M3 Ultra

as donto believed on

From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-06-06.)

M3 Ultra has 96 facts recorded in Dontopedia across 38 references, with 2 live disagreements.

96 facts·70 predicates·38 sources·2 in dispute

Mostly:rdf:type(15), has theoretical max bandwidth(2), has headroom(2)

Rdf:typein dispute`rdf:type`

Hardware Model^[21]all time · 45
Hardware^[22]all time · 17
Device^[23]all time · 19
Hardware^[24]all time · 58
Hardware^[25]all time · 61
Hardware^[27]all time · 88
Hardware^[28]all time · 592
Hardware Model^[29]all time · 593
Hardware Processor^[30]all time · 594
Hardware^[32]sourceall time · 596

Inbound mentions (63)

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.

`runsOnHardware`Runs on Hardware(4)

Cl6 Field Cpu Pathex:cl6-field-cpu-path
Finding 4ex:finding-4
Smoke Resultsex:smoke-results
Trainingex:training

`hardwarePlatform`Hardware Platform(2)

M3 Ultra F32 Resultex:m3-ultra-f32-result
M3 Ultra F64 Resultex:m3-ultra-f64-result

`isFasterThan`Is Faster Than(2)

Best Waverotor Portex:best-waverotor-port
Gb10 Pytorch Portex:gb10-pytorch-port

`measuredOnHardware`Measured on Hardware(2)

Measurement H40ex:measurement-h40
Performance Metric 1ex:performance-metric-1

`mentionsHardware`Mentions Hardware(2)

Message 1ex:message-1
Utterance 2026 04 15 20 13ex:utterance-2026-04-15-20-13

`occursOnHardware`Occurs on Hardware(2)

Bandwidth Wallex:bandwidth-wall
Performance Measurementex:performance-measurement

`performedOnHardware`Performed on Hardware(2)

Per Config Benchmarkex:per-config-benchmark
This Training Runex:this-training-run

`achievableIn20Hours`Achievable In20 Hours(1)

500k Sweepsex:500k-sweeps

`achievableIn24Minutes`Achievable In24 Minutes(1)

10k Sweepsex:10k-sweeps

`achievableIn4Hours`Achievable In4 Hours(1)

100k Sweepsex:100k-sweeps

`achievedOnHardware`Achieved on Hardware(1)

Semi Fused Approach Resultsex:semi-fused-approach-results

`achievesEffectiveBandwidthOn`Achieves Effective Bandwidth on(1)

Semi Fused Approachex:semi-fused-approach

`addressesAudienceWith`Addresses Audience With(1)

Xenonfunex:xenonfun

`areSignificant`Are Significant(1)

Cache Savingsex:cache-savings

`cannotPinCores`Cannot Pin Cores(1)

Macosex:macos

`cannotUseAllGpuCores`Cannot Use All Gpu Cores(1)

Fused Single Tgex:fused-single-tg

`compares`Compares(1)

Performance Gap Comparisonex:performance-gap-comparison

`degradesAtLargeN`Degrades at Large N(1)

Fused Pathex:fused-path

`executedOn`Executed on(1)

Training Session 2026 03 07ex:training-session-2026-03-07

`executedOnHardware`Executed on Hardware(1)

M3 Ultra Setupex:m3-ultra-setup

`fasterPerAmxUnitThan`Faster Per Amx Unit Than(1)

M4 Proex:m4-pro

`fitsComfortablyOnM3Ultra`Fits Comfortably on M3 Ultra(1)

L4096ex:l4096

`hardwareTarget`Hardware Target(1)

Rayon Parallel Sweepex:rayon-parallel-sweep

`hasNoMemoryConcernOn`Has No Memory Concern on(1)

Anchor V3ex:anchor-v3

`hasZeroDegradationWithScale`Has Zero Degradation With Scale(1)

Tiling Strategyex:tiling-strategy

`hostsExperiments`Hosts Experiments(1)

