How to choose a GPU, and a way to buy it
Two decisions. You do not need to know hardware: answer three questions on /start and we make both for you, with a price and a start date. This page explains what it does.
1. Which GPU, and how many
- Start from the job. Running a model for your users needs a few GPUs that answer requests. Fine-tuning or training a model needs many GPUs working together for weeks.
- Memory decides what fits. A model needs about 1.3 GB of GPU memory for each billion parameters to run (at FP8: Storing each number of a model in 8 bits. Half the memory of the usual 16 bits, nearly the same answers., with room to work). A 30B model needs about 39 GB, so a 48 GB L40 fits it and a 24 GB L40G does not.
- Running a model: one GPU per copy, behind an Endpoint: A web address (HTTPS) your app sends requests to. The supplier runs the GPU behind it; you never manage a server.. Start with two copies, so one can restart while the other answers, and add more as traffic grows. You pay by the hour, with no term.
- Training or fine-tuning: whole Server: One machine with 8 GPUs in it, sometimes called a node. of 8 GPUs, with fast links between them (InfiniBand: A very fast network between servers that lets them work as one. Needed to train across servers.), all on one Fabric island: A group of servers on the same fast network, so they can work on one job together. A training block sits on one island.. You reserve them for dates: a Window: Whole servers on one fabric island, reserved for you for a set range of dates. Training capacity is sold this way..
| GPU | Memory each | Usually for |
|---|---|---|
| L40G | 24 GB | Running models up to about 18B parameters, one GPU at a time |
| L40 | 48 GB | Running models up to about 36B parameters, one GPU at a time |
| H100 | 80 GB | Training and fine-tuning, in servers of 8; large models that need several GPUs at once |
| H200 | 141 GB | Training and fine-tuning, in servers of 8; large models that need several GPUs at once |
| B200 | 180 GB | Training and fine-tuning, in servers of 8; large models that need several GPUs at once |
| B300 | 288 GB | Training and fine-tuning, in servers of 8; large models that need several GPUs at once |
Not sure? Answer three questions and we pick the GPU and the count. Prices for each are on GPU prices.
2. Which way to buy
It depends on how much you need and for how long. Buy now and Request a block are the Open market: Spare GPUs that AI companies and other GPU owners list for weeks or a few months. Fewer servers, booked fast.: spare GPUs from AI companies and other GPU owners, for weeks or a few months. For bulk on a long contract, go Data-centre direct: Bulk GPUs on a long contract, usually a year or more, signed straight with a partner data centre. We find the options and run the paperwork; you pay the data centre.. In every way, you pay the supplier the price shown and Silicon Liquid never holds your money.
- Buy now: Order GPUs listed for sale at the price shown. We hold them for 12 hours while we confirm with the supplier, usually the same business day.
Buy now
Best for: GPUs to run a model, or up to 8 servers for up to 28 days, from what is listed today.
How it goes: Pick from the list and order in a minute. We Hold: GPUs kept out of anyone else's order while yours is confirmed: 12 hours on Buy now, 5 business days after an LOI. them for 12 hours while we confirm, then the supplier invoices you.
Usually running within 1 day of payment
See what is on sale → - RFQ: Request for quotes: you describe the block you need, suppliers send prices, you pick. We aim for three firm quotes within 24 hours.
Request a block
Best for: Whole servers for weeks or months, or anything that is not listed: a size, a place, a date.
How it goes: Describe it once. Suppliers send Firm quote: A price the supplier is bound to until the date on it. You can accept it as it stands., each with its total. You accept one and sign with the supplier.
Quotes within 24 hours; running in about 2 weeks
Request a block → - Data-centre direct: Bulk GPUs on a long contract, usually a year or more, signed straight with a partner data centre. We find the options and run the paperwork; you pay the data centre.
Data-centre direct
Best for: 16 servers or more for 12 months or more: a long contract only a data centre can fill.
How it goes: We show you 3 to 5 options without names, run the NDA: Non-disclosure agreement: both sides keep what they learn private, and agree not to go around Silicon Liquid for 24 months. and the paperwork, and test the servers before your contract starts.
Signed in about 6 weeks
Buy direct →
Planning ahead? The availability calendar shows what is listed for the next 26 weeks.
Let us choose for you
Three questions: what you want to do, how big, and for how long. You get one recommendation with an all-in monthly price, a start date and a named person.
Find the right GPUs →Words on this page are explained when you tap them, and all together in the glossary. How we check suppliers is on Trust; who pays what is on Fees.