---
title: "How to choose a GPU topology for inference | Aguasto"
description: "Use per-device memory, topology, and a representative load test to choose between one GPU, a sharded replica, and multiple replicas."
canonical: "https://www.aguasto.com/blog/choose-gpu-topology-for-inference"
last-updated: "2026-11-17"
---

# How to choose a GPU topology for inference | Aguasto

Canonical HTML: [How to choose a GPU topology for inference](https://www.aguasto.com/blog/choose-gpu-topology-for-inference)

> Use per-device memory, topology, and a representative load test to choose between one GPU, a sharded replica, and multiple replicas.

[Aguasto](https://www.aguasto.com/) [Blog](https://www.aguasto.com/blog)  How to choose a GPU topology for inference

[Model serving & inference](https://www.aguasto.com/topics/model-serving)

## How to choose a GPU topology for inference

Use per-device memory, topology, and a representative load test to choose between one GPU, a sharded replica, and multiple replicas.

Aguasto · November 17, 2026

### Size one replica before choosing a GPU count

Start with the exact checkpoint, precision and maximum useful sequence length. A 70-billion-parameter checkpoint stored at two bytes per parameter has a theoretical weight payload of 140 decimal GB. NVIDIA lists 141 GB per H200 SXM and 80 GB per H100 SXM. On paper, the H200 example leaves about 1 GB before overhead. A hypothetical even split across two H100s assigns about 70 GB of raw weights to each device. Both are candidate calculations with different operating constraints.

Replace the weight estimate with observed allocation per device. Allow for cache, activations, workspace, metadata and framework allocations. A sharded runtime can load unequal amounts on different devices or retain tensors on each device. The device with the least remaining headroom can limit the entire replica. Measure a cold load and then repeat at your longest intended context and realistic concurrency.

### Choose how each request uses the hardware

If one device has enough memory and meets the quality and latency targets, use it as the baseline. When a replica needs several GPUs, tensor parallelism splits tensor operations across devices and requires frequent communication. Pipeline parallelism assigns different layers to different stages and passes activations between them. The engine and exact topology determine which split is practical. vLLM documents separate paths for one GPU, multiple GPUs on a node, and multiple nodes.

When the model already fits on one device, an independent replica gives the scheduler another place to send a request. Two replicas can serve separate requests, while a two-GPU sharded replica distributes the work of a single request. Compare them at the same accepted request mix and quality floor. For multiple nodes, request the exact network and test the communication path your software will use.

### Set the acceptance criteria before requesting capacity

Keep the workload fixed across candidates. A result should record the bottleneck and the conditions under which the configuration passed. Add the accepted topology and fallback to the capacity request.

- Record the model revision, container or engine version, precision and parallelism settings.

- Measure peak memory per GPU, time to first token, tail latency, successful throughput and errors under sustained load.

- Verify the device model, device count and fabric on the actual configuration before treating the trial as accepted.

### Sources

- [NVIDIA HGX hardware reference](https://docs.nvidia.com/enterprise-reference-architectures/hgx-ai-factory-h100-h200-b200/latest/components.html)

- [vLLM parallelism and scaling](https://docs.vllm.ai/en/latest/serving/parallelism_scaling/)

- [CoreWeave instance selection guide](https://docs.coreweave.com/platform/instances/selecting-an-instance)

### Continue the research

[Compare hardware specifications](https://www.aguasto.com/gpus/compare) [Model weights and runtime memory](https://www.aguasto.com/learn/model-weights-and-runtime-memory) [Prepare your capacity request](https://www.aguasto.com/blog/write-a-gpu-capacity-request)

### On this page

[Size one replica before choosing a GPU count](https://www.aguasto.com/blog/choose-gpu-topology-for-inference#part-1) [Choose how each request uses the hardware](https://www.aguasto.com/blog/choose-gpu-topology-for-inference#part-2) [Set the acceptance criteria before requesting capacity](https://www.aguasto.com/blog/choose-gpu-topology-for-inference#part-3) [Sources](https://www.aguasto.com/blog/choose-gpu-topology-for-inference#sources) [All blog posts](https://www.aguasto.com/blog)