Apple Siliconex:apple-silicon

`improvesThroughputOn`Improves Throughput on(1)

Compiled Stepex:compiled-step

`includesBandwidthWallFinding`Includes Bandwidth Wall Finding(1)

Kuramoto Findings Mdex:kuramoto-findings-md

`isFasterPerAmxUnitThan`Is Faster Per Amx Unit Than(1)

M4 Proex:m4-pro

`isLessDramatic`Is Less Dramatic(1)

F64 Collapseex:f64-collapse

`isNegligible`Is Negligible(1)

Parallel Measurement Overheadex:parallel-measurement-overhead

`isOutperformedByM3Ultra`Is Outperformed by M3 Ultra(1)

Legacyex:legacy

`isTheBottleneck`Is the Bottleneck(1)

Memory Bandwidthex:memory-bandwidth

`isUntappedOpportunity`Is Untapped Opportunity(1)

Multi Cluster Parallelismex:multi-cluster-parallelism

`keepsFullParallelism`Keeps Full Parallelism(1)

Multi Threadgroup Semi Fusionex:multi-threadgroup-semi-fusion

`killsPerformance`Kills Performance(1)

Single Threadgroup Fusionex:single-threadgroup-fusion

`limitsCorePinning`Limits Core Pinning(1)

Macosex:macos

`locatedOn`Located on(1)

Anchorkan Modelex:anchorkan-model

`location`Location(1)

F64 Collapseex:f64-collapse

`maintainsMultiCoreParallelism`Maintains Multi Core Parallelism(1)

Semi Fused Approachex:semi-fused-approach

`measuredOn`Measured on(1)

Sim Size 512x512ex:sim-size-512x512

`nonIssue`Non Issue(1)

Memoryex:memory

`nonIssueOnM3Ultra`Non Issue on M3 Ultra(1)

Memory for L1024 L2048ex:memory-for-l1024-l2048

`projectsHardware`Projects Hardware(1)

Performance Projection Tableex:performance-projection-table

`recommendCpuOnly`Recommend Cpu Only(1)

Production Runsex:production-runs

`reducesDispatchOverhead`Reduces Dispatch Overhead(1)

Multi Threadgroup Semi Fusionex:multi-threadgroup-semi-fusion

`sharesBenchmarks`Shares Benchmarks(1)

Xenonfunex:xenonfun

`slowerThanM3UltraAtSameTask`Slower Than M3 Ultra at Same Task(1)

M4 Proex:m4-pro

`superiorInPerUnitSpeed`Superior in Per Unit Speed(1)

M4 Proex:m4-pro

`supportsParallelUpdatesAcrossPcores`Supports Parallel Updates Across Pcores(1)

Rayonex:rayon

`targetsHardware`Targets Hardware(1)

Rust Accelerate Benchmarkex:rust-accelerate-benchmark

`trainedOnHardware`Trained on Hardware(1)

H 40 Presetex:h-40-preset

`worksPerfectly`Works Perfectly(1)

Tiling Strategyex:tiling-strategy

`wouldDoublePerformance`Would Double Performance(1)

Parallel Tilingex:parallel-tiling

Other facts (71)

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.

