spectral

as donto believed on

From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-07-03.)

spectral has 106 facts recorded in Dontopedia across 13 references, with 6 live disagreements.

106 facts·85 predicates·13 sources·6 in dispute

Mostly:has ppl at step(7), rdf:type(6), has tokens per second(2)

Inbound mentions (53)

Other subjects in dontopedia point AT this entity as a value. These are inverse relationships — e.g. "X motherOf this subject" — and answer questions the forward facts can't. Grouped by predicate.

`comparedTo`Compared to(2)

Ffn Dftex:ffn-dft
Topo Anchor Kan Quality Advantageex:topo-anchor-kan-quality-advantage

`hasFewerParametersThan`Has Fewer Parameters Than(2)

Anchor Kanex:anchor_kan
Lohe Spherical Speed Advantageex:lohe-spherical-speed-advantage

`believesSpectralBetterLm`Believes Spectral Better Lm(1)

Xenonfunex:xenonfun

`comparesPpl`Compares Ppl(1)

Row 1ex:row-1

`comparesTokPerS`Compares Tok Per S(1)

Row 2ex:row-2

`containsModel`Contains Model(1)

Code Blockex:code-block

`describesForwardThroughput`Describes Forward Throughput(1)

Lisamegawatts Message 18 58ex:lisamegawatts-message-18-58

`evaluatesAsBetterLanguageModel`Evaluates As Better Language Model(1)

Key Takeawayex:key-takeaway

`exhibitsMuchRicherSpecializationThan`Exhibits Much Richer Specialization Than(1)

Ffn Dftex:ffn-dft

`expertType`Expert Type(1)

Latentmoeex:latentmoe

`fasterThan`Faster Than(1)

Lohe Spherical Speed Advantageex:lohe-spherical-speed-advantage

`follows`Follows(1)

Anchor Kanex:anchor-kan

`has13PercentFewerParamsThan`Has13 Percent Fewer Params Than(1)

Lohe Sphericalex:lohe-spherical

`has19PercentSpeedAdvantage`Has19 Percent Speed Advantage(1)

Anchor Kanex:anchor-kan

`has27PercentFewerParamsThan`Has27 Percent Fewer Params Than(1)

Lohe Sphericalex:lohe-spherical

`hasHighestRIn`Has Highest R in(1)

Block L2ex:block-l2

`hasMember`Has Member(1)

Code Blockex:code-block

`hasRicherSpecialization`Has Richer Specialization(1)

Ffn Dftex:ffn-dft

`hasStrongDcSyncHub`Has Strong DC Sync Hub(1)

Row 4ex:row-4

`hasSuperiorQuality`Has Superior Quality(1)

Topo Anchor Kanex:topo-anchor-kan

`hasSyncHubRoleForConfig`Has Sync Hub Role for Config(1)

Block 10ex:block-10

`hasWeakerSync`Has Weaker Sync(1)

Lohe Sphericalex:lohe-spherical

`haveParamCountsNearby`Have Param Counts Nearby(1)

Other Attention Typesex:other-attention-types

`improvesSpeed`Improves Speed(1)

Parallel Prefillex:parallel-prefill

`includesModel`Includes Model(1)

Test Execution Planex:test-execution-plan

`involvesComponent`Involves Component(1)

Step Compare Form1ex:step-compare-form1

`is120xFaster`Is120x Faster(1)

Parallel Prefillex:parallel-prefill

`is25PercentFasterThan`Is25 Percent Faster Than(1)

Lohe Sphericalex:lohe-spherical

`is47PercentFasterThan`Is47 Percent Faster Than(1)

Lohe Sphericalex:lohe-spherical

`isConsistentWithBetterPpl`Is Consistent With Better Ppl(1)

Row 5ex:row-5

`isFaster`Is Faster(1)

Anchor Kanex:anchor-kan

`isHigherIn`Is Higher in(1)

Thermodynamic Betaex:thermodynamic-beta

`isMoreOrderedAtSameCoupling`Is More Ordered at Same Coupling(1)

Row 5ex:row-5

`isNotablyFlatterThan`Is Notably Flatter Than(1)

Beta Gate of Run 2ex:beta-gate-of-run-2

`isOutputOf`Is Output of(1)

Spectral Output Sampleex:spectral-output-sample

`isPatternOf`Is Pattern of(1)

Word Like Runsex:word-like-runs

`isSparser`Is Sparser(1)

Anchor Kanex:anchor-kan

`isStrictlyBetterThan`Is Strictly Better Than(1)

Lohe Sphericalex:lohe-spherical

`measuredFor`Measured for(1)

Spectral Throughput Measurementex:spectral-throughput-measurement

`measuresDifferenceBetween`Measures Difference Between(1)

Perplexity Gapex:perplexity-gap

`mentions`Mentions(1)