## Company navigation

- [API versioning policy](https://www.aguasto.com/developers/versioning)
- [Developer sandbox](https://www.aguasto.com/sandbox)
- [Home](https://www.aguasto.com/)
- [Infrastructure](https://www.aguasto.com/infrastructure)
- [GPU Cloud](https://www.aguasto.com/gpu-cloud)
- [AI Inference](https://www.aguasto.com/ai-inference)
- [Development & Investment](https://www.aguasto.com/development)
- [Company](https://www.aguasto.com/company)
- [Contact](https://www.aguasto.com/contact)
- [Privacy](https://www.aguasto.com/privacy)
- [Terms of use](https://www.aguasto.com/terms)
- [GPU operations](https://www.aguasto.com/resources/gpu-operations)
- [Instance access](https://www.aguasto.com/resources/gpu-operations/instance-access)
- [Storage and checkpoints](https://www.aguasto.com/resources/gpu-operations/storage-and-checkpoints)
- [Networking](https://www.aguasto.com/resources/gpu-operations/networking)
- [Monitoring](https://www.aguasto.com/resources/gpu-operations/monitoring)
- [Interruption recovery](https://www.aguasto.com/resources/gpu-operations/interruption-recovery)
- [Get started](https://www.aguasto.com/gpu-cloud/get-started)
- [Troubleshooting](https://www.aguasto.com/support/troubleshooting)
- [Account access](https://www.aguasto.com/support/account-access)
- [GPU capacity requests](https://www.aguasto.com/support/gpu-capacity)
- [API access and billing](https://www.aguasto.com/support/api-and-billing)
- [Release notes](https://www.aguasto.com/release-notes)
- [Solutions](https://www.aguasto.com/solutions)
- [AI training](https://www.aguasto.com/solutions/ai-training)
- [AI inference](https://www.aguasto.com/solutions/ai-inference)
- [Fine-tuning](https://www.aguasto.com/solutions/fine-tuning)
- [Agentic AI](https://www.aguasto.com/solutions/agentic-ai)
- [AI agents](https://www.aguasto.com/solutions/ai-agents)
- [Large language models](https://www.aguasto.com/solutions/large-language-models)
- [Long-context inference](https://www.aguasto.com/solutions/long-context-inference)
- [Video generation](https://www.aguasto.com/solutions/video-generation)
- [Rendering](https://www.aguasto.com/solutions/rendering)
- [Scientific computing](https://www.aguasto.com/solutions/scientific-computing)
- [Engineering simulation](https://www.aguasto.com/solutions/engineering-simulation)
- [High-performance computing](https://www.aguasto.com/solutions/high-performance-computing)
- [Reinforcement learning](https://www.aguasto.com/solutions/reinforcement-learning)
- [Data science](https://www.aguasto.com/solutions/data-science)
- [Startups](https://www.aguasto.com/solutions/startups)
- [Enterprise](https://www.aguasto.com/solutions/enterprise)
- [Pricing](https://www.aguasto.com/gpu-cloud/pricing)
- [Availability](https://www.aguasto.com/gpu-cloud/availability)
- [Flex](https://www.aguasto.com/gpu-cloud/flex)
- [On-Demand](https://www.aguasto.com/gpu-cloud/on-demand)
- [Reserved capacity](https://www.aguasto.com/gpu-cloud/reserved)
- [Regions](https://www.aguasto.com/gpu-cloud/regions)
- [United States](https://www.aguasto.com/gpu-cloud/regions/united-states)
- [Europe](https://www.aguasto.com/gpu-cloud/regions/europe)
- [Mission](https://www.aguasto.com/company/mission)
- [Modular data centers](https://www.aguasto.com/company/modular-data-centers)
- [Technology](https://www.aguasto.com/company/technology)
- [Company roadmap](https://www.aguasto.com/company/roadmap)
- [Trust](https://www.aguasto.com/trust)
- [Security](https://www.aguasto.com/security)
- [Status](https://www.aguasto.com/status)
