Sizing Guides
37 articlesTiering Architecture for AI Data: Hot NVMe, Warm Capacity, Cold Object
A sizing guide for the three-tier AI data hierarchy: a hot TLC NVMe tier sized to the active working set, a warm QLC capacity tier, and cold object storage for…
CDUs Explained: Sizing Coolant Distribution for 50-200 kW AI Racks
A CDU is the pump, heat exchanger and control loop between your GPU cold plates and the building. Liquid-to-liquid versus liquid-to-air, in-rack, in-row and sidecar options, and why supply temperature…
Sizing On-Prem AI by Model Class: From 3B Laptops to 671B Servers
Pick the machine by the model class you intend to run and fine-tune, not by GPU brand. A tiered reference range from a 3B laptop to a 671B eight-GPU server,…
Fine-Tuning Capacity: What 8x RTX PRO 6000 Blackwell Plus aiDAPTIV+ Actually Runs
On 8x RTX PRO 6000 Blackwell (768 GB VRAM), a 70B FP32 fine-tune returns zero concurrent sessions on GPU alone. With aiDAPTIV+ NVMe offload the same box reports seven. Published…
Conversational Commerce in Indian Languages: GPU Sizing for Voice Retail
Voice and chat commerce in Indian languages runs a three-model pipeline per turn: speech recognition, language model, speech synthesis. Latency is the product requirement, and code-mixed Indian speech is where…
Multimodal RAG: GPU Planning for Document, Image and Video Retrieval
Most enterprise knowledge is in scanned documents, diagrams, screenshots and recordings, not clean prose. Multimodal RAG indexes those directly, and the GPU cost sits overwhelmingly in ingestion rather than in…
Local RAG on a Mid-Tier Workstation: Index Size and Honest Limits
A mid-tier workstation can host a genuinely useful private RAG system over a few hundred thousand documents. This sets out where the memory goes between generation model, embeddings and index,…
The Multi-Agent Developer Desk: Sizing a Mid-Tier Workstation
Developers now run several concurrent agent sessions rather than one chat. Concurrency multiplies KV cache rather than weights, which is why a 96 GB mid-tier workstation behaves very differently from…
Sizing the RL Fine-Tuning Loop: Rollouts, Verifiers and GPU Split
Reinforcement learning with verifiable rewards is now standard post-training, and it inverts the usual sizing assumption: most of the GPU budget goes to generating rollouts, not to gradient steps. That…
Digital Pathology at Scale: Whole-Slide Image AI GPU Planning
A whole-slide image is gigapixels, not megapixels, so pathology AI is a tiling and throughput problem before it is a model problem. Storage and scanning capacity usually constrain an Indian…
AI Drug Discovery: GPU Planning for Protein and Molecular Models
Structure prediction, docking and molecular dynamics have very different GPU profiles, and Indian pharma teams frequently size for the wrong one. This sets out where the compute actually goes across…
Imaging Foundation Models: GPU Sizing for Radiology in 2026
Radiology moved from single-finding algorithms to foundation models covering many conditions in one pass, with the first such device cleared in January 2026. One model replacing fourteen changes both the…
AI Restoration at Archive Scale: GPU Throughput Planning
Restoring a film archive is a throughput problem measured in frames, not files. A two-hour feature is around 172,800 frames, and a thousand-title library is billions. This shows how to…
Virtual Production and LED Volumes: Real-Time GPU Sizing
An LED volume is a hard real-time system: the wall must render camera-correct perspective every frame or the illusion breaks. This sets out how to size render nodes from wall…
Quality Inspection at Line Rate: Latency Budgets and Edge GPU Choice
Line-rate inspection is a deadline problem, not a throughput problem. Parts per minute sets a hard cycle time, and every stage from trigger to reject actuator must fit inside it.…
Time-Series Foundation Models for Predictive Maintenance: Sizing the Stack
Pretrained time-series models now forecast machine behaviour zero-shot, with reported throughput above 300 forecasts per second on a single GPU. That removes the per-asset training burden that stalled most Indian…
Vision AI on a Single 32 GB Card: What an Entry Workstation Handles
A 32 GB Blackwell-class workstation card runs detection, segmentation, OCR and mid-size vision-language models comfortably. This sets out realistic throughput expectations, where 32 GB stops being enough, and how to…
Insurance Claims Automation: GPU Planning for Document AI at Scale
Claims automation is a document AI problem before it is an LLM problem. Vision-language models now read scanned forms, prescriptions and estimates end-to-end, which changes GPU sizing: throughput is governed…
Fraud Detection at UPI Scale: GPU Sizing for Sub-100 ms Decisions
