--- title: Behind the Scenes of Distributed Training and Why Your... description: A deep dive into distributed training strategies for large models, covering DDP, FSDP, and the ZeRO stages (1, 2, 3) from DeepSpeed as a spectrum between full... canonical: https://daily.dev/posts/behind-the-scenes-of-distributed-training-and-why-your-gpu-wiring-matters-as-much-as-your-strategy-hxdhri7gq twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Behind the Scenes of Distributed Training and Why Your GPU Wiring Matters as Much as Your Strategy | daily.dev og:description: A deep dive into distributed training strategies for large models, covering DDP, FSDP, and the ZeRO stages (1, 2, 3) from DeepSpeed as a spectrum between full... og:url: https://daily.dev/posts/behind-the-scenes-of-distributed-training-and-why-your-gpu-wiring-matters-as-much-as-your-strategy-hxdhri7gq og:image: https://api.daily.dev/og/posts/HxdHri7GQ.png og:image:alt: Behind the Scenes of Distributed Training and Why Your GPU Wiring Matters as Much as Your Strategy og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Behind the Scenes of Distributed Training and Why Your GPU Wiring Matters as Much as Your Strategy **[Towards Data Science](https://daily.dev/sources/tds)** · 27 min read · 1 upvotes · 0 comments ## Summary A deep dive into distributed training strategies for large models, covering DDP, FSDP, and the ZeRO stages (1, 2, 3) from DeepSpeed as a spectrum between full replication and full sharding. The author explains what each GPU stores under each strategy, the communication primitives involved (all-reduce, all-gather, reduce-scatter), and measured memory/throughput tradeoffs on four A100 GPUs. The second half examines GPU interconnect hardware — PCIe, NVLink bridges (NVL), and NVSwitch — showing that the same FSDP code can run 3–5x slower when a job spans NVL bridge groups versus staying within one group or using NVSwitch. Practical guidance is given on choosing a strategy based on model size and fabric topology, including using `nvidia-smi topo -m` to inspect wiring before launching jobs. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Should I use DDP or FSDP when fine-tuning Mistral-7B on A100 GPUs? Use FSDP if the model does not fit on a single GPU. Fine-tuning Mistral-7B with Adam in BF16 requires roughly 87GB per GPU for parameters, gradients, and optimizer states, which exceeds an A100's 80GB VRAM and causes DDP to crash with an out-of-memory error, while FSDP shards the state across 4 GPUs down to about 22-29GB per GPU, allowing it to run. _Anyone weighing memory limits against training speed can track distributed training tradeoffs like these on daily.dev._ ### How much does GPU interconnect topology affect all-reduce bandwidth in multi-GPU training? Interconnect topology causes roughly a 10x bandwidth difference between NVLink and PCIe on the same H200 GPUs, and up to a 9x difference within the same node depending on GPU placement. An NVSwitch fabric stays flat at 330-467 GB/s regardless of GPU count, a full four-GPU NVL bridge group reaches about 322 GB/s, but a job split across two NVL bridge groups falls to only 28-39 GB/s because it must cross PCIe. _Teams debugging inconsistent multi-GPU throughput can check topology-driven bottlenecks like these on daily.dev._ ### Why does FSDP training slow down more than DDP when GPUs are connected over a slow interconnect? FSDP communicates at every layer via all-gather and reduce-scatter operations, while DDP only performs one all-reduce per training step, so FSDP's per-layer transfers cannot hide behind compute on a slow wire. In measured Mistral-7B fine-tuning, throughput dropped 3 to 5x when a job spanned two NVL bridge groups instead of staying within one, with FSDP falling the hardest of the strategies tested. _Engineers choosing a sharding strategy under hardware constraints can weigh tradeoffs like these on daily.dev._ ## Similar posts on daily.dev - [AI in Multiple GPUs: ZeRO & FSDP](https://daily.dev/posts/ai-in-multiple-gpus-zero-fsdp-zvbbwsoip) · Towards Data Science · 0 upvotes · 0 comments - [Train Your Large Model on Multiple GPUs with Fully Sharded Data Parallelism](https://daily.dev/posts/train-your-large-model-on-multiple-gpus-with-fully-sharded-data-parallelism-cvv6hy9uk) · Machine Learning Mastery · 0 upvotes · 0 comments - [Optimizing Data Transfer in Distributed AI/ML Training Workloads](https://daily.dev/posts/optimizing-data-transfer-in-distributed-ai-ml-training-workloads-c015cc2oy) · Towards