- [Support](https://www.aguasto.com/support)
- [Acceptable use](https://www.aguasto.com/legal/acceptable-use)
- [Data processing agreements](https://www.aguasto.com/legal/dpa)
- [Subprocessors](https://www.aguasto.com/legal/subprocessors)
- [Cookies](https://www.aguasto.com/legal/cookies)
- [REST API](https://www.aguasto.com/developers/api)
- [Quickstart](https://www.aguasto.com/developers/quickstart)
- [Authentication](https://www.aguasto.com/developers/authentication)
- [API keys](https://www.aguasto.com/developers/api-keys)
- [GPU catalogue data](https://www.aguasto.com/developers/gpu-catalog)
- [Availability](https://www.aguasto.com/developers/availability)
- [Pricing](https://www.aguasto.com/developers/pricing)
- [Deployments](https://www.aguasto.com/developers/deployments)
- [Instances](https://www.aguasto.com/developers/instances)
- [Usage](https://www.aguasto.com/developers/usage)
- [Billing and wallet](https://www.aguasto.com/developers/billing)
- [Webhooks](https://www.aguasto.com/developers/webhooks)
- [SDKs and CLI](https://www.aguasto.com/developers/sdks)
- [Errors](https://www.aguasto.com/developers/errors)
- [Rate limits](https://www.aguasto.com/developers/rate-limits)
- [Build with AI agents](https://www.aguasto.com/developers/agents)
- [MCP](https://www.aguasto.com/developers/mcp)
- [A2A](https://www.aguasto.com/developers/a2a)
- [GPU roadmap](https://www.aguasto.com/gpu-roadmap)
- [Research topics](https://www.aguasto.com/topics)
- [GPU systems & local clusters](https://www.aguasto.com/topics/gpu-systems)
- [Model serving & inference](https://www.aguasto.com/topics/model-serving)
- [Modular infrastructure & resources](https://www.aguasto.com/topics/modular-infrastructure)
- [Compute economics & evidence](https://www.aguasto.com/topics/compute-economics)
- [Glossary](https://www.aguasto.com/glossary)
- [Methodologies](https://www.aguasto.com/methodologies)
- [Blog](https://www.aguasto.com/blog)
- [How to choose a GPU topology for inference](https://www.aguasto.com/blog/choose-gpu-topology-for-inference)
- [How to verify GPU placement before booking](https://www.aguasto.com/blog/verify-gpu-placement-before-booking)
- [How to compare GPU quotes and reconcile usage](https://www.aguasto.com/blog/compare-gpu-quotes-and-usage)
- [How to write a GPU capacity request](https://www.aguasto.com/blog/write-a-gpu-capacity-request)
- [How to read a GPU hardware roadmap](https://www.aguasto.com/blog/read-a-gpu-hardware-roadmap)
- [Read the GPU catalogue from a script](https://www.aguasto.com/blog/read-the-gpu-catalogue-from-a-script)
- [Developers](https://www.aguasto.com/developers)
- [Resources](https://www.aguasto.com/resources)
- [GPU catalogue](https://www.aguasto.com/gpus)
- [Model Index](https://www.aguasto.com/models)
- [Inference Stack](https://www.aguasto.com/inference-stack)
- [Learn](https://www.aguasto.com/learn)
- [Insights](https://www.aguasto.com/insights)
- [Research](https://www.aguasto.com/research)
- [Industry cases](https://www.aguasto.com/research/cases)
- [Why Modular](https://www.aguasto.com/infrastructure/why-modular)
- [Cooling & Resource Efficiency](https://www.aguasto.com/infrastructure/cooling-resource-efficiency)
- [Partners](https://www.aguasto.com/partners)
- [GPU comparison](https://www.aguasto.com/gpus/compare)
- [Model comparison](https://www.aguasto.com/models/compare)
- [Engineering tools](https://www.aguasto.com/tools)
- [Development friction](https://www.aguasto.com/research/hyperscale-development-friction)
- [NVIDIA A100 40GB](https://www.aguasto.com/gpus/a100-40gb)
