# Dynamo Published articles for Dynamo. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut DevFeed: [NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut]() Original publisher: [Read original article]() Author: Zhihan Jiang Published: 2026-09-16T15:00:48Z Content type: article Language: en Sources: [NVIDIA Blog]() Topics: [NVIDIA Vera Rubin](), [Vera Rubin NVL72](), [benchmarking](), [Inference Performance](), [Dynamo](), [vllm](), [Nvidia](), [TensorRT-LLM]() Tags: [ai-infrastructure](), [benchmarks](), [dynamo](), [efficiency](), [hardware](), [inference](), [mlperf](), [networking](), [nvidia](), [nvidia-vera-rubin](), [nvl72](), [optimization](), [performance](), [software](), [tensorrt](), [tensorrt-llm](), [vera-rubin-nvl72](), [vllm]() ### AI overview NVIDIA reports MLPerf Inference v6.1 preview results for Vera Rubin NVL72 and GB300 NVL72 systems. Vera Rubin NVL72 delivered up to 3.7x higher throughput than GB300 NVL72 on Qwen3-VL and up to 2.5x higher throughput on DeepSeek-R1, while a four-rack GB300 NVL72 submission achieved 99% scaling efficiency. The results used vLLM, NVIDIA Dynamo, and TensorRT-LLM. ### Source excerpt System performance, efficient infrastructure scaling and continuous software optimization are key levers that determine AI inference economics. Higher system performance means more tokens generated, resulting in higher revenue. Efficient scaling means throughput grows proportionally as hardware gets added, requiring fewer resources to serve users at scale. Continuous optimization means generating more value from infrastructure investments. [...] ## Building Pinterest's VLM Serving Stack on NVIDIA Dynamo DevFeed: [Building Pinterest's VLM Serving Stack on NVIDIA Dynamo]() Original publisher: [Read original article]() Author: Pinterest Engineering Published: 2026-09-10T23:08:16Z Content type: article Language: en Sources: [Pinterest Engineering Blog - Medium]() Topics: [vlm](), [Inference](), [model-deployment](), [Orchestration](), [Latency](), [Caching](), [Open Source Models & Datasets]() Tags: [blackwell](), [cache](), [dynamo](), [embeddings](), [engineering](), [inference](), [latency](), [llm](), [multimodal](), [multimodal-ai](), [nvidia](), [nvidia-blackwell](), [open-source](), [orchestration](), [performance](), [pinterest](), [vllm](), [vlm](), [vlm-serving]() ### AI overview Pinterest describes its VLM serving stack built on NVIDIA Blackwell GPUs and NVIDIA Dynamo. The stack addresses multimodal inference demands such as image processing, variable prefill costs, KV-cache pressure, routing, and cache offloading. ### Source excerpt Lei Pan | Senior Software Engineer; Salina Wu | Senior Software Engineer; Cristian Lopez | Software Engineer I; Guangtong Bai | Staff Software Engineer; Soam Acharya | Principal Engineer; Saurabh Vishwas Joshi | Principal Engineer; Chia-Wei Chen | Staff Software Engineer; Ambud Sharma | Principal Engineer Why VLM Serving Matters at Pinterest Pinterest is a visual search and discovery platform, so its AI systems must reason over both language and visual content. Vision-language models (VLMs), which can interpret images, compare visual candidates, and respond naturally to user intent, are becoming the foundation for the next generation of Pinterest experiences: Pinterest Assistant, hybrid search, multimodal reranking, content understanding, signal generation, content safety, and more. This direction also reflects Pinterest's broader strategy to customize open-source models to meet its product & scale needs. Pinterest Assistant is a standout example. This multi-turn conversational experience covers both user language and visual content. Serving it requires low-latency VLM inference over rich multimodal context as well as reworking Qwen3-VL with proprietary multimodal embeddings to cut runtime cost while improving performance. Serving VLMs, however, introduces more challenges compared to text-only LLM workloads. Requests may carry multiple images, require extra vision encoder computation, incur larger and more variable prefill cost, and create higher KV cache pressure. To support this new class of models & product experiences, we built Pinterest's VLM serving stack on top of NVIDIA Blackwell GPUs and NVIDIA Dynamo. Blackwell GPUs incorporate many architectural innovations that are uniquely positioned for today's most demanding AI workloads -- including higher BF16/FP8 compute throughput, increased memory bandwidth, and larger HBM memory capacity -- that enable dramatically higher performance for inference. Dynamo provides a distributed inference orchestration layer that g ## When to Use Encode-Prefill-Decode Disaggregation to Accelerate Multimodal Model Serving DevFeed: [When to Use Encode-Prefill-Decode Disaggregation to Accelerate Multimodal Model Serving]() Original publisher: [Read original article]() Author: Tanya Lenz Published: 2026-09-09T20:31:04Z Content type: article Language: en Sources: [NVIDIA Developer](), [NVIDIA