Mochi’s pocket notebook
One small explanation. A story when you want more.
Context window
The working space a model can use for one response.
Read Mochi’s illustrated story →GPU
A graphics processing unit: a chip that performs the same operation on many data values in parallel.
Read Mochi’s illustrated story →HBM
High-bandwidth memory: stacked memory for moving lots of data to a processor.
Read Mochi’s illustrated story →Inference
A trained model generating an output from a new input.
Read Mochi’s illustrated story →MCP
Model Context Protocol: a shared standard for connecting AI apps to outside information and tools.
Read Mochi’s illustrated story →NPU
A neural processing unit: a chip tailored to accelerate AI tasks and workloads.
Read Mochi’s illustrated story →Photonics
Moving data with light: using light pulses to transmit information.
Read Mochi’s illustrated story →PUE
Power usage effectiveness: a ratio that compares a data center’s non-computing overhead energy, like cooling and power distribution, with the energy its computers use.
Read Mochi’s illustrated story →RAG
Retrieval-augmented generation: the AI retrieves relevant information from outside documents first, then grounds its answer in it.
Read Mochi’s illustrated story →SOCAMM
Small outline compression attached memory module: low-power mobile memory on a small module for data center servers.
Read Mochi’s illustrated story →Tokens
The small units of text a language model reads and writes.
Read Mochi’s illustrated story →TPU
A tensor processing unit: Google’s custom chip for large matrix operations.
Read Mochi’s illustrated story →VRAM
Video random access memory: the high-speed memory available directly to the GPU, where a model’s weights have to fit.
Read Mochi’s illustrated story →Capital spending
Cash spent buying long-lived assets, such as buildings or equipment.
Read Mochi’s illustrated story →Price and volume
Two different ingredients in sales: what each unit earns and how many units sell.
Read Mochi’s illustrated story →