Source Utteranceex:source-utterance

`overtakes`Overtakes(1)

Lohe Sphericalex:lohe-spherical

`property`Property(1)

Audio Visual Representationsex:audio-visual-representations

`reportsConvergence`Reports Convergence(1)

Row 1ex:row-1

`reportsHigherThermodynamicBeta`Reports Higher Thermodynamic Beta(1)

Row 5ex:row-5

`reportsSysR`Reports Sys R(1)

Row 2ex:row-2

`representsSameSignal`Represents Same Signal(1)

Anchor Kanex:anchor-kan

`risesIn`Rises in(1)

Sys Rex:sys_r

`slowerDueTo5DTensors`Slower Due To5 D Tensors(1)

Anchorkanex:anchorkan

`tendedTowardBy`Tended Toward by(1)

Run on Noun Phrasesex:run-on-noun-phrases

`underperforms`Underperforms(1)

Lohe Sphericalex:lohe-spherical

Other facts (101)

The long tail: predicates that appear too rarely to warrant their own section. Filter or scroll to find a specific one. Each row links to its source.

101 facts

Predicate	Value	Ref
Has Ppl at Step	1140	[8]
Has Ppl at Step	561	[8]
Has Ppl at Step	501	[8]
Has Ppl at Step	373	[8]
Has Ppl at Step	260	[8]
Has Ppl at Step	219	[8]
Has Ppl at Step	239	[8]
Rdf:type	Model Component	[9]
Rdf:type	Attention Configuration	[10]
Rdf:type	Model Architecture	[11]
Rdf:type	Attention Type	[12]
Rdf:type	Model Kind	[13]
Rdf:type	Model	[13]
Has Tokens Per Second	12815	[6]
Has Tokens Per Second	12815	[12]
Has Loss	9.02	[6]
Has Loss	9.02	[12]
Has Param Count	107.2M	[6]
Has Param Count	107200000	[12]
Achieved Throughput	43700	[11]
Achieved Throughput	46300	[11]
Throughput Unit	tok/s	[13]
Throughput Unit	tokens per second	[13]
Better in Training Than	Pytorch Mps	[1]
Peaks at Forward Throughput	365000	[1]
Is Four Times Faster Than	Pytorch Mps	[1]
Has Training Throughput	84800	[1]
Has Latency Scaling	7.9	[1]
Exhibits Scaling	Super Linear Ish	[1]
Presupposes Competitor to Gpt	{}	[2]
Has Essential Stability	Memory Standard	[2]
Outperforms	Gpt	[2]
Was Stable	Memory Standard	[2]
Reaches Same DC Sync Strength As	Anchor Kan	[3]
Has Tok S	12000	[3]
Has R Hub	0.56	[3]
Has Sync Hub	Block 10	[3]
Sync Hub Precedes	Anchor Kan	[3]
Has DC Mode	0.315	[3]
Tends Toward Run on Noun Phrases	true	[4]
Lacks5 D Tensor	true	[4]
Faster Than Anchorkan	Anchorkan	[4]
References in Comparison	Anchorkan	[4]
Outputs Longer Word Like Runs	"The not used in the be 2001 3 -5 of a same"	[4]
Has Speed	~400 tok/s	[4]
Includes Numbers and Partial Structure	true	[4]
Has Dominant H0 in L2	true	[5]
Forms Sync Hub	L2	[5]
Has26 Percent Ppl Gap	Anchor Kan	[5]
Has DC Synchronization Hub	L2 H0	[5]
Is More Ordered	true	[5]
Reaches Higher Final Order	Anchor Kan	[5]
Has Smoother Convergence	Anchor Kan	[5]
Has Higher Beta	Anchor Kan	[5]
Optimizes Ppl	true	[5]
Implements True Rope in Qk	true	[5]
Has Ppl Per Token	0.64	[6]
Has Same Model Quality As	Lohe Spherical	[6]
Converges Identically to	Lohe Spherical	[6]
Has Param Count Approx	17M	[6]
Has Ppl	8228	[6]
Has Best Quality	true	[7]
Has Ppl	6.03	[7]
Has Tokens Per Second at Temp256	210000	[7]
Has Quality Adjusted Tokens Per Second	8008	[7]
Has Tok Per S	56931	[8]
Has Final Ppl	239	[8]
Has Params	16.9M	[8]
Second Best in Ppl	null	[8]
Uses	Funk Hecke Decomposition	[8]
Has Sync Hub Location	Block 10	[10]
Has Sync Strength	0.56	[10]
Has Baseline Throughput	12000	[10]
Has Throughput Unit	Tok S	[10]
Achieved Perplexity	116.4	[11]
Has Speedup Factor	120	[11]
Benefited From	Parallel Prefill	[11]
Compatible With	Rope Py	[11]
Uses Rope Mode	True Rope Qk	[11]
Validation Perplexity	116.4	[11]
Converges Smoothly	true	[11]
Has Final Order	0.36	[11]
Phase State	disordered	[11]
Classified As	Better Language Model	[11]
Has Perplexity	8228	[12]
Has Ppl Per Token	0.64	[12]
Outputs Pattern	Word Like Runs	[13]
Outputs Sample Text	Spectral Output Sample	[13]
Tends Toward Output Pattern	Run on Noun Phrases	[13]
Described Speed	Faster	[13]
Approximate Throughput	400	[13]
Output Length	longer	[13]
Appears in	Code Block	[13]
Outputs Sample	Spectral Output Sample	[13]
Precedes	Anchor Kan	[13]
Output Characteristic	longer word-like runs	[13]
Example Output	The not used in the be 2001 3 -5 of a same	[13]
Includes	numbers and partial structure	[13]
Tends Toward	run-on noun phrases	[13]
Performance	faster	[13]