Technical Blog]() Topics: [Optimization](), [model-deployment](), [multimodal](), [Dynamo](), [Inference Performance](), [Embeddings](), [Mixture of Experts (MoE)](), [GPU](), [Open Source](), [Artificial Intelligence]() Tags: [agentic-ai-generative-ai](), [ai-agent](), [ai-inference](), [computer-vision-video-analytics](), [developer-tools-techniques](), [dynamo](), [embeddings](), [gpu](), [inference](), [low-latency-inference](), [mixture-of-experts-moe](), [model-serving](), [multimodal](), [nvfp4](), [open-source](), [optimization]() ### AI overview This article explains when encode-prefill-decode disaggregation improves multimodal model serving. Using NVIDIA Dynamo, it separates vision encoding from LLM prefill and decode so the stages can batch, schedule, and scale independently. The approach is most effective for image-heavy prompts, short-to-medium outputs, and quantized mixture-of-experts models, with reported gains of up to 5x in time to first token and 7x in end-to-end response time. ### Source excerpt Encode-prefill-decode (EPD) disaggregation is an inference optimization technique for multimodal models that separates the vision encoder stage from the prefill... ## Introducing CUDA Rust: Two Tracks for Writing GPU Kernels DevFeed: [Introducing CUDA Rust: Two Tracks for Writing GPU Kernels]() Original publisher: [Read original article]() Author: Elizabeth Goodman Published: 2026-09-08T12:00:00Z Content type: article Language: en Sources: [NVIDIA Developer](), [NVIDIA Technical Blog]() Topics: [GPU](), [Code generation](), [toolchain](), [C++]() Tags: [c-plus-plus](), [cuda](), [cuda-tile](), [data-science](), [developer-tools-techniques](), [dynamo](), [gpu](), [kernels](), [nemo-retriever](), [open-source](), [programming-languages-compilers](), [rust](), [toolchains]() ### AI overview NVIDIA introduces CUDA Rust for writing GPU kernels natively in Rust and compiling them to PTX. The article explains the SIMT and Tile programming tracks, recommends Tile as the default, and describes cuda-oxide's Rust-based code-generation path. ### Source excerpt In September 2026, NVIDIA announced it is leaning into native GPU programming in Rust. CUDA C++ and CUDA Python are mature, enterprise-grade toolchains, and... ## NVIDIA Dynamo 1.0 Is Available to DigitalOcean Customers for Inference Performance and Cost Efficiency DevFeed: [NVIDIA Dynamo 1.0 Is Available to DigitalOcean Customers for Inference Performance and Cost Efficiency]() Original publisher: [Read original article]() Author: Waverly Swinton Published: 2026-03-19T22:13:37Z Content type: release Language: en Sources: [DigitalOcean]() Topics: [Dynamo](), [Digital Ocean](), [Inference Performance](), [Artificial Intelligence](), [Generative AI](), [Inference](), [Cloud](), [GB200](), [sglang](), [TensorRT-LLM](), [vllm](), [Kubernetes]() Tags: [ai-ml](), [cloud](), [cost](), [deployment](), [digitalocean](), [dynamo](), [gb200](), [inference](), [inference-performance](), [kubernetes](), [latency](), [llm](), [load-balancing](), [nvidia](), [product-updates](), [routing](), [sglang](), [tensorrt-llm](), [vllm]() ### AI overview NVIDIA Dynamo 1.0 is now available to DigitalOcean customers as an inference service framework for large-scale generative AI workloads. The article describes claimed performance improvements on NVIDIA GB200 NVL systems, cost-efficiency benefits, deployment options, and features including GPU-aware routing and disaggregated serving. ### Source excerpt NVIDIA Dynamo 1.0, which was released on Monday at NVIDIA GTC, is now available to DigitalOcean customers to help drive performance enhancements and cost efficiency. NVIDIA Dynamo 1.0 offers a 7x inference performance increase on NVIDIA GB200 NVL systems, and by pairing it with DigitalOcean's Agentic Inference Cloud, customers can achieve higher performance at lower costs while benefiting from seamless deployment. Working together, DigitalOcean's optimizations with NVIDIA have already achieved a 67% cost savings for customers like Workato, and this new generation of Dynamo can unlock even greater gains for businesses who run production-grade agentic workflows. DigitalOcean customers can get access to NVIDIA Dynamo 1.0 as a container image that can be run on a Droplet or can deploy directly on DigitalOcean Kubernetes with an inference runtime (vLLM, SGlang, TensorRT). What is NVIDIA Dynamo 1.0? NVIDIA Dynamo is a cutting-edge, high-performance inference service framework specifically designed to accelerate and optimize large-scale generative AI and inference models. Dynamo is an orchestration layer that sits above engines like vLLM, SGLang, and NVIDIA TensorRT-LLM. Think of it as the distributed traffic controller for your GPU fleet, seamlessly orchestrating GPU and memory resources across a cluster and reducing bottleneck by intelligently routing requests Key technical breakthroughs offered by Dynamo 1.0 include: 7x Performance Boost: When paired with NVIDIA