols into ClickHouse, with a MergeTree table ordered by (symbol, ts) and codecs including DoubleDelta and ZSTD. SQL queries calculate VWAP, hourly OHLC candles, and as-of joins, while ClickHouse’s sparse primary index skips most granules for symbol-and-time filters. Across ten months, 16.2 billion rows compress to 76 GiB on disk; the post reports a ClickHouse Cloud cost of $1.88 per month and argues that typical symbol-and-time queries retain low latency at that scale, while broader scans read more data.


### [Building Faster Cryptography with Carryless Multiplication in NVIDIA CUDA 13.3](https://yomu.fyi/post/building-faster-cryptography-with-carryless-multiplication-in-nvidia-c.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Michelle Horton
- Published: Jul 15, 2026
- Type: Announcement

NVIDIA CUDA 13.3 introduces clmad, a native PTX carryless multiply-accumulate instruction for Ampere and newer GPU architectures. Prior to this release, GPUs lacked dedicated hardware for carryless multiplication, requiring developers to emulate binary extension field arithmetic with ALU operations or bitsliced circuits. The new primitive performs 64-bit by 64-bit carryless multiplication into 128-bit results with accumulation, allowing full 128-bit field multiplications via Karatsuba in six instructions. Benchmarks demonstrate that clmad achieves up to 6.3 TB/s throughput for GHASH on the NVIDIA B200, representing an 18.8x improvement over bitslicing. Additionally, sum-check protocol latency in zero-knowledge proving workloads decreases by up to 13x on the B200 and 4x on the RTX 5090.


### [Model Routing Is Simple. Until It Isn’t.](https://yomu.fyi/post/model-routing-is-simple-until-it-isn-t.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Yara Rizk, Eyal Shnarch, Jason Tsay, Merve Unuvar
- Published: Jul 15, 2026
- Type: Explainer

Model routing in agentic systems is presented as a systems optimization problem rather than a simple classification task. On 417 AppWorld Test Challenge tasks with the same CodeAct agent, Claude Sonnet 4.6 cost $79 total ($0.19 per task), versus $155 ($0.37 per task) for GPT-4.1, despite GPT-4.1 having lower listed token prices; cache-read pricing and repeated context changed the effective result. The discussion also shows why task difficulty and model speed are incomplete signals: hidden execution work, specialization, governance constraints, routing overhead, cache state, hardware, and endpoint load all affect decisions. The described router optimizes cost, quality, and latency together, producing configurable cost-accuracy operating points; one latency-oriented configuration reached 84% accuracy for $93 and 83 seconds, while adding about 6 ms and 2 kB per task.


### [Welcome Inkling by Thinking Machines](https://yomu.fyi/post/welcome-inkling-by-thinking-machines.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: ben burtenshaw, merve, Pedro Cuenca, Aritra Roy Gosthipaty, Andres Marafioti
- Published: Jul 15, 2026
- Type: Announcement

Thinking Machines Lab has released Inkling and the smaller Inkling-Small family as open multimodal models that accept image, text, and audio inputs. Inkling is described as a roughly 1T-parameter decoder-only Mixture-of-Experts model, with 975B total and 41B active parameters, a 1M-token context window, and training on 45 trillion multimodal tokens. Its architecture combines relative attention, alternating global and sliding-window attention, short convolution, shared experts, hierarchical image patchification, and discretized mel-spectrogram audio processing. The release includes BF16 and NVFP4 checkpoints, speculative MTP layers, day-zero support in Transformers, SGLang, vLLM, and llama.cpp, plus deployment configurations and interactive voice and image demos. Reported benchmarks show Inkling and Inkling-Small achieving strong but varied results across reasoning, coding, vision, audio, factuality, chat, and safety evaluations.


