Local Hardware and Model Bounds Database

This is the interactive version of the local hardware writeup. It tracks Apple, NVIDIA, and AMD systems from 2020-2026 by complete-machine price band, memory capacity, bandwidth, and complete-system USD per GB.

The local copy adds model metadata and theoretical memory-side decode bounds: single-session, KV-aware aggregate, and loose Dennard memory-power ceilings.

Chart Explorer

Switch metric and price band. Hover for source rows; click a point to open its item page when the point comes from a specific device.

What This Tracks

The data tracks accelerator-accessible memory in one physical machine, grouped by complete-machine price band. It covers Apple, NVIDIA, and AMD from 2020 through 2026.

The core question is practical: what can someone buy or configure locally, and how do affordability, capacity, bandwidth, and vendor tradeoffs move over time?

How To Read The Symbols

U
Unified or coherent memory available to the accelerator.
V
VRAM on one discrete GPU.
Σ
Aggregate installed VRAM across multiple GPUs in one chassis.
*
Price band inferred from a documented component bill of materials.
Estimated, reconstructed, or bill-of-materials inferred price.

Aggregate VRAM is useful, but it is not the same thing as one unified memory pool. Treat aggregate points as chassis capacity, not a guarantee that one process can use it as one contiguous model memory space.

Main Read

Under $3,000, AMD changes the shape of the local market in 2025 and 2026 with 128GB unified memory through Framework Desktop-class hardware. Apple is strong on usable unified memory, but the current sub-$3,000 Apple ceiling in this dataset is 64GB. NVIDIA stays at 24GB in this band because the data tracks complete machines, not used cards or component-only builds.

In the $3,000-$5,000 band, both AMD and NVIDIA reach 128GB unified or coherent memory by 2025-2026. This is the band that matters most for consumer and prosumer local model work because it is expensive but still within a serious personal hardware budget.

In the $5,000-$10,000 band, Apple has the most dramatic historical point: the 512GB M3 Ultra Mac Studio launch configuration in 2025. The best-so-far memory charts keep that 512GB point through 2026 because the historical record should not go down just because the current Apple configuration changed.

Above $75,000, NVIDIA dominates the memory and bandwidth records. DGX Station-class systems are local in the physical sense, but not local in the affordability sense.

Local Hardware Database

A TechPowerUp-style catalog for complete local systems and accelerator configurations, including Apple unified-memory machines.

Method

The editable source of truth is data/items/*.json. The generator expands each item price history into CSVs, derives annual and best-so-far records, and renders the Space from those generated artifacts.

The complete-system price-per-memory metric is total configured machine price divided by accelerator-accessible memory. The extracted component memory-price file is separate and should not be confused with complete-machine affordability.

The projection view fits simple least-squares linear trends to 2020-2026 best-so-far memory series. It is a scenario view, not verified future product data.