Blackwell Ultra GPUs, Dynamo can increase inference performance by up to 7x, significantly lowering your cost per token. KV-Aware Routing: Instead of simple round-robin load balancing, Dynamo routes requests to the specific GPUs that already have the relevant "memory" from previous turns of a conversation. Disaggregated Serving: Dynamo splits the "prefill" (reading the prompt) and "decode" (generating the answer) phases across different GPUs to maximize utilization and reduce latency. Memory Offloading: The KV B ## Dynamo, DynamoDB, and Aurora DSQL DevFeed: [Dynamo, DynamoDB, and Aurora DSQL]() Original publisher: [Read original article]() Author: Marc Brooker Published: 2025-08-15T00:00:00Z Content type: article Language: en Sources: [Marc Brooker's Blog](), [Marc Brooker's Blog]() Topics: [Amazon DynamoDB](), [DSQL](), [Dynamo](), [NoSQL](), [Replication](), [Availability](), [Serverless](), [SQL](), [Amazon Web Services](), [Transactions]() Tags: [amazon-dynamodb](), [availability](), [aws](), [data](), [database](), [databases](), [dsql](), [dynamo](), [replication](), [serverless](), [sql]() ### AI overview This article compares the architectures of Amazon Dynamo, Amazon DynamoDB, and Aurora DSQL, beginning with how they provide durability and availability when a single host fails. It explains Dynamo's replication across successor nodes in a consistent-hashing ring and contrasts it with DynamoDB's replica groups, multi-AZ servers, Paxos replication, quorum acknowledgments, and scaling behavior. ### Source excerpt Dynamo, DynamoDB, and Aurora DSQL Names are hard, ok? People often ask me about the architectural relationship between Amazon Dynamo (as described in the classic 2007 SOSP paper), Amazon DynamoDB (the serverless distributed NoSQL database from AWS), and Aurora DSQL (the serverless distributed SQL database from AWS). There's a ton to say on the topic, but I'll start off on comparing how the systems achieve a few key properties. The key references for this post are: For Dynamo, Dynamo: Amazon's Highly Available Key-value Store from SOSP'07. For DynamoDB, Amazon DynamoDB: A Scalable, Predictably Performant, and Fully Managed NoSQL Database Service from ATC'22, Distributed Transactions at Scale in Amazon DynamoDB from ATC'23, and Lessons learned from 10 years of DynamoDB from the Amazon Science blog. For DSQL, my blog series on DSQL. Durability The databases we're looking at offer different levels of durability, but all three are designed not to lose data when a single host fails. Dynamo does this by taking advantage of its consistent hashing approach, replicating the data across multiple hosts in order in the hash ring: To achieve high availability and durability, Dynamo replicates its data on multiple hosts. Each data item is replicated at N hosts. ... Each key, k, is assigned to a coordinator node[]. The coordinator is in charge of the replication of the data items that fall within its range. In addition to locally storing each key within its range, the coordinator replicates these keys at the N-1 clockwise successor nodes in the ring. Like Dynamo, DynamoDB assigns a node in a hash ring to each individual item. But that's where the similarities stop. Instead of replicating across multiple nodes in the ring, in DynamoDB each node consists of a replica group with multiple servers in multiple AZs using Paxos to replicate the data. Instead of appearing the ring N times, each item appears once, and takes advantage of fault-tolerant nodes rather than spreading over multiple ## The DynamoDB paper DevFeed: [The DynamoDB paper]() Original publisher: [Read original article]() Author: Marc Brooker Published: 2022-07-12T00:00:00Z Content type: article Language: en Sources: [Marc Brooker's Blog](), [Marc Brooker's Blog]() Topics: [Amazon DynamoDB](), [NoSQL](), [Database](), [systems](), [API](), [Dynamo](), [Amazon Web Services]() Tags: [amazon-dynamodb](), [api](), [architecture](), [aws](), [customers](), [database](), [dynamo](), [performance](), [scale](), [systems]() ### AI overview The article discusses the DynamoDB paper, highlighting the database service's massive scale, evolution to meet customer needs, and lessons from real-world distributed-system design. It explains how non-uniform workloads caused partition hot spots and throttling despite sufficient total provisioned throughput, and contrasts DynamoDB's architecture with the earlier Dynamo system. ### Source excerpt The DynamoDB paper The other database called Dynamo This week at USENIX ATC'22, a group of my colleagues1 from the AWS DynamoDB team are going to be presenting their paper Amazon DynamoDB: A Scalable, Predictably Performant, and Fully Managed NoSQL Database Service. This paper is a rare look at a real-world distributed system that runs at massive scale. From the paper: In 2021, during the 66-hour Amazon Prime Day shopping event, Amazon systems ... made trillions of API calls to DynamoDB, peaking at 89.2 million requests