Faciens Technologies · technical evaluation build

Quester GX-1 demonstrator

Describe a workload. The demonstrator characterises each stage as an orientation in Orientational Connective Logic, chooses its number format and representation, generates hardware organisations for it, rejects what cannot hold it, and runs it on the Quester GX-1 specification.

Analytical model. No silicon, no measurement.self-test …

Workload and assumptions

Start from a preset or edit the stages. Every result on this page updates as you type.

StageStepFormatOperationsGB touchedRegularForwardedFeedsDense product
Operations accept forms like 2.3e13. “Forwarded” is the share of touched bytes passed to another stage. “Auto” picks the cheapest format that meets the stage’s error tolerance.
Assumed gain from specialising hardware

The one free parameter of the design comparison. A fully specialised element is taken to be this many times better than the reference GPU. No measurement stands behind it. At 1.0× the uniform design cannot be beaten.

Memory available to every generated design

Character of each stage

Four signs per stage: number precision, access pattern, what bounds it, and locality. Stages with the same signs are Parallel and form one class. Stages whose signs differ on three or four counts Resist each other.

StageFormatCharacterPrecisionAccessBound byLocalityClassResists

Adaptive precision under an error constraint

A format is allowed only if its worst-case rounding error stays within the stage’s tolerance. Tighten the tolerance and the format widens; ask for the impossible and the stage is rejected.

FormatBitsWorst-case errorVerdict

Bound: γk·κ with γk = k·u/(1−k·u), u the unit roundoff, k the accumulation length (or the tree depth plus one). Worst case, so formats that work in practice for statistical reasons can fail it. An accumulation length of 1 means sums are held in a wider format.

Representation transformation

A dense matrix product can be rewritten as a low-rank, sparse or banded one when the data has that structure. The change is accepted only if the stage, conversion included, gets faster. An accepted change is applied to the workload above and can alter the stage’s character.

RepresentationOperationsGB touchedConversion shareTime on the reference GPUDecision

Generated designs

Speed against the uniform reference design

ceiling set by the assumed gainuniform design = 1.00

The fastest generated design

ElementServesFree signsArea

Admissibility over time

Each stage of the fastest generated design carries one of TSAM’s five values at every moment. Set a deadline, or break a stage and watch the failure reach everything that depends on it. With recovery on, a broken stage is rolled back and run again, at the cost of its time.

Deadline
Break a stage
not yet located (ε∞)in transition (κ rising)trueboundary returnfalsedeadline
Inspect at time
StageValue nowκLimit ρAt the deadlineWhy

Numerical invariant verification and recovery

This panel executes a real computation in your browser: a 32 × 32 matrix product in double precision, checked against a checksum identity that must hold up to rounding. Inject an error and see whether the check catches it and the transaction recovers.

Invariant: for every column j, Σi Cij = Σk (Σi Aik)·Bkj. Tolerance: 5·N·u·Σ|A||B|. An error smaller than the rounding tolerance cannot be told from rounding, and the check says so.

Quester GX-1 organisation, proposal v0.2

Eight heterogeneous chiplets. Areas were sized by search over 150 random workloads and scored on 400 others. An equal split was tested and dropped: on the six reference workloads it left half the compute area idle.

ChipletTakesFree signsAreaStages routed here now

Execution pathway chosen for each stage

StageChiplets able to take itChosenBecause

A proposed architecture. The mapping from each role to the four signs, the routing rule and the area floor of 2% are modelling choices. Designs are compared at equal total area, so the model cannot value the main practical reason for chiplets: a package of several dies holds more silicon than one die can be made with.

Quester GX-1 at its published targets

Here the published specification is the device. The rates and memory figures below are design targets, entered as the model’s inputs. The table shows what a device with those figures would do with your workload, and where its peak rates would and would not be reached.

SpecificationReference GPUQuester GX-1 targetRatio
StageReference GPUQuester targetFaster byBound byRate reachedOf its targetData placed inWith cache

Counting rules, to be confirmed: “120 PFLOPS sparse NVFP4” is read as NVIDIA reads its own “with sparsity” figures, the dense-equivalent rate at 2:4 structured sparsity, which is twice the operations performed. The other four targets are read as dense rates. A device without a format runs it at its next wider one. “Faster by” compares both devices without caches, because the reference GPU’s cache sizes are not public.

The published specification, line by line

Where each line of the Quester GX-1 announcement appears on this page, and in what sense.

Published lineOn this pageSense
runs live computed in your browser from your inputmodelled represented in the analytical modeldesign target an input, not a result

What the tests showed

Each test had a failure condition fixed before it ran. T-2c and T-4 were pre-registered and run once on fresh instances.

TestQuestionOutcomeResult

Independent cost check

The designs above assume that an element serving more kinds of work costs more per operation. On 15 November that assumption was tested against a component cost model built separately, with no geometric quantity in it. Its method and pass mark were fixed before the run, and the run was repeated independently with identical numbers.

pre-registered verdict Partial at 45 nm and not supported at 7 nm, against a rank-correlation bar of 0.6.

Cost never fell when a character was added to an element (25,440 checks). The count of signs on which characters disagree ranked designs better than precision count or pairwise distance, with a rank correlation of about 0.3 to 0.6 depending on set size. It falls short of the bar because the count stops distinguishing sets once they span all four signs.

Finding in the independent modelEnergy per operationArea per unit rate
Mixing compute-bound and memory-bound work in one element×1.18×1.87
Mixing high and reduced precision×1.11×1.37
Mixing regular and irregular access×1.10×1.39
Mixing local and forwarded output×1.05×1.16
Largest gain from specialising (tensor role against a fully general element)×1.71×2.86

A fully general element in that model costs ×1.64 in energy and ×3.26 in area against a specialised one. On this page’s gain slider that is a setting between 1.6× and 3.3×; the default is 2.0×. These gains are well below the 8× an equal split into eight chiplets would need. One value does not describe both energy and area, while this page uses a single value.

The cost model is analytical: about half of its 58 coefficients are estimates, three cited figures did not match their sources when checked, and no circuit was synthesised. An exploratory follow-up suggests the mean orientation of a set predicts this cost much better than the count does; that has not been tested on a separate model. Files and the verification record are in the package under records/external/argon_cost_test.

What the published targets require

Arithmetic on public figures, not a result of this model. The last column follows the gain slider above.

Published targetTimes the referencePower at reference efficiencyAt the assumed gain

Reference: NVIDIA H100 SXM, dense peak rates, 700 W (datasheet), a 2022 device, so these powers are upper estimates. Memory: JEDEC HBM4, 2 TB/s and 64 GB per stack (Semiconductor Engineering). For scale: NVIDIA’s Rubin GPU is announced at 50 PFLOPS NVFP4 inference with 288 GB and 22 TB/s of HBM4 (Tom’s Hardware); the Blackwell B200 at 9 PFLOPS dense FP4, 4.5 PFLOPS dense FP8, 40 TFLOPS FP64 and 1 kW (Tom’s Hardware).

Limits of this build