### [Introducing Real World VoiceEQ: Measuring the human quality of voice AI](https://yomu.fyi/post/introducing-real-world-voiceeq-measuring-the-human-quality-of-voice-ai.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: David Ayllon, Alice, Jeff Brooks, Franc Camps Febrer, Jakub Piotr Cłapa, Theo Lebryk, Jens Madsen, Olya Ossipova, Sharath Rao, Hoon Shin, Tigran, Rashish Tandon, Panagiotis Tzirakis
- Published: Jul 15, 2026
- Type: Announcement

Existing voice-AI benchmarks increasingly suggest near-human performance, while real-world conversations still expose failures involving speaker consistency, hesitation, accents, noise, emotion, and natural responses. Real World VoiceEQ measures human quality in voice interaction by testing whether systems recognize, produce, and respond to acoustic information omitted by transcripts, including tone, emotion, speaker identity, and background context. It evaluates more than 40 proprietary and open-source models across 15+ dimensions and 60+ metrics covering ASR, TTS, S2S, and Speech Understanding, using more than 1 million human ratings, including 785,000 TTS and 48,000 STS ratings. Findings indicate that capabilities are specialized, speech-to-speech systems often underuse paralinguistic cues, and traditional benchmarks can overestimate performance in noise and other conversational conditions. The research also reports that human agreement remains essential for subjective acoustic and social judgments because SLM evaluators were less reliable outside clear, verifiable tasks.


### [How Bullet uses ClickHouse Cloud to give DeFi's fastest exchange real-time analytics](https://yomu.fyi/post/how-bullet-uses-clickhouse-cloud-to-give-defi-s-fastest-exchange-real.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Published: Jul 15, 2026
- Type: Problem & solution

Bullet operates a DeFi exchange for perpetuals, spot trading, and lending, with trades settled on a blockchain while users retain custody of their assets. Its sub-millisecond trading system produces more than 150 million event rows per hour, creating a need for historical indexing, real-time analytics, and operational monitoring beyond the blockchain’s state-machine model. ClickHouse Cloud consumes events from Kafka, stores them in dedicated tables, and supports user features, leaderboard and reward calculations, P&L charts, audit logs, and Grafana-backed alerts. Replacing Databricks reduced query latency from 10–15 seconds to milliseconds, improved data freshness from 1–2 hours to under 5 seconds, and removed a separate DynamoDB serving layer at comparable cost while handling 1,000 times more data across staging, testnet, and mainnet.


### [@clickhouse/rowbinary: when your library is also a parser compiler](https://yomu.fyi/post/clickhouse-rowbinary-when-your-library-is-also-a-parser-compiler.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Published: Jul 14, 2026
- Type: Explainer

ClickHouse's @clickhouse/rowbinary is a Node.js reader and writer for the RowBinary, RowBinaryWithNames, and RowBinaryWithNamesAndTypes formats, offering both a generic parser and an accompanying Agent Skill. RowBinary avoids per-row overhead but is difficult to consume in JavaScript because nested types, LEB128 varints, precision-sensitive DateTime64 values, and recursive self-describing types require extensive runtime dispatch; JSON alternatives can also round large UInt64 values. The package provides small, composable, monomorphizable reading primitives, while SKILL.md instructs a coding agent to assemble query-specific parsers with inlined operations and selected policies such as BigInt handling and array materialization. Generated parsers run 1.5–3.4x faster than composing library functions in loops, with a fixed-width benchmark parser reaching 3.41x faster, and produce reviewable source that can be tested and benchmarked like ordinary application code.


### [Post-Train NVIDIA Cosmos 3 in One Day Using Agent Skills](https://yomu.fyi/post/post-train-nvidia-cosmos-3-in-one-day-using-agent-skills.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Tanya Lenz
- Published: Jul 14, 2026
- Type: Tutorial

Adapting vision reasoning models for specialized physical AI video tasks often demands days of manual container configuration, training script authoring, and hyperparameter tuning. NVIDIA Cosmos 3 Nano, an omnimodal mixture-of-transformers model, pairs an autoregressive Reasoner tower with an iterative diffusion Generator tower for video question answering. By integrating NVIDIA TAO agent skills with a coding agent like Codex, developers can automate data patching, evaluation, and Low-Rank Adaptation (LoRA) post-training. In experiments using the Woven Traffic Safety dataset, LoRA fine-tuning boosted zero-shot baseline accuracy from 54.41% to 87.14% while using approximately seven times fewer GPU hours than full-parameter supervised fine-tuning. Combining this agent workflow with TAO AutoML further elevated model accuracy to 93.35% before automated deployment via NVIDIA NIM.