per second 89 million requests per second is a big database by any standards (and that's just Amazon's use of DynamoDB)! What's exciting for me about this paper is that it covers DynamoDB's journey, and how it has changed over time to meet customers' needs. There are relatively few papers that cover this kind of change over time. For example: The uniform distribution of throughput across partitions is based on the assumptions that an application uniformly accesses keys in a table and the splitting a partition for size equally splits the performance. However, we discovered that application workloads frequently have non-uniform access patterns both over time and over key ranges. When the request rate within a table is non-uniform, splitting a partition and dividing performance allocation proportionately can result in the hot portion of the partition having less available performance than it did before the split. Since throughput was allocated statically and enforced at a partition level, these non- uniform workloads occasionally resulted in an application's reads and writes being rejected, called throttling, even though the total provisioned throughput of the table was sufficient to meet its needs. This is the kind of assumption in a system design--that splitting makes performance better--that's really easy to overlook when designing a system, and potentially difficult to fix when you're in production. A lot of what makes systems like DynamoDB so useful is ## The Consistency Series DevFeed: [The Consistency Series]() Original publisher: [Read original article]() Author: Nicolas Liochon (noreply@blogger.com) Published: 2015-05-12T15:43:00Z Content type: article Language: en Sources: [Nicolas Liochon]() Topics: [consistency](), [Databases](), [acid](), [Dynamo](), [CAP theorem]() Tags: [acid](), [availability](), [cap-theorem](), [consistency](), [database](), [databases](), [distributed-system](), [durability](), [dynamo](), [latency](), [network](), [partition](), [series]() ### AI overview The article introduces a new series on consistency models, focusing on practical details and their implementations. It distinguishes the topic from a simple strong-versus-eventual consistency trade-off and previews posts about durability, ACID, and the Dynamo model. ### Source excerpt "Those are my consistency models, and if you don't like them... well, I have others." - Groucho Marx, quoted from memory The CAP theorem series is coming to an end. I have a few posts planned, but they will be mainly about wrapping things-up-- I hope these posts will be great, though. Anyway, here starts a new series: the consistency series. This series is about getting into some practical details about the consistency models and their implementations. You think that the trade-off is only or mainly on strong consistency vs. eventual consistency? That everything is said with "W+R" vs. "N"? Then read this series. Settling on the color may not be enough when choosing a car. Posts, in recommended reading order The posts published so far are: Eventual Consistency and Durability: shows the link between durability and consistency, and why traditional databases have the 'D' in ACID. Tells you what 'W=1' implies in the Dynamo model. ## Android Architecture: Introducing Dynamo DevFeed: [Android Architecture: Introducing Dynamo]() Original publisher: [Read original article]() Author: SystemDotRun Published: 2015-04-03T00:00:00Z Content type: article Language: en Sources: [SystemDotRun]() Topics: [Dynamo](), [Library](), [Development](), [Code]() Tags: [android](), [android-architecture](), [app-architecture](), [asynchronous](), [code](), [dynamo](), [libraries](), [lifecycle]() ### AI overview This introductory post presents Dynamo, a small library and wiki for Android application architecture. It focuses on managing application state and asynchronous code, including their interaction with the Android lifecycle, and aims to decouple UI behavior from core application logic. ### Source excerpt Dynamo has been superseded by Pilot Welcome, welcome. Come in, it looks mighty cold out there. The code is laying heavy across the hills, disorder rules the day and the nights are long. We yearn for a day where the suns shines forth once more. This is post #2 of my very short Android Architecture series (can two be a series?). Post #1 was a primer for this post. It contained A brief outline of the problems when defaulting to stock Android architecture. An overview of common general architectural concepts. This post is really a short intro to a small library + Wiki I have created. This library fills an architectural hole which I see in the Android dev world. There is no unified approach to generic Android app architecture and this creates a vacuum which sucks in well meaning devs and results is messy codebases strewn across the land. It is also my reponse to the many MVP blog posts out there, which have really helped me think about this issue but also