UPI processed 23.2 billion transactions in a single month of 2026, over 66 crore a day. Scoring that volume with deep models inside a sub-100 ms budget is a throughput…
Checkpoint Storage at Frontier Scale: Sizing for 2026-2028
Published work puts checkpoint overhead at 12-43 percent of total training time, with a 16,000-accelerator cluster taking roughly 155 checkpoints a day. This explains how to size checkpoint capacity, bandwidth…
Vector Index Sizing: From 10 Million to 1 Billion Embeddings
A billion 1536-dimension vectors in a plain HNSW index needs terabytes of RAM. Quantization cuts that by 4x to 32x, and disk-based indexes move the graph to NVMe. This is…
GPU-Accelerated Genomics: Sizing Secondary Analysis for Clinical Labs in 2026
Turning raw sequencer reads into variants is now GPU-bound work. NVIDIA Parabricks 4.6 with DeepVariant runs short-read whole-genome secondary analysis up to ~100x faster than CPU pipelines. How to size…
Media GenAI Workstation Sizing Guide for India
A media GenAI workstation in India must be sized to the heaviest concurrent task — video diffusion, audio synthesis, or multi-modal editing — not the average workload. GPU VRAM, NVMe…
Beyond SFT: Sizing DPO and RLVR Post-Training Infrastructure
The 2026 post-training recipe is SFT, then DPO, then RL with verifiable rewards. DPO doubles resident model copies; GRPO halved RL memory by dropping the critic, putting 7-32B reasoning training…
FP8 to FP4: How Low-Precision Training Reshapes Cluster Sizing
FP8 pretraining is the 2026 default and NVFP4 4-bit recipes are validated to 120B scale with FP8-matching accuracy, doubling arithmetic and halving memory on Blackwell-class silicon. Size clusters in tokens-per-day…
Demand Forecasting GPUs: Sizing for Retail and Quick Commerce
Demand forecasting is retrain-dominated: size GPUs for the nightly training window across millions of SKU-location series, not for serving. CPUs suffice below a million series; quick commerce forces 4-8 GPU…
GPU Sizing for E-commerce Recommendation and Personalization (2026)
E-commerce recommendation sizing in 2026 hinges on peak requests per second, a sub-100 ms latency budget, and embedding-table memory. Most Indian retailers need one L4/L40S-class inference server to a small…
GPU Server India Sizing Checklist for AI Inference
When sizing GPU servers for AI inference in India, consider memory capacity, workload requirements, and compliance with data protection regulations. The NVIDIA H200's 141 GB HBM3e memory and adherence to…
RAG Storage and Retrieval Sizing for Enterprise Search
Sizing RAG storage and retrieval for enterprise search requires careful consideration of GPU capabilities, data governance, and workload benchmarks. The NVIDIA H200's 141 GB HBM3e memory enhances data-center acceleration, while…
Fine-Tuning GPU Server Sizing for Enterprise LLMs
Fine-tuning infrastructure should be sized around dataset shape, experiment cadence, checkpoint strategy, and governance controls before GPU count. RDP GPU Mart can help Indian teams compare DRACO GPU server options…
Sizing 70B LLM Inference on GPU Servers in India
A 70B LLM inference plan should start with memory, concurrency, latency, power, and data-residency constraints, not only GPU count. For Indian teams, RDP GPU Mart can turn those constraints into…
GPU Sizing for Agentic AI Workloads (2026)
Agentic AI changes the sizing question: it isn't just a GPU problem. The GPU still runs the model, but the orchestration layer — scheduling sub-tasks, routing tool calls, passing state…
How Many GPUs Do You Need for a 100-User Private ChatGPT (On-Prem)?
A private, on-prem ChatGPT-style assistant for 100 users typically runs on one to two datacenter GPUs. Because "100 users" rarely means 100 people typing at once — active-to-total ratios are…
Sizing GPU Compute for Computer Vision / Video Analytics
Computer-vision and video-analytics GPU sizing is throughput-driven, not memory-driven: capacity is set by streams × resolution × frame rate × model complexity, and constrained by the GPU's video-decode and inference…
Right-Sizing a GPU Server for Enterprise RAG
In a Retrieval-Augmented Generation (RAG) stack, the generator LLM dominates GPU sizing — a 70B model needs ~1 GPU (H200) at FP8/4-bit — while the embedding model and vector database…
How Much VRAM Does a 70B / 405B LLM Need for Inference?
For inference, a 70B-parameter model needs ~140 GB of GPU memory at FP16, ~70 GB at FP8, and ~43 GB at 4-bit — so one to two datacenter GPUs. A…
How Many GPUs Do You Need to Fine-Tune a 70B LLM On-Prem?
Fine-tuning a 70-billion-parameter LLM on-premises needs roughly one GPU for QLoRA, two GPUs for standard LoRA, and eight or more for full FP16 fine-tuning. The deciding variable is the tuning…