### [How we configure huge pages in ClickHouse Managed Postgres](https://yomu.fyi/post/how-we-configure-huge-pages-in-clickhouse-managed-postgres.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Published: Jul 14, 2026
- Type: Tutorial

ClickHouse Managed Postgres configures Linux huge pages to reduce the per-connection page-table cost and translation overhead of a large PostgreSQL shared\_buffers cache. The configuration reserves a quarter of machine memory before PostgreSQL starts, verifies HugePages\_Total, sets huge\_pages = 'on' to prevent silent fallback, and sizes shared\_buffers against the complete shared-memory segment, including buffer descriptors, WAL buffers, and lock tables. On an r7i.4xlarge with 32GB shared\_buffers, 4KB pages added 31.1MB of page tables per connection and reached 6.12GB at 200 connections, while 2MB huge pages reached 111MB. A select-only pgbench test at 100 clients also rose from 373,083 TPS to 418,087 TPS, a 12% increase, while startup failure without a sufficient pool made misconfiguration explicit.


### [Systems Engineering Playbook: Optimizing Qwen 3.5-397B MoE on Ironwood (TPU7x)](https://yomu.fyi/post/systems-engineering-playbook-optimizing-qwen-3-5-397b-moe-on-ironwood.md)
- Company: [Google](https://yomu.fyi/company/google.md)
- Author: Google for Developers
- Published: Jul 14, 2026
- Type: Benchmark

Serving Qwen3.5-397B, a 397-billion-parameter sparse Mixture-of-Experts model, on Ironwood (TPU v7x) requires fitting roughly 400 GB of weights into HBM while sustaining utilization across contrasting workloads. The report applies a modular, model-agnostic playbook built from reusable JAX/Pallas kernels and hardware-aware cost models, focusing new work on Gated DeltaNet linear attention and Attention Data Parallelism while integrating changes into vLLM and SGLang. The architecture activates exactly 17 billion parameters per token and combines 15 repeating blocks of three GDN layers and one GQA layer. Benchmarks use a single host with four Ironwood chips, eight logical devices, 8K/1K prefill-heavy and 1K/8K decode-heavy workloads, and concurrency tiers from 64 to 512. Between April and June 2026, stated optimizations improved inference performance by approximately 4.7x for 512-concurrency prefill-heavy workloads and 3.1x for decode-heavy workloads.


### [NVIDIA Ising Decoding Cuts Color Code Logical Error Rates by Over 300x](https://yomu.fyi/post/nvidia-ising-decoding-cuts-color-code-logical-error-rates-by-over-300x.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Elizabeth Goodman
- Published: Jul 13, 2026
- Type: Announcement

Fault-tolerant quantum computing requires effective quantum error correction, but color codes have historically been constrained by difficult decoding demands. NVIDIA introduced the Ising Decoder ColorCode 1 Fast, a pre-decoder designed to accelerate and improve color code logical error rates. The framework uses a 3D Convolutional Neural Network trained via cuStabilizer in NVIDIA cuQuantum alongside PyTorch to handle localized error syndromes. In benchmark testing at code distance d=31 with a 0.3% physical error rate, the decoder achieved over 347.7x lower logical error rates and 7.3x faster runtime compared to the Chromobius decoder. These local, space-time predictions scale to arbitrary distances, supporting parallel blockwise decoding for real-time quantum algorithm execution.


### [Ultra-Fast Anomaly Detection using Apache Spark Real-Time Mode](https://yomu.fyi/post/ultra-fast-anomaly-detection-using-apache-spark-real-time-mode.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Jitesh Soni
- Published: Jul 13, 2026
- Type: Benchmark

The post presents a reusable real-time guardrail pattern for flagging suspicious Ethereum blockchain transactions and routing them for downstream action. Its rules identify impossible blocks where gas\_used exceeds gas\_limit and scan extra\_data for email addresses, JWT tokens, or AWS access-key patterns, producing ALLOW or QUARANTINE decisions with reasons. The implementation uses Apache Spark Structured Streaming Real-Time Mode, whose continuous data flow, pipeline scheduling, streaming shuffle, pre-allocated execution pipelines, and asynchronous checkpointing target millisecond latency without a separate streaming engine. On a four-worker DBR 16.4 LTS cluster, the stateless Kafka-to-Kafka test processed about 23 million records at 69,713 rows per second, with P95 below 0.5 milliseconds and P99 at 1 millisecond. The post notes that RTM with a Kafka sink provides at-least-once delivery and that more complex stateful workloads may incur higher latency.