I have my own take on. Plus, I feel sorry for new Android devs who are turning up later to the party and are overwhealmed by the Android ecosystem and need to get up to speed quickly on what these issues are before even getting to the point of thinking about potential solutions. Some of the issues addressed by this approach are solved by fantastic existing libraries, but again, I think for a large chunk of devs these can be difficult concepts to grasp and get running with. These are mentioned in the projects wiki for further reading. In my experience at least there are two core ideas around the center of most apps codebases. From the codebases I have seen & inherited both ideas seem to have too few brain cycles spent on them. 1. State Most apps and views are state-based however the spaghetti-level that captures the state logic is generally very high. 2. Asynchronous code Pretty much every app involves some asynchronous code. Pretty much every app allows this to touch the Android lifecycle. This always leads to sadness. ## The Space Between Theory and Practice in Distributed Systems DevFeed: [The Space Between Theory and Practice in Distributed Systems]() Original publisher: [Read original article]() Author: Marc Brooker Published: 2014-08-10T00:00:00Z Content type: opinion Language: en Sources: [Marc Brooker's Blog](), [Marc Brooker's Blog]() Topics: [distributed-systems](), [systems](), [Apache Cassandra](), [Dynamo]() Tags: [architecture](), [blog-post](), [cassandra](), [distributed-systems](), [dynamo](), [systems]() ### AI overview The article examines the gap between theory and practice in distributed-systems education and literature. It argues that both theoretical understanding and practical experience are valuable, while more material is needed to connect the two. ### Source excerpt The Space Between Theory and Practice in Distributed Systems How do we learn synthesis? Teaching and learning about distributed systems, like any complex topic, requires real thought about what to teach and what to learn. It would be great to have enough time to teach and learn everything, but there's just too much material out there. Even if we did have the time to cover every paper, result, code base, experience report and blog post in the field, we'd still need to chose an order to cover them in. There's a natural order to things to the learned that makes them much easier to learn. That's why I was excited to see Henry Robinson's Distributed systems theory for the distributed systems engineer. It's a good list of things to learn about, from the practical to the theoretical. I like the way it's broken up into clear sections, and uses examples from many real-world systems from industry. This is a list I'm going to be recommending to people for a while, and going to be working through myself. Unfortunately, Henry's list reflects a greater gap in the overall literature: the gap between theory and practice. From Dynamo and Cassandra to Chubby and ZooKeeper there's a wealth of content available on the design and implementation of real systems. Some of these papers go into real depth on seemingly small details (like Paxos Made Live) while others concern themselves with high-level architecture. Combined with things like Deutsch's 8 Fallacies and Jeff Hodges Notes on Distributed Systems for Young Bloods there's a lot of practical advice available to learn about the practical side of distributed systems. On the theoretical side, there's also a wealth of material. Robinson points to the CAP proof and the FLP result, and admits that he's only just scratching the surface. There are thousands of good theoretical results out there, from the usual suspects like Lamport and Lynch to areas like topology and game theory. I feel like if I went through everything I've read on distrib ## Anti-RDBMS: A list of distributed key-value stores DevFeed: [Anti-RDBMS: A list of distributed key-value stores]() Original publisher: [Read original article]() Author: {"image"=\>"rj-headshot.jpg", "greetings"=\>nil} Published: 2009-01-19T06:00:00Z Content type: opinion Language: en Sources: [Richard Jones]() Topics: [Databases](), [Replication](), [data](), [Algorithms](), [cloud-computing]() Tags: [algorithm](), [algorithms](), [couchdb](), [data](), [database](), [databases](), [dht](), [distributed](), [dynamo](), [erlang](), [hadoop](), [hashing](), [http](), [java](), [nosql](), [programming](), [rdbms](), [replication]() ### AI overview This article presents research notes on distributed key-value and document stores that could replace groups of relational database shards under suitable conditions. It discusses use cases, replication and partitioning concerns, related systems and background research, while noting that the material was written in January 2009 and may be outdated. ### Source excerpt Please Note: this was written January 2009 - see the comments for updates and additional information. A lot has changed since I wrote this.