Data Science · 0 upvotes · 0 comments - [AI parallelism training strategies: Data, model, tensor and pipeline parallelism](https://daily.dev/posts/ai-parallelism-training-strategies-data-model-tensor-and-pipeline-parallelism-yeav4ssfg) · Data Science Central · 0 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#llm](https://daily.dev/tags/llm), [#gpu](https://daily.dev/tags/gpu), [#pytorch](https://daily.dev/tags/pytorch) [View this post on daily.dev](https://daily.dev/posts/behind-the-scenes-of-distributed-training-and-why-your-gpu-wiring-matters-as-much-as-your-strategy-hxdhri7gq) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Behind the Scenes of Distributed Training and Why Your GPU Wiring Matters as Much as Your Strategy","url":"https://daily.dev/posts/behind-the-scenes-of-distributed-training-and-why-your-gpu-wiring-matters-as-much-as-your-strategy-hxdhri7gq","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/behind-the-scenes-of-distributed-training-and-why-your-gpu-wiring-matters-as-much-as-your-strategy-hxdhri7gq"},"datePublished":"2026-07-09T17:08:17.329Z","dateModified":"2026-09-14T06:36:17.096Z","description":"A deep dive into distributed training strategies for large models, covering DDP, FSDP, and the ZeRO stages (1, 2, 3) from DeepSpeed as a spectrum between full...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/fd565f6665b9edb5a843119dfd571fe4?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/fd565f6665b9edb5a843119dfd571fe4?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Towards Data Science","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Towards Data Science","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/tds","url":"https://daily.dev/sources/tds"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/behind-the-scenes-of-distributed-training-and-why-your-gpu-wiring-matters-as-much-as-your-strategy-hxdhri7gq","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,llm,gpu,pytorch","timeRequired":"PT27M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Towards Data Science","item":"https://daily.dev/sources/tds"},{"@type":"ListItem","position":3,"name":"Behind the Scenes of Distributed Training and Why Your GPU Wiring Matters as Much as Your Strategy"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/behind-the-scenes-of-distributed-training-and-why-your-gpu-wiring-matters-as-much-as-your-strategy-hxdhri7gq#faq","mainEntity":[{"@type":"Question","name":"Should I use DDP or FSDP when fine-tuning Mistral-7B on A100 GPUs?","acceptedAnswer":{"@type":"Answer","text":"Use FSDP if the model does not fit on a single GPU. Fine-tuning Mistral-7B with Adam in BF16 requires roughly 87GB per GPU for parameters, gradients, and optimizer states, which exceeds an A100's 80GB VRAM and causes DDP to crash with an out-of-memory error, while FSDP shards the state across 4 GPUs down to about 22-29GB per GPU, allowing it to run. Anyone weighing memory limits against training speed can track distributed training tradeoffs like these on daily.dev."}},{"@type":"Question","name":"How much does GPU interconnect topology affect all-reduce bandwidth in multi-GPU training?","acceptedAnswer":{"@type":"Answer","text":"Interconnect topology causes roughly a 10x bandwidth difference between NVLink and PCIe on the same H200 GPUs, and up to a 9x difference within the same node depending on GPU placement. An NVSwitch fabric stays flat at 330-467 GB/s regardless of GPU count, a full four-GPU NVL bridge group reaches about 322 GB/s, but a job split across two NVL bridge groups falls to only 28-39 GB/s because it must cross PCIe. Teams debugging inconsistent multi-GPU throughput can check topology-driven bottlenecks like these on daily.dev."}},{"@type":"Question","name":"Why does FSDP training slow down more than DDP when GPUs are connected over a slow interconnect?","acceptedAnswer":{"@type":"Answer","text":"FSDP communicates at every layer via all-gather and reduce-scatter operations, while DDP only performs one all-reduce per training step, so FSDP's per-layer transfers cannot hide behind compute on a slow wire. In measured Mistral-7B fine-tuning, throughput dropped 3 to 5x when a job spanned two NVL bridge groups instead of staying within one, with FSDP falling the hardest of the strategies tested. Engineers choosing a sharding strategy under hardware constraints can weigh tradeoffs like these on daily.dev."}}]} ```