### [Modernizing the Meta Ads Service With an Open-Source Kernel Scheduler](https://yomu.fyi/post/modernizing-the-meta-ads-service-with-an-open-source-kernel-scheduler.md)
- Company: [Meta](https://yomu.fyi/company/meta.md)
- Published: Jul 13, 2026
- Type: Problem & solution

Meta’s ads serving fleet processes more than 5 million requests per second, so millisecond-scale latency changes can affect ads relevance and advertiser ROI. During a move from kernel 6.4 with CFS to kernel 6.9, the team observed a latency regression associated with EEVDF and used sched\_ext, an upstream BPF-based framework, to encode ads-specific scheduling priorities. The policy soft-partitions CPUs between latency-critical request-path threads and less-sensitive work, adjusts pool sizes with load-based heuristics, and runs as a user-space binary that loads BPF, enabling policy changes without rebuilding or reinstalling the kernel. Initial results were a 28% reduction in ads retrieval p99 latency, 3.28 megawatts of fleet power savings, and a 1.1% increase in weighted ads ranked; two later user-space updates delivered an additional 60% p99 reduction and an 18% decrease in critical-path timeout errors, with iterations shipping in days rather than months.


### [How Insider One uses ClickHouse Cloud to power real-time customer engagement at scale](https://yomu.fyi/post/how-insider-one-uses-clickhouse-cloud-to-power-real-time-customer-enga.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Published: Jul 13, 2026
- Type: Problem & solution

Insider One’s Unified Customer Database (UCD) unifies event data, user attributes, identity resolution, and profiles across channels to support real-time segmentation, personalization, and analytics. The platform initially ran ClickHouse in a self-managed AWS deployment, but growth brought larger datasets, more concurrent queries, and increasing work tuning clusters and balancing performance against cost. In a proof of concept, the team migrated an anonymized dataset from one of its largest customers to ClickHouse Cloud and replayed UI segmentation, audience-count, and scheduled-campaign workloads. Average query latency fell 50–60% without significant changes, reaching 70–75% for the largest customers, while latency percentiles became more predictable; the managed service also reduced operational burden and gave the team more capacity for product development and its AI-oriented roadmap.


### [Reducing High-Bandwidth Memory Bottlenecks in JAX-Based LLM Training with Host Offloading](https://yomu.fyi/post/reducing-high-bandwidth-memory-bottlenecks-in-jax-based-llm-training-w.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Tanya Lenz
- Published: Jul 10, 2026
- Type: Benchmark

Large language model training workloads frequently face GPU high-bandwidth memory (HBM) capacity constraints before compute limits are reached. To alleviate this memory pressure in JAX training workloads, host offloading transfers selected forward-pass activations to pinned host memory over high-speed NVLink-C2C interconnects and reloads them during the backward pass instead of recomputing them. Benchmarks executed on 128 GPUs within NVIDIA GB200 NVL72 systems using MaxText demonstrated significant improvements. For DeepSeek-V3 671B, host offloading combined with the Latency Hiding Scheduler and pipelined offloading reached 908.2 TFLOPs/s/device, surpassing activation rematerialization throughput by 57% and unlocking larger batch configurations. Similarly, QKV activation offloading on Llama 3.1 405B achieved a 2.9% throughput increase when transfer latency was hidden behind computation.


### [Amazon and University of Michigan give robots a sense of touch](https://yomu.fyi/post/amazon-and-university-of-michigan-give-robots-a-sense-of-touch.md)
- Company: amazon.science
- Author: Mani Nambi, Nima Fazeli
- Published: Jul 10, 2026
- Type: Problem & solution

Robotic manipulation in contact-rich settings requires accurate tactile force perception, but existing physics simulators either run too slowly for reinforcement learning or oversimplify shear forces. HydroShear addresses this tactile reality gap by incorporating path-dependent force tracking into hydroelastic contact models. By remembering object motion history across soft elastomer sensors, the GPU-parallelizable simulator computes full 3-D contact deformations, friction, and slipping dynamics. Reinforcement learning policies trained entirely in HydroShear transferred directly to a real Franka robot equipped with GelSight Mini sensors without fine-tuning. Across four challenging manipulation tasks, HydroShear achieved a 93 percent average success rate, outperforming baseline models TacSL and FOTS.