- [NVIDIA A100 80GB](https://www.aguasto.com/gpus/a100-80gb)
- [NVIDIA H100 PCIe](https://www.aguasto.com/gpus/h100-pcie)
- [NVIDIA H100 SXM](https://www.aguasto.com/gpus/h100)
- [NVIDIA GH200](https://www.aguasto.com/gpus/gh200)
- [NVIDIA H200 SXM](https://www.aguasto.com/gpus/h200)
- [NVIDIA H200 NVL](https://www.aguasto.com/gpus/h200-nvl)
- [NVIDIA B200 SXM](https://www.aguasto.com/gpus/b200)
- [NVIDIA B300](https://www.aguasto.com/gpus/b300)
- [NVIDIA GB200](https://www.aguasto.com/gpus/gb200)
- [NVIDIA GB300](https://www.aguasto.com/gpus/gb300)
- [NVIDIA GB200 NVL72](https://www.aguasto.com/gpus/gb200-nvl72)
- [NVIDIA GB300 NVL72](https://www.aguasto.com/gpus/gb300-nvl72)
- [NVIDIA Vera Rubin Platform](https://www.aguasto.com/platforms/nvidia-vera-rubin)
- [NVIDIA Rubin GPU](https://www.aguasto.com/gpus/nvidia-rubin)
- [NVIDIA Vera Rubin NVL72](https://www.aguasto.com/systems/nvidia-vera-rubin-nvl72)
- [NVIDIA DGX Rubin NVL8](https://www.aguasto.com/systems/nvidia-dgx-rubin-nvl8)
- [NVIDIA HGX Rubin NVL8](https://www.aguasto.com/systems/nvidia-hgx-rubin-nvl8)
- [NVIDIA Vera Rubin NVL4](https://www.aguasto.com/systems/nvidia-vera-rubin-nvl4)
- [NVIDIA Rubin CPX](https://www.aguasto.com/gpus/nvidia-rubin-cpx)
- [NVIDIA Vera Rubin NVL144 CPX](https://www.aguasto.com/systems/nvidia-vera-rubin-nvl144-cpx)
- [NVIDIA Rubin HPC](https://www.aguasto.com/gpus/nvidia-rubin-hpc)
- [NVIDIA Vera CPU](https://www.aguasto.com/components/nvidia-vera-cpu)
- [NVIDIA Groq 3 LPX](https://www.aguasto.com/components/nvidia-groq-3-lpx)
- [NVIDIA Vera BlueField-4 STX](https://www.aguasto.com/components/nvidia-bluefield-4-stx)
- [NVIDIA Spectrum-6 SPX Ethernet](https://www.aguasto.com/components/nvidia-spectrum-6-spx-ethernet)
- [AMD Instinct MI300X](https://www.aguasto.com/gpus/mi300x)
- [AMD Instinct MI325X](https://www.aguasto.com/gpus/mi325x)
- [AMD Instinct MI355X](https://www.aguasto.com/gpus/mi355x)
- [AMD Instinct MI455X](https://www.aguasto.com/gpus/mi455x)
- [AMD Helios](https://www.aguasto.com/gpus/helios)
- [NVIDIA L40](https://www.aguasto.com/gpus/l40)
- [NVIDIA L40S](https://www.aguasto.com/gpus/l40s)
- [NVIDIA RTX PRO 6000 Blackwell Server Edition](https://www.aguasto.com/gpus/rtx-pro-6000)
- [NVIDIA RTX 6000 Ada Generation](https://www.aguasto.com/gpus/rtx-6000-ada)
- [GLM-4.5](https://www.aguasto.com/models/glm-4-5)
- [Qwen3-235B-A22B](https://www.aguasto.com/models/qwen3-235b-a22b)
- [DeepSeek-R1](https://www.aguasto.com/models/deepseek-r1)
- [Llama 3.3 70B Instruct](https://www.aguasto.com/models/llama-3-3-70b-instruct)
- [Gemma 3 27B IT](https://www.aguasto.com/models/gemma-3-27b-it)
- [gpt-oss-120b](https://www.aguasto.com/models/gpt-oss-120b)
- [Kimi-K2-Instruct](https://www.aguasto.com/models/kimi-k2-instruct)
- [GLM-5.3](https://www.aguasto.com/models/glm-5-3)
- [Qwen3.8-27B](https://www.aguasto.com/models/qwen3-8-27b)
- [DeepSeek-V4.1-Flash](https://www.aguasto.com/models/deepseek-v4-1-flash)
- [Llama 4 Scout 17B-16E Instruct](https://www.aguasto.com/models/llama-4-scout-17b-16e-instruct)
- [Gemma 4 31B IT](https://www.aguasto.com/models/gemma-4-31b-it)
- [Kimi K3](https://www.aguasto.com/models/kimi-k3)
- [vLLM](https://www.aguasto.com/inference-stack/vllm)
- [SGLang](https://www.aguasto.com/inference-stack/sglang)
- [TensorRT-LLM](https://www.aguasto.com/inference-stack/tensorrt-llm)
- [NVIDIA Dynamo](https://www.aguasto.com/inference-stack/dynamo)
- [NVIDIA Triton Inference Server](https://www.aguasto.com/inference-stack/triton)
- [AMD ROCm](https://www.aguasto.com/inference-stack/rocm)