71 facts

Predicate	Value	Ref
Has Theoretical Max Bandwidth	800	[1]
Has Theoretical Max Bandwidth	800	[21]
Has Headroom	true	[14]
Has Headroom	true	[37]
Achieves Bandwidth Utilization	39%	[1]
Has Unified Memory	M Series	[1]
Bottlenecked by Memory Bandwidth	Memory Bandwidth	[1]
Has Gpu Cores	76	[1]
Is Lisa Hardware	null	[2]
Has Activity Duration	5 days	[2]
Has Cache Limit	40GB	[3]
Has No Memory Concern	true	[4]
Has No Memory Concern for	Anchor V3	[4]
Power Usage Watts	75	[5]
Faster Than For100 Iters	Rtx 3060 12gb	[5]
Time For100 Iters Seconds	22.3	[5]
Will Benchmark Changes	Code Changes	[6]
Has Pcores Count	20	[6]
Runs	500k Sweeps 1024 Squared	[7]
Is High End Hardware	true	[7]
Has Memory Bandwidth Gbs	800	[8]
Benefits From Parallelism	Rayon Threads	[8]
Has Advantage in	More Amx Units Across Clusters	[8]
Should Pull Ahead	M4 Pro	[8]
Has Number of Pclusters	2	[8]
Lacks Advantage in	Speed Per Unit	[8]
Is Clear Choice	Production 1024 Squared Runs	[9]
Faster Than	M4 Pro	[9]
Advocated for Production	Xenonfun	[9]
Enables Large Scale Cpu Compute	1024x1024 Cl60 Metropolis Experiment	[10]
Has28 Threads Configuration	null	[10]
Has Batch Time in Seconds At1024 Squared	14.4	[10]
Outperforms M4 Pro At1024 Squared	M4 Pro	[10]
Achieves Performance At512x512	843 GFLOP/s	[11]
Superior to	Rtx 3060	[12]
Has Throughput Target	20-40K tok/s	[12]
Achieves Tok Per S	1771	[13]
Expected to Outperform	Rtx 3060	[13]
Should Be Competitive With	Rtx 3060	[13]
Has Performance Gap to	Rtx 3060	[13]
Targets Model Size	25M	[13]
Presupposes25m Model	F32 Model	[13]
Supports Concurrent Runs	Above Prediction	[14]
Has24 Cores	true	[15]
Supports Gpu Training	true	[16]
Hardware Context	Bandwidth Saturation	[17]
Has Time Period	2026-02-27 to 2026-03-02	[18]
Has Total Api Spending	513.84	[18]
Is Machine	true	[18]
Is Hardware for Config	Spectralreservoir Gptv3	[19]
Supports Training	Training	[20]
Has Total Cores	76	[21]
Has Duration of Activity	5 days	[23]
Has Saturation Throughput	56000	[25]
Has Memory Bandwidth Wall	~56K tok/s	[26]
Per Amx Unit Performance	~125 GFLOP/s	[29]
Advantage Source	more AMX units across clusters	[29]
Disadvantage	not faster ones	[29]
Performance Characteristic	remarkably flat	[29]
Performance Value	125 GFLOP/s	[29]
Fused F64 Performance At1 M	44	[29]
Memory Bandwidth	800 GB/s	[29]
Memory Bandwidth Effect	helps the untiled path survive longer	[29]
P Cluster Count	2	[29]
Projected Performance	~250 GFLOP/s f32	[29]
Projected Performance Reason	Proposed Execution Mode	[29]
Relative Performance Result	flip the comparison	[29]
Expected Outcome	actually pull ahead	[29]
Clear Choice for	production 1024² runs	[31]
Benchmark Task	Current Code State	[32]
Core Count	20	[32]

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.

