HPC & AI Mining Glossary

High-Performance Computing: computing performed on systems designed to process scientific, industrial or engineering workloads quickly.

A set of methods that enable a system to perform tasks such as classification, content generation or decision-making.

Learning automatically from data, without explicitly programming every decision rule.

A subfield of machine learning that uses neural networks with many layers.

The phase in which a model learns from a dataset; it often requires a lot of compute and memory.

Using an already trained model to produce a prediction or a response.

A mathematical structure trained to perform a specific task.

An internal value learned by a model. The number of parameters affects, among other things, the memory and resources required.

A general-purpose processor suited to a wide range of sequential computations and control tasks.

A massively parallel processor used for graphics rendering, AI and some HPC workloads.

An integrated circuit designed for one specific task. A mining ASIC runs a given algorithm efficiently but cannot replace a general-purpose GPU.

A reconfigurable circuit whose hardware logic can be adapted to a particular workload.

A specialized component that offloads specific computations from the CPU, such as a GPU or an ASIC.

A GPU’s onboard memory. Its capacity and bandwidth limit the size of the models and data batches it can process.

The amount of data memory can transfer per second. It can become the limiting factor before compute power does.

The link between accelerators, servers or racks. Its latency and throughput have a major impact on distributed workloads.

A computer system designed to provide compute, storage or network resources on a continuous basis.

An individual machine that is part of an HPC or AI cluster.

A group of coordinated nodes that handles larger workloads or provides higher availability.

A standardized enclosure that houses servers, networking equipment and power distribution.

Power Distribution Unit: equipment that distributes electrical power to the devices in a rack or an installation.

Duplicating critical components, such as power or networking, to reduce outages.

The time needed to transmit information or get a response.

The amount of data transferred between systems per unit of time.

The instantaneous demand of a piece of equipment. It determines how circuits, circuit protection, cables and PDUs are sized.

The energy used over a period of time. It is used to calculate the actual electricity cost.

Power Usage Effectiveness: the ratio of a site’s total energy use to the energy used by its IT equipment. A value close to 1 means lower overhead losses.

Removing heat with airflow, fans and an air-conditioning or exhaust system.

Transferring heat through a fluid circulating in cold plates or a dedicated loop.

Cooling equipment by submerging it in a compatible dielectric fluid.

The amount of heat concentrated in a given area or volume. It influences the choice of cooling.

Reusing the heat produced to warm a building, a fluid or a process.

The share of time a system remains available and operational.

Tracking performance, temperatures, power consumption, errors and availability status.

Measurements sent by equipment to a monitoring or alerting system.

The set of computations and data assigned to an infrastructure.

Spreading a workload across multiple resources to avoid bottlenecks.

A tool that schedules, deploys and supervises tasks or services across multiple machines.

Service Level Agreement: a measurable commitment to availability or service level.

Total Cost of Ownership: the full cost, including purchase, energy, cooling, maintenance, networking and operations.

A consensus mechanism in which machines perform verifiable computations to secure a blockchain network.

The number of hash calculations performed per second by a miner or a network.

The difference between estimated mining revenue and costs, including electricity, hosting and hardware purchases.

Performance measured relative to the power consumed. The relevant unit depends on the workload and the hardware.

Equipment optimized for a defined task. Higher efficiency generally comes at the cost of versatility.

Putting equipment to a different use. It is limited by the architecture: a mining ASIC generally cannot be converted into a GPU server.

Upfront capital expenditures, such as buying machines, racks, transformers and cooling systems.

Recurring operating expenses, including electricity, connectivity, maintenance, staff and hosting.