- [Ray Serve](https://www.aguasto.com/inference-stack/ray-serve)
- [Kubernetes GPU orchestration — NVIDIA device plugin](https://www.aguasto.com/inference-stack/kubernetes)
- [Model weights are only the beginning of the memory budget](https://www.aguasto.com/learn/model-weights-and-runtime-memory)
- [Why tokens per second is not enough to benchmark inference](https://www.aguasto.com/research/inference-benchmark-methodology)
- [Prefill, decode and the cost of separating them](https://www.aguasto.com/learn/prefill-decode-and-disaggregated-serving)
- [Choosing between vLLM, SGLang and TensorRT-LLM](https://www.aguasto.com/learn/choosing-an-inference-serving-framework)
- [From Power to Inference—Building a Distributed AI Compute Network](https://www.aguasto.com/learn/from-power-to-inference)
- [Dedicated compute and shared inference: compare the whole workload](https://www.aguasto.com/insights/dedicated-and-shared-inference-economics)
- [Distributed inference reliability begins with the failure boundary](https://www.aguasto.com/learn/distributed-inference-reliability)
- [Core Scientific: reading contracted, energized and billed capacity](https://www.aguasto.com/research/cases/core-scientific-capacity-disclosures)
- [Why Modular AI Infrastructure Matters Now](https://www.aguasto.com/research/why-modular-ai-infrastructure-matters-now)
- [What Is a Modular AI Data Center?](https://www.aguasto.com/learn/what-is-a-modular-ai-data-center)
- [How Modular AI Data Centers Are Built and Commissioned](https://www.aguasto.com/learn/modular-data-center-factory-to-commissioning)
- [Liquid Cooling for High-Density GPU Systems: Follow the Interfaces](https://www.aguasto.com/learn/liquid-cooling-gpu-rack-interfaces)
- [Why Electricity Availability Shapes AI Infrastructure Deployment](https://www.aguasto.com/research/electricity-availability-and-time-to-power)
- [What Is KV Cache? Capacity, Reuse and Isolation](https://www.aguasto.com/learn/kv-cache-capacity-and-policy)
- [Continuous Batching: Throughput Without Hiding Latency](https://www.aguasto.com/learn/continuous-batching-and-latency)
- [Applied Digital: Reading the Polaris Forge 1 Delivery Milestones](https://www.aguasto.com/research/cases/applied-digital-polaris-forge-capacity-milestones)
- [Crusoe Abilene: Separate Delivered Infrastructure from Expansion Plans](https://www.aguasto.com/research/cases/crusoe-abilene-delivery-and-expansion)
- [Data Center Cooling: Choosing Air, Liquid, Dry and Hybrid Architectures](https://www.aguasto.com/learn/data-center-cooling-air-liquid-dry-hybrid)
- [Can Data Centers Reduce Water Consumption? Define the Water Boundary First](https://www.aguasto.com/research/data-center-water-consumption-and-withdrawal)
- [Measuring PUE, WUE and AI Compute Efficiency at the Right Boundary](https://www.aguasto.com/learn/measuring-pue-wue-and-ai-compute-efficiency)
- [The Environmental Trade-offs of High-Density AI Infrastructure](https://www.aguasto.com/research/environmental-tradeoffs-and-infrastructure-reuse)
- [Selecting GPU Systems for Enterprise Inference](https://www.aguasto.com/learn/selecting-enterprise-inference-gpu-systems)
- [Networking Rack-Scale AI Systems: Define the Communication Boundary](https://www.aguasto.com/learn/rack-scale-ai-networking-boundaries)
- [Official LinkedIn](https://www.linkedin.com/company/aguasto)

This is a generated representation of the public HTML page. Diagrams are conceptual. GPU and inference services are planned and are not currently available to book or call. The contact page describes the available inquiry flow.