Timeline axis is valid_time — when each source says the fact was true in the world, not when Dontopedia learned about it. Retracted rows are kept for provenance; coloured stripes indicate the context kind.

106 undated

—

betterInTrainingThanblah/watt-activation/part-104

ex:pytorch-mps

—

peaksAtForwardThroughputblah/watt-activation/part-104

365000

—

isFourTimesFasterThanblah/watt-activation/part-104

ex:pytorch-mps

—

hasTrainingThroughputblah/watt-activation/part-104

84800

—

hasLatencyScalingblah/watt-activation/part-104

7.9

—

exhibitsScalingblah/watt-activation/part-104

ex:super-linear-ish

—

presupposesCompetitorToGptblah/watt-activation/part-125

{}

—

hasEssentialStabilityblah/watt-activation/part-125

ex:memory-standard

—

outperformsblah/watt-activation/part-125

ex:gpt

—

wasStableblah/watt-activation/part-125

ex:memory-standard

—

reachesSameDcSyncStrengthAsblah/watt-activation/part-221

ex:anchor-kan

—

hasTokSblah/watt-activation/part-221

12000

—

hasRHubblah/watt-activation/part-221

0.56

—

hasSyncHubblah/watt-activation/part-221

ex:block-10

—

syncHubPrecedesblah/watt-activation/part-221

ex:anchor-kan

—

hasDcModeblah/watt-activation/part-221

0.315

—

tendsTowardRunOnNounPhrasesblah/watt-activation/part-232

true

—

lacks5DTensorblah/watt-activation/part-232

true

—

fasterThanAnchorkanblah/watt-activation/part-232

ex:anchorkan

—

referencesInComparisonblah/watt-activation/part-232

ex:anchorkan

—

outputsLongerWordLikeRunsblah/watt-activation/part-232

"The not used in the be 2001 3 -5 of a same"

—

hasSpeedblah/watt-activation/part-232

~400 tok/s

—

includesNumbersAndPartialStructureblah/watt-activation/part-232

true

—

hasDominantH0InL2blah/watt-activation/part-236

true

—

formsSyncHubblah/watt-activation/part-236

ex:l2

—

has26PercentPplGapblah/watt-activation/part-236

ex:anchor-kan

—

hasDcSynchronizationHubblah/watt-activation/part-236

ex:l2-h0

—

isMoreOrderedblah/watt-activation/part-236

true

—

reachesHigherFinalOrderblah/watt-activation/part-236

ex:anchor-kan

—

hasSmootherConvergenceblah/watt-activation/part-236

ex:anchor-kan

—

hasHigherBetablah/watt-activation/part-236

ex:anchor-kan

—

optimizesPplblah/watt-activation/part-236

true

—

implementsTrueRopeInQKblah/watt-activation/part-236

true

—

hasPPLPerTokenblah/watt-activation/part-317

0.64

—

hasSameModelQualityAsblah/watt-activation/part-317

ex:lohe-spherical

—

hasTokensPerSecondblah/watt-activation/part-317

12815

—

convergesIdenticallyToblah/watt-activation/part-317

ex:lohe-spherical

—

hasLossblah/watt-activation/part-317

9.02

—

hasParamCountblah/watt-activation/part-317

107.2M

—

hasParamCountApproxblah/watt-activation/part-317

17M

—

hasPPLblah/watt-activation/part-317

8228

—

hasBestQualityblah/watt-activation/part-103

true

—

hasPplblah/watt-activation/part-103

6.03

—

hasTokensPerSecondAtTemp256blah/watt-activation/part-103

210000

—

hasQualityAdjustedTokensPerSecondblah/watt-activation/part-103

8008

—

hasTokPerSblah/watt-activation/part-319

56931

—

hasFinalPplblah/watt-activation/part-319

239

—

hasParamsblah/watt-activation/part-319

16.9M

—