96 undated

—

achievesBandwidthUtilizationblah/resources/part-45

39%

—

hasUnifiedMemoryblah/resources/part-45

ex:m-series

—

bottleneckedByMemoryBandwidthblah/resources/part-45

ex:memory-bandwidth

—

hasTheoreticalMaxBandwidthblah/resources/part-45

800

—

hasGpuCoresblah/resources/part-45

—

isLisaHardwareblah/watt-activation/part-19

null

—

hasActivityDurationblah/watt-activation/part-19

5 days

—

hasCacheLimitblah/watt-activation/part-57

40GB

—

hasNoMemoryConcernblah/watt-activation/part-58

true

—

hasNoMemoryConcernForblah/watt-activation/part-58

ex:anchor-v3

—

powerUsageWattsblah/watt-activation/part-587

—

fasterThanFor100Itersblah/watt-activation/part-587

ex:rtx-3060-12gb

—

timeFor100ItersSecondsblah/watt-activation/part-587

22.3

—

willBenchmarkChangesblah/watt-activation/part-599

ex:code-changes

—

hasPcoresCountblah/watt-activation/part-599

—

runsblah/watt-activation/part-600

ex:500k-sweeps-1024-squared

—

isHighEndHardwareblah/watt-activation/part-600

true

—

hasMemoryBandwidthGbsblah/watt-activation/part-596

800

—

benefitsFromParallelismblah/watt-activation/part-596

ex:rayon-threads

—

hasAdvantageInblah/watt-activation/part-596

ex:more-amx-units-across-clusters

—

shouldPullAheadblah/watt-activation/part-596

ex:m4-pro

—

hasNumberOfPclustersblah/watt-activation/part-596

—

lacksAdvantageInblah/watt-activation/part-596

ex:speed-per-unit

—

isClearChoiceblah/watt-activation/part-601

ex:production-1024-squared-runs

—

fasterThanblah/watt-activation/part-601

ex:m4-pro

—

advocatedForProductionblah/watt-activation/part-601

ex:xenonfun

—

enablesLargeScaleCpuComputeblah/watt-activation/part-602

ex:1024x1024-cl60-metropolis-experiment

—

has28ThreadsConfigurationblah/watt-activation/part-602

null

—

hasBatchTimeInSecondsAt1024Squaredblah/watt-activation/part-602

14.4

—

outperformsM4ProAt1024Squaredblah/watt-activation/part-602

ex:m4-pro

—

achievesPerformanceAt512x512blah/watt-activation/part-597

843 GFLOP/s

—

superiorToblah/watt-activation/part-631

ex:rtx-3060

—

hasThroughputTargetblah/watt-activation/part-631

20-40K tok/s

—

achievesTokPerSblah/watt-activation/part-641

1771

—

expectedToOutperformblah/watt-activation/part-641

ex:rtx-3060

—

shouldBeCompetitiveWithblah/watt-activation/part-641

ex:rtx-3060

—

hasPerformanceGapToblah/watt-activation/part-641

ex:rtx-3060

—

targetsModelSizeblah/watt-activation/part-641

25M

—

presupposes25mModelblah/watt-activation/part-641

ex:f32-model

—

hasHeadroomblah/watt-activation/part-671

true

—

supportsConcurrentRunsblah/watt-activation/part-671

ex:above-prediction

—

has24Coresblah/watt-activation/part-678

true

—

supportsGpuTrainingblah/watt-activation/part-680

true

—

hardwareContextblah/watt-activation/part-61

ex:bandwidth-saturation

