New DirectX and lower memory use: what will actually change in games?

Primary sources: the Agility SDK announcement of October 28, 2026 and the Clustered Geometry and PTLAS technical explanation and test. This is an independent analysis of that release. The figures below come from Microsoft’s sample, not our own tests or measurements across retail games.

“DirectX cuts graphics memory use” is an appealing headline. But which memory, in which scene, compared with what? Those questions separate an important engineering advance from a claim mistakenly applied to every game.

First identify what was released

The announcement distinguishes retail Agility SDK 1.620 from 1.722 preview. New Clustered Geometry and Partitioned TLAS features belong to the preview branch. An experimental capability is not automatically available in all shipped games. Agility SDK is a development and runtime-distribution tool: downloading a file as a player does not rewrite every older game engine.

Why do ray-tracing structures matter?

Engines use acceleration structures to find ray intersections with geometry. Building or updating those structures has its own cost in a changing world. The clustered approach works with smaller, reusable pieces. PTLAS concentrates update work on changed scene regions. Applications can use the features separately or together.

My interpretation is that the main opportunity is less repeated work and new engine designs. But an engine may spend the saved capacity on richer geometry. Greater efficiency therefore does not necessarily mean lower final consumption under different visual settings.

What does the official test show?

Microsoft compares traditional BLAS rebuilds with clustered geometry in one configuration of its experimental sample on an AMD RX 9070:

Total structure memory falls by 89.7 percent in this example; the animated portion falls by 96.4 percent. Build scratch, however, increases. The source explicitly cautions against expecting gains of this magnitude in games. This is a test playground compared with traditional rebuilding, not average results across commercial titles.

Why is this not an equivalent reduction in total VRAM?

Graphics memory also contains textures, render targets, geometry, buffers and other resources. The Direct3D 12 memory-management guidance discusses budgets and residency: one allocation number does not describe the entire situation.

In a purely hypothetical case, if acceleration structures are a small share of a game’s memory, a large saving there produces a smaller overall saving. If the engine also increases geometric detail, total consumption might remain unchanged. Compare the same scene and image quality, and record peak use rather than only a quiet instant.

Questions PC players should ask

Will my low-memory card stop running out of VRAM?

That has not been established. If textures or output settings cause the limit, changing ray-tracing structures may not resolve it. Identify the game, engine support and actual source of pressure first.

Will a Windows update alone increase FPS?

There is no guarantee. Adoption needs suitable game or engine code, the relevant runtime and a supporting driver. A “DirectX 12” label in dxdiag does not show that every new feature is active in a game.

Will every ray-tracing card support it?

The features introduce no new hardware capability requirement beyond existing ray tracing, but vendor implementation and driver validation determine availability by generation. Base functionality does not guarantee support for every model today. Consult the release’s feature table and driver notes.

Where does Advanced Shader Delivery fit?

ASD concerns delivering precompiled shaders and can reduce startup waits and compilation stutter. That is different from reducing acceleration-structure memory. Download size, disk-cache capacity, RAM and VRAM are not interchangeable measures either.

Are the new debugging tools cost-free at runtime?

DirectX Dump Files help diagnose GPU crashes. Higher-overhead diagnostic modes can consume more memory or reduce frame rate. Tests using those modes should be reported separately from ordinary play.

What would a credible comparison look like?

Record game, Windows, driver, GPU and settings. Repeat a fixed route; separate cold and warm caches. Examine frame times, noticeable drops and memory peaks alongside average FPS. Disclose any quality or settings changes. A smaller number in one tool does not establish that all resource use fell.

My assessment

Treat this release as a serious opportunity for future engines, not a universal PC upgrade. Reducing redundant rebuilding and managing changing scenes more efficiently are promising. Purchasing claims should follow real-game tests at equal quality, rather than percentages from a demonstration.

Related news: the new Agility SDK release.

Sources checked: November 2, 2026. No independent benchmark, universal FPS promise or specific GPU purchase recommendation is offered. The illustration is conceptual.