Faciens Technologies · Quester GX-1

Quester logic lab

Four explorers of the mathematics inside the Quester GX-1 GPU architecture: Orientational Connective Logic, the time-dependent truth of TSAM, conflicts that only appear when parts of a design meet, and how area is shared between chiplets. Each one runs the same rules as the Quester software, in your browser.

Logic playground

In Orientational Connective Logic a statement is more than true or false. It also carries a context, an orientation and, in Quester’s use, the set of orientations it can ground and a load. Here four statements, p, q, r and s, each describe a stage of work. Set their properties, write a formula, and see its truth at each moment.

StatementTrue at t0 · t1 · t2OrientationContextCan groundLoad
The frame is re-orienting atNow is

Formula

Syntax: ~ & | -> <-> for not, and, or, implies, iff; P G C R L for Parallel, Ground, Context, Resist and Caelvorn between two formulas; S shift, A anchor (evaluate now); (…)@1 evaluate at t1; forall t. (…)@t and exists t. … over the three moments.

The five relations at

  • P Parallel: both true and identically oriented.
  • R Resist: both true and orientations opposed (negative inner product: three or four features differ).
  • G Ground: both true, same context, and the second can ground the first’s orientation.
  • C Context: the first is true and both share a context.
  • L Caelvorn: Resist, and the first carries more load.

Rules as in the Quester verification of the OCL manuscript, with the owner’s approved amendments for ground sets and load (A-1, A-2) and the literal reading of neutral values (DR-001). One world, three moments, orientations on the four-feature sphere.

Fault spread over time

TSAM gives every stage one of five values at each moment: not yet located (ε∞), in transition (κ rising towards its limit ρ), true (1), false (0), or returned to the boundary. Break a stage, move the deadline, or let a stage fail its accuracy check, and watch the consequences travel along the data flow.

Workload
StageTime slotDurationNumber formatBreak itRecoverNot servedFails accuracy
Deadline
not yet locatedin transitiontrueboundaryfalsedeadline
Inspect at time
StageValue nowκLimit ρAt the deadlineWhy

Rules as in Quester’s TSAM layer: a fault returns the boundary value and it spreads to every consumer; a stage that is never produced leaves its consumers unlocated; a stage that would finish at or after the deadline returns the boundary value; with recovery, a broken stage is rolled back and run again, taking twice as long. Stages in one time slot share it, and the slot lasts as long as its slowest stage. ρ is one minus the worst rounding error of a single operation in the stage’s format.

Hidden-conflict finder

A design can pass every check between two of its parts and still be impossible as a whole. Quester’s co-admissibility test finds such cases and names the smallest set of parts that cannot be built together. Choose an instance and switch stages in and out.

Instance

Every pair, checked on its own

Green: the two stages can be built together within the budget. The diagonal checks each stage alone.

The stages

StageCharacterTime slotLoadFeeds

Each verdict is the result of Quester’s design generator searching every organisation within the budget, computed in advance for every subset of stages. The random instances are from Quester’s test T-3, which found 35 such cases in 500.

Chiplet area planner

Quester’s proposal divides a GPU package among eight chiplet roles, each a face of the character cube. Divide the area yourself and see how the package compares with one uniform design of the same total area.

ChipletTakesFree featuresAreaShareShare × gainStages routed here

Rules as in Quester’s organisation model: a chiplet with f free features runs (1 + 4g) / (1 + g·f) times faster than a fully general one; each time slot lasts as long as its slowest chiplet, its memory traffic at 16 TB/s, or its traffic between chiplets at 7.2 TB/s, whichever is longest. A specialised chiplet can only beat the uniform design on work it holds alone if its share of the area times its gain exceeds one. Workloads, rates and memory as in the demonstrator; the gain is assumed.