—

hasTimePeriodblah/watt-activation/part-17

2026-02-27 to 2026-03-02

—

hasTotalApiSpendingblah/watt-activation/part-17

513.84

—

isMachineblah/watt-activation/part-17

true

—

isHardwareForConfigblah/watt-activation/part-103

ex:spectralreservoir-gptv3

—

supportsTrainingblah/watt-activation/part-605

ex:training

—

typeblah/resources/45

ex:HardwareModel

—

hasTheoreticalMaxBandwidthblah/resources/45

800

—

hasTotalCoresblah/resources/45

—

labelblah/watt-activation/17

M3 Ultra

—

typeblah/watt-activation/17

ex:Hardware

—

typeblah/watt-activation/19

ex:Device

—

labelblah/watt-activation/19

M3 Ultra

—

hasDurationOfActivityblah/watt-activation/19

5 days

—

labelblah/watt-activation/58

M3 Ultra

—

typeblah/watt-activation/58

ex:Hardware

—

typeblah/watt-activation/61

ex:Hardware

—

hasSaturationThroughputblah/watt-activation/61

56000

—

hasMemoryBandwidthWallblah/watt-activation/62

~56K tok/s

—

typeblah/watt-activation/88

ex:Hardware

—

typeblah/watt-activation/592

ex:Hardware

—

labelblah/watt-activation/593

M3 Ultra

—

typeblah/watt-activation/593

ex:HardwareModel

—

perAmxUnitPerformanceblah/watt-activation/593

~125 GFLOP/s

—

advantageSourceblah/watt-activation/593

more AMX units across clusters

—

disadvantageblah/watt-activation/593

not faster ones

—

performanceCharacteristicblah/watt-activation/593

remarkably flat

—

performanceValueblah/watt-activation/593

125 GFLOP/s

—

fusedF64PerformanceAt1Mblah/watt-activation/593

—

memoryBandwidthblah/watt-activation/593

800 GB/s

—

memoryBandwidthEffectblah/watt-activation/593

helps the untiled path survive longer

—

pClusterCountblah/watt-activation/593

—

projectedPerformanceblah/watt-activation/593

~250 GFLOP/s f32

—

projectedPerformanceReasonblah/watt-activation/593

ex:proposed-execution-mode

—

relativePerformanceResultblah/watt-activation/593

flip the comparison

—

expectedOutcomeblah/watt-activation/593

actually pull ahead

—

typeblah/watt-activation/594

ex:HardwareProcessor

—

labelblah/watt-activation/594

M3 Ultra

—

clearChoiceForblah/watt-activation/598

production 1024² runs

—

typeblah/watt-activation/596

ex:Hardware

—

labelblah/watt-activation/596

M3 Ultra

—

benchmarkTaskblah/watt-activation/596

ex:current-code-state

—

coreCountblah/watt-activation/596

—

typeblah/watt-activation/597

ex:Hardware

—

typeblah/watt-activation/599

ex:Processor

—

typeblah/watt-activation/629

ex:Hardware

—

labelblah/watt-activation/629

M3 Ultra

—

typeblah/watt-activation/638

ex:Hardware

—

labelblah/watt-activation/638

M3 Ultra

—

labelblah/watt-activation/668

M3 Ultra

—

typeblah/watt-activation/668

ex:Hardware

—

hasHeadroomblah/watt-activation/668

true

—

labelblah/watt-activation/677

M3 Ultra

References (38)

