OMNAR
How to write a GPU server specification
The workload, operating requirements and acceptance evidence needed to compare proposals before placing an order.
Read the guide: How to write a GPU server specificationOMNAR / Knowledge centre
Understand before you order.
Practical guides for planning, selecting and deploying GPU infrastructure.
How much GPU memory does a language model need?
Separate weights, KV cache and operating headroom, account for MoE, and validate the estimate against a real workload.
Read the guide: How much GPU memory does a language model need?Buy or rent a GPU server
Compare both options against one workload: utilisation, duration, control and the complete cost.
Read the guide: Buy or rent a GPU server
Prepare your site for a GPU server
From rack space and power to access, software versions and acceptance: agree the essentials before delivery.
Read the guide: Prepare your site for a GPU serverOne million tokens: budgeting for long context
Calculate KV memory, distinguish MHA, GQA and MLA, and account for several long requests running together.
Read the guide: One million tokens: budgeting for long contextHGX, NVL and NVL72: understanding GPU connectivity
Why GPU count and total memory do not establish communication speed: NVLink domains, CPU RAM and networking between nodes.
Read the guide: HGX, NVL and NVL72: understanding GPU connectivityHow to compare international GPU server prices
Check the sale unit, currency, taxes, shipping and configuration to compare offers without inventing missing costs.
Read the guide: How to compare international GPU server prices