Skip to main content

Loading...

Home/Glossary

Reference

GPU Rental
Glossary

Rigorous definitions for the 36 terms that shape GPU rental market analysis — precision formats, memory technologies, compute units, and category boundaries.

GPU Categories

3 terms in this section.

GPUs designed primarily for gaming and personal workstations.

Server-grade GPUs designed for sustained workloads in cooled, racked environments.

Workstation GPU

GPUs in between consumer and datacenter tiers, used in professional creative and engineering settings.

Numerical Precision

4 terms in this section.

BF16 (bfloat16)

Alternative 16-bit floating-point format with wider exponent range than FP16.

16-bit half-precision floating-point format used heavily in deep learning.

FP32 (Single Precision)

32-bit floating-point format — the traditional default precision for scientific computing.

8-bit floating-point format introduced for high-throughput inference and training.

Memory

2 terms in this section.

HBM (High Bandwidth Memory)

Stacked DRAM technology used for high-throughput GPU memory in datacenter parts.

VRAM (Video RAM)

On-package GPU memory that holds model weights, activations, and intermediate tensors during compute.

Compute Units

2 terms in this section.

General-purpose parallel processors in NVIDIA GPUs.

NVIDIA's specialized matrix-multiplication accelerators used heavily in deep learning.

Architecture

4 terms in this section.

Attention (Self-Attention)

The mechanism that lets transformer models weigh which parts of the input matter most for each output.

Diffusion Model

Generative model architecture that produces images, audio, or video by iteratively denoising from random noise.

Memory caching technique that stores transformer attention keys and values to avoid recomputing them at every token generation step.

Neural network architecture that underpins modern large language models and many vision models.

Interconnect

3 terms in this section.

NVIDIA's high-bandwidth GPU-to-GPU interconnect for multi-GPU systems.

The standard expansion bus used to connect GPUs to host systems.

NVIDIA's high-density socket form factor for datacenter GPUs.

Hardware

1 term in this section.

High-bandwidth, low-latency interconnect used for multi-node GPU training in HPC environments.

Performance Metric

5 terms in this section.

End-to-end time between a request and a complete response.

Memory Bandwidth

How fast a GPU can read and write its VRAM, measured in gigabytes per second.

Trillion floating-point operations per second — the standard GPU compute throughput metric.

How many operations (tokens, images, queries) a system processes per second.

TTFT (Time to First Token)

Latency from receiving an LLM prompt to producing the first output token.

Workload Concept

9 terms in this section.

Distributed Training

Training large models across multiple GPUs or machines simultaneously.

Serving trained large language models to user requests, often memory-bandwidth-bound.

Compute-intensive process of training large language models from scratch or via continued pre-training.

LoRA (Low-Rank Adaptation)

Parameter-efficient fine-tuning method that adds small trainable matrices to a frozen base model.

Mixed Precision

Training technique using FP16/BF16 for most operations while keeping critical paths in FP32.

PEFT (Parameter-Efficient Fine-Tuning)

Umbrella term for techniques that fine-tune large models by updating only a small fraction of parameters.

QLoRA (Quantized LoRA)

LoRA fine-tuning combined with 4-bit quantization of the base model — dramatically reduces VRAM needs.

Reducing the numerical precision of model weights to lower memory and compute requirements.

RAG (Retrieval-Augmented Generation)

Architecture pattern that grounds LLM responses in retrieved documents from a knowledge base.

Pricing Model

3 terms in this section.

On-Demand Pricing

Standard pay-as-you-go hourly rate with no upfront commitment.

Reserved Instance

Discounted hourly rates obtained by pre-committing to multi-month or multi-year usage.

Discounted hourly rates with the marketplace's right to reclaim capacity at short notice.

See these concepts in action

Every fit-analysis page on AIMC uses these definitions — VRAM, FP16 TFLOPS, GPU class — to score GPU suitability for specific workloads.

Browse GPUs Read methodology

The definitive source for GPU rental market data.

Price•Power•Carbon

Product

Home
AIMC Market Intelligence
AIMC GPU Global Map
AIMC Market Index
GPU Watchlist
GPU Cost Calculator
GPU Marketplace Directory
GPU Hyperscaler Reference
GPU Power Tracking
GPU Carbon Tracking
GPU Comparisons
GPU Use Cases
GPU Prices
Our Pricing

Company

About
Methodology
Institutional License
AIMC Insights
AIMC Press
Data Coverage
Our Data Partners
Why Marketplaces Join
Our Security
FAQ
Contact

Dashboards

AIMC Market Index
AIMC Market Intelligence
AIMC GPU Global Map

Your Account

Account
API Docs
API Keys
API Usage
GPU Watchlist
Settings
Log In

© 2026 AIMC Technologies, LLC. All rights reserved.

Terms Privacy SLA