38 references

[1]Part 455 facts
ctx:discord/blah/resources/part-45
[2]Part 192 facts
ctx:discord/blah/watt-activation/part-19
[3]Part 571 fact
ctx:discord/blah/watt-activation/part-57
[4]Part 582 facts
ctx:discord/blah/watt-activation/part-58
[5]Part 5873 facts
ctx:discord/blah/watt-activation/part-587
[6]Part 5992 facts
ctx:discord/blah/watt-activation/part-599
[7]Part 6002 facts
ctx:discord/blah/watt-activation/part-600
[8]Part 5966 facts
ctx:discord/blah/watt-activation/part-596
[9]Part 6013 facts
ctx:discord/blah/watt-activation/part-601
[10]Part 6024 facts
ctx:discord/blah/watt-activation/part-602
[11]Part 5971 fact
ctx:discord/blah/watt-activation/part-597
[12]Part 6312 facts
ctx:discord/blah/watt-activation/part-631
[13]Part 6416 facts
ctx:discord/blah/watt-activation/part-641
[14]Part 6712 facts
ctx:discord/blah/watt-activation/part-671
[15]Part 6781 fact
ctx:discord/blah/watt-activation/part-678
[16]Part 6801 fact
ctx:discord/blah/watt-activation/part-680
[17]Part 611 fact
ctx:discord/blah/watt-activation/part-61
[18]Part 173 facts
ctx:discord/blah/watt-activation/part-17
[19]Part 1031 fact
ctx:discord/blah/watt-activation/part-103
[20]Part 6051 fact
ctx:discord/blah/watt-activation/part-605
[21]453 facts
ctx:discord/blah/resources/45
- full textresources-45
  text/plain2 KBdoc:agent/resources-45/f97a124a-8c9b-4da1-a8a2-b10c6d609fbf
  Show excerpt
  [2026-03-02 23:55] xenonfun: Excellent results. The semi-fused approach is the winner: - Semi-fused: 130 tok/s (7.69 ms/tok, 311 GB/s effective bandwidth) - Non-fused: 93 tok/s (10.78 ms/tok, 222 GB/s) - Fused single-TG: 11 tok/s (de…
[22]172 facts
ctx:discord/blah/watt-activation/17
- full textwatt-activation-17
  text/plain3 KBdoc:agent/watt-activation-17/34c034a2-297a-49fd-8499-a54ef15da9a5
  Show excerpt
  [2026-03-03 17:55] xenonfun: 2026-02-27 to 2026-03-02 (just the M3 Ultra) Total API spending on this machine: ~$513.84 (all-time, cached data through Mar 2) Breakdown by major project | Project | Total | |---------|-------| | symbiogenesis…
[23]193 facts
ctx:discord/blah/watt-activation/19
- full textwatt-activation-19
  text/plain2 KBdoc:agent/watt-activation-19/e74bc25c-aab8-43ac-90e0-2f036b5a9627
  Show excerpt
  [2026-03-05 22:21] xenonfun: Both started from the same checkpoint, so same baseline: - Start checkpoint ./philosophy_model_fresh/checkpoint_iter_9198.npz - Baseline on same eval slice/settings: val_loss=5.355859, val_ppl=211.85 So …
[24]582 facts
ctx:discord/blah/watt-activation/58
- full textwatt-activation-58
  text/plain3 KBdoc:agent/watt-activation-58/2b893cf3-b6f9-4fdb-a94d-202a521b9459
  Show excerpt
  [2026-03-07 09:29] xenonfun: ``` Anchor v3 running. On memory — the anchor overhead is minimal: Anchor v3 extra parameters per layer (m=32): - anchor_re/im: 32 × 4 phases = 256 floats - anchor_omega: 256 floats - W_anchor_q: 320 ×…
[25]612 facts
ctx:discord/blah/watt-activation/61
- full textwatt-activation-61
  text/plain3 KBdoc:agent/watt-activation-61/3d9a288b-f7ca-4227-8d92-0bca73a33496
  Show excerpt
  [2026-03-07 10:25] xenonfun: ``` Config: anchor_v3_m32_L2048 (seq_len=2048, kwargs={'use_anchor': True, 'n_anchors': 32}) …
[26]621 fact
ctx:discord/blah/watt-activation/62
- full textwatt-activation-62
  text/plain2 KBdoc:agent/watt-activation-62/a54e16c9-84e0-4825-8097-9ebef4dc861d
  Show excerpt
  [2026-03-07 14:44] xenonfun: L512/L1024/L2048 data [2026-03-07 14:50] xenonfun: (files: phase_diagram_r_vs_ppl.png, training_trajectories_r_vs_ppl.png, scaling_r_and_ppl_vs_seqlen.png) [2026-03-07 14:51] xenonfun: ``` +### Per-configuratio…
[27]881 fact
ctx:discord/blah/watt-activation/88
- full textwatt-activation-88
  text/plain3 KBdoc:agent/watt-activation-88/0299a1f9-423c-4481-997f-02f201cc0796
  Show excerpt
  [2026-03-07 22:01] xenonfun: ``` ⏺ Training complete. Final results: ┌──────────────────┬───────────────────────────────────────┐ │ Metric │ Value │ ├──────────────────┼──────────────────────…
