Glossary

SlotTransactions a second (TPS)Landed and failedPriority feeLamport and micro-lamportCompute unit (CU)Jito tipVRAM (GPU memory)fp16, fp8 and int4Context (tokens)NOSMCP (Model Context Protocol)

Slot

Solana's turn to make the next block. A new slot starts about every 0.4 seconds.

For example: "Block time 0.42 s" means slots are running at about the normal pace.

Transactions a second (TPS)

How many transactions the whole Solana network processes each second.

For example: 1,200 TPS is a normal working day for Solana.

Landed and failed

A transaction lands when it makes it into a block. It can still fail inside the block (for example a trade whose price moved), and a failed one still pays its fee.

For example: "1 in 4 swaps failed" means a quarter of the swaps that reached a block did not go through.

Priority fee

An optional extra fee that moves your transaction ahead in line when blocks are busy. Wallets and bots set it in micro-lamports per compute unit.

For example: Setting "normal" on the fees page pays what the middle transaction paid lately.

Lamport and micro-lamport

A lamport is the smallest piece of SOL: one billionth. A micro-lamport is a millionth of a lamport, the unit priority fees are priced in.

For example: A priority fee of 10,000 micro-lamports per compute unit on a 200,000-unit swap is 2,000 lamports, a tiny fraction of a cent.

Compute unit (CU)

A measure of how much work a transaction asks the network to do. Bigger transactions use more units and pay more priority fee.

For example: A simple SOL transfer uses a few hundred units; a swap through several pools can use 200,000 or more.

Jito tip

A tip paid to the validator through Jito to get a transaction or bundle in fast. It is on top of the priority fee.

For example: "In a hurry" prices on the fees page include a tip.

VRAM (GPU memory)

The GPU's own memory. An AI model has to fit in it completely to run on that GPU.

For example: A 24 GB card holds a model of about 40 billion parameters at 4-bit.

fp16, fp8 and int4

How finely a model's numbers are stored. fp16 is full size, fp8 is half, int4 is a quarter. Smaller fits a cheaper GPU, with a small loss in quality.

For example: Llama 3.3 70B needs about 140 GB at fp16 but about 43 GB at int4.

Context (tokens)

How much text a model keeps in mind at once. A token is about three quarters of a word.

For example: 8,192 tokens is roughly 6,000 words.

NOS

Nosana's token. GPU rentals on Nosana are paid in NOS, straight from your wallet into Nosana's program.

For example: A one-hour rental of a small GPU deposits about 0.1 NOS.

MCP (Model Context Protocol)

A standard way for AI agents to use outside tools. With it, your own agent can call Shello's tools.

For example: In Claude Code, one line connects Shello.