### [Kernel Fusion in NVIDIA CUDA: Optimizing Memory Traffic and Launch Overhead](https://yomu.fyi/post/kernel-fusion-in-nvidia-cuda-optimizing-memory-traffic-and-launch-over.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Michelle Horton
- Published: Jul 10, 2026
- Type: Explainer

GPU compute often outpaces memory bandwidth, causing separate kernels to suffer from intermediate global memory round-trips and launch overhead. Kernel fusion resolves this by executing multiple operations within a single kernel to keep intermediate data inside registers. Using the sum(abs(x)) operation as a benchmark on an NVIDIA GeForce RTX 4090, a naive two-kernel setup moved 3 GB of global memory over 3.51 ms. In contrast, manual CUDA C++ fusion, PyTorch compiler fusion via torch.compile, and explicit Python fusion using cuda.compute all achieved a 3x speedup. Across all three approaches, global memory transfers dropped to 1 GB by computing absolute values inline during reduction.


### [AI Model Co-Design: Hardware-Friendly LLM Design](https://yomu.fyi/post/ai-model-co-design-hardware-friendly-llm-design.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Elizabeth Goodman
- Published: Jul 10, 2026
- Type: Explainer

Efficient large language model serving requires co-designing architectures to balance datacenter token throughput and user interactivity within hardware constraints. Roofline analysis shows that linear layers become memory-bound when hidden or intermediate projection dimensions are small, leaving arithmetic hardware underutilized even at large token counts. For prefill workloads, Chunked Pipeline Parallelism reduces first-token latency on long contexts when models use regular, repeatable layer patterns that split into balanced stages. For latency-critical decoding, decoupling attention and feed-forward network parallelization enables independent optimization, applying expert parallelism for feed-forward networks and Helix parallelism to shard the key-value cache across sequences. Following architectural design rules—such as dimension alignment, width prioritization, and low-precision NVFP4 compatibility—allows models to achieve higher utilization on modern accelerator hardware.


### [Better tools made Copilot code review worse. Here’s how we actually improved it.](https://yomu.fyi/post/better-tools-made-copilot-code-review-worse-here-s-how-we-actually-imp.md)
- Company: [Github](https://yomu.fyi/company/github.md)
- Author: Napalys Klicius
- Published: Jul 10, 2026
- Type: Problem & solution

When migrating GitHub Copilot code review from custom navigation tools to shared Unix-style CLI utilities—grep, glob, and view—benchmarks revealed higher review costs and fewer caught issues. Rather than an issue with the underlying tools, trace analysis showed the agent used general-purpose coding assistant instructions that triggered expansive repository browsing loops. In response, the team rewrote the tool guidance to enforce a review-specific workflow anchored to pull request diffs. The updated instructions direct the agent to narrow candidate call sites and files using batched discovery before reading minimal line ranges with view. This workflow tuning reduced average review costs by approximately 20% in production while maintaining review quality.


### [One Driver, One Format, Every Language: ADBC](https://yomu.fyi/post/one-driver-one-format-every-language-adbc.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Published: Jul 10, 2026
- Type: Announcement

ClickHouse announces an official ADBC driver, a modern Arrow-native alternative to ODBC and JDBC for analytics and AI applications. It provides end-to-end columnar movement: query results leave ClickHouse as Apache Arrow and arrive in applications without a row-oriented conversion or staging step. A single Rust-built implementation connects through ClickHouse's HTTP interface, supports TLS, and serves languages and tools with ADBC bindings, including Ruby, R, and C. Users can install precompiled, signed binaries through the ADBC Driver Foundry and connect with a clickhouse:// URI; dbt Fusion distributes the driver directly. The driver also supports streaming Arrow inserts, while schema discovery, improved bulk ingest, usage examples, and JSON support remain on the roadmap.


[Newer posts](https://yomu.fyi/topic/performance/page/11.md) · [Older posts](https://yomu.fyi/topic/performance/page/13.md)