[28]5921 fact
ctx:discord/blah/watt-activation/592
- full textwatt-activation-592
  text/plain3 KBdoc:agent/watt-activation-592/64459ee2-1124-47ae-b744-747c970a98d6
  Show excerpt
  [2026-04-04 13:47] lisamegawatts: The pattern is clear: EVENT TIMELINE — COMET POSITION AT EACH INTERACTION: 01:17 UTC R2 radio blackout Comet at 14.7 R_sun (streamer crossing) 06:45 UTC CCOR-1 flash/whiteout Comet at 9.9 R_sun …
[29]59315 facts
ctx:discord/blah/watt-activation/593
- full textwatt-activation-593
  text/plain3 KBdoc:agent/watt-activation-593/5c383b0f-5e7c-4f28-b8e6-6c00caa3ef98
  Show excerpt
  [2026-04-04 17:10] xenonfun: ⏺ Tiled is a massive win. Now let me wire it into the production code and commit. ``` ┌────────┬─────────────┬─────────────┬─────────────┐ │ N │ Fused f32 │ Tiled f32 │ Improvement │ ├────────┼───…
[30]5942 facts
ctx:discord/blah/watt-activation/594
- full textwatt-activation-594
  text/plain3 KBdoc:agent/watt-activation-594/d9974ad5-bce6-44fb-a387-861b63507258
  Show excerpt
  [2026-04-04 17:21] xenonfun: ``` The headline numbers ┌──────────────────────┬─────────────────┬───────────────┬─────────┐ │ │ Scalar (before) │ Par tiled AMX │ Speedup │ ├──────────────────────┼───────────────…
[31]5981 fact
ctx:discord/blah/watt-activation/598
- full textwatt-activation-598
  text/plain2 KBdoc:agent/watt-activation-598/3c9c31a0-fbf1-4132-9ad5-b18c96a16586
  Show excerpt
  [2026-04-04 19:12] xenonfun: ⏺ Found it. The site index collection allocates a Vec<usize> of 512K elements every half-sweep, every sweep — that's 100 sweeps × 2 parities × 512K × 8 bytes = 800 MB of allocation churn per batch at 1024². Pl…
[32]5964 facts
ctx:discord/blah/watt-activation/596
- full textwatt-activation-596
  text/plain3 KBdoc:agent/watt-activation-596/83561bf3-adfc-4a74-8293-71c29d442bdf
  Show excerpt
  [2026-04-04 18:11] xenonfun: ⏺ scalar_product went from 10,789 µs to 142 µs — 76× speedup. The precomputed scalar pairs reduced ~1024 iterations to ~32. Overall Metropolis per-site: 4.8× speedup, from 21 µs/site to 4.3 µs/site. For a …
[33]5971 fact
ctx:discord/blah/watt-activation/597
[34]5991 fact
ctx:discord/blah/watt-activation/599
[35]6292 facts
ctx:discord/blah/watt-activation/629
- full textwatt-activation-629
  text/plain3 KBdoc:agent/watt-activation-629/23bf04a4-e754-4ece-8525-729dd63d8627
  Show excerpt
  [2026-04-15 22:07] xenonfun: ⏺ The output is strikingly coherent for a 25M-param byte-level model trained on dialogue: ``` ▎ Alan Turing: What do you think about consciousness? ▎ I find your consciousness in a principle in the world's w…
[36]6382 facts
ctx:discord/blah/watt-activation/638
- full textwatt-activation-638
  text/plain3 KBdoc:agent/watt-activation-638/539b1483-cc08-42f3-9ea1-6a7f6ae47ab8
  Show excerpt
  [2026-04-16 00:45] xenonfun: ``` ⏺ Let me answer the Python question first, then look at rayon parallelism. Python training setup (from scaling_wave_native_curriculum.py): - BATCH_SIZE = 32 (base script has BATCH_SIZE=16 with GRAD_ACCU…
[37]6683 facts
ctx:discord/blah/watt-activation/668
- full textwatt-activation-668
  text/plain3 KBdoc:agent/watt-activation-668/724edb5f-0ab5-4a49-a13a-d6074cbe8065
  Show excerpt
  [2026-04-20 12:46] xenonfun: ``` ⏺ Acknowledged — #34 is next up after PR #13 returns (that subagent is ~halfway through its budget, leaving it to complete avoids wasted work). No new Clifford subagents until #34 lands. Updated #34 scope…
[38]6771 fact
ctx:discord/blah/watt-activation/677
- full textwatt-activation-677
  text/plain2 KBdoc:agent/watt-activation-677/0a3c1f90-3dc6-48cb-b08e-24d6d4275ede
  Show excerpt
  [2026-04-22 13:11] xenonfun: • For the H=40 preset (D=560, ~19.0M params), current measured GPU training is: ``` 21,642 tok/s at B=16, T=256 on M3 Ultra. I also tried nearby H values in multiples of 4: | H | D | Params | GPU tok/s |…

Keep researching

Missing something or suspicious of what's here? Kick off a research session — a Claude agent will investigate, cite its sources, and file new facts into a dedicated context you can review before accepting into the shared view.

M3 Ultra

Rdf:typein disputerdf:type

Inbound mentions (63)

runsOnHardwareRuns on Hardware(4)

hardwarePlatformHardware Platform(2)

isFasterThanIs Faster Than(2)

measuredOnHardwareMeasured on Hardware(2)

mentionsHardwareMentions Hardware(2)

occursOnHardwareOccurs on Hardware(2)

performedOnHardwarePerformed on Hardware(2)

achievableIn20HoursAchievable In20 Hours(1)

achievableIn24MinutesAchievable In24 Minutes(1)

achievableIn4HoursAchievable In4 Hours(1)

achievedOnHardwareAchieved on Hardware(1)

achievesEffectiveBandwidthOnAchieves Effective Bandwidth on(1)

addressesAudienceWithAddresses Audience With(1)

areSignificantAre Significant(1)

cannotPinCoresCannot Pin Cores(1)

cannotUseAllGpuCoresCannot Use All Gpu Cores(1)

comparesCompares(1)

degradesAtLargeNDegrades at Large N(1)

executedOnExecuted on(1)

executedOnHardwareExecuted on Hardware(1)

fasterPerAmxUnitThanFaster Per Amx Unit Than(1)

fitsComfortablyOnM3UltraFits Comfortably on M3 Ultra(1)

hardwareTargetHardware Target(1)

hasNoMemoryConcernOnHas No Memory Concern on(1)

hasZeroDegradationWithScaleHas Zero Degradation With Scale(1)

hostsExperimentsHosts Experiments(1)

improvesThroughputOnImproves Throughput on(1)

includesBandwidthWallFindingIncludes Bandwidth Wall Finding(1)

isFasterPerAmxUnitThanIs Faster Per Amx Unit Than(1)

isLessDramaticIs Less Dramatic(1)

isNegligibleIs Negligible(1)

isOutperformedByM3UltraIs Outperformed by M3 Ultra(1)

isTheBottleneckIs the Bottleneck(1)

isUntappedOpportunityIs Untapped Opportunity(1)

keepsFullParallelismKeeps Full Parallelism(1)

killsPerformanceKills Performance(1)

limitsCorePinningLimits Core Pinning(1)

locatedOnLocated on(1)

locationLocation(1)

maintainsMultiCoreParallelismMaintains Multi Core Parallelism(1)

measuredOnMeasured on(1)

nonIssueNon Issue(1)

nonIssueOnM3UltraNon Issue on M3 Ultra(1)

projectsHardwareProjects Hardware(1)

recommendCpuOnlyRecommend Cpu Only(1)

reducesDispatchOverheadReduces Dispatch Overhead(1)

sharesBenchmarksShares Benchmarks(1)

slowerThanM3UltraAtSameTaskSlower Than M3 Ultra at Same Task(1)

superiorInPerUnitSpeedSuperior in Per Unit Speed(1)

supportsParallelUpdatesAcrossPcoresSupports Parallel Updates Across Pcores(1)

targetsHardwareTargets Hardware(1)

trainedOnHardwareTrained on Hardware(1)

worksPerfectlyWorks Perfectly(1)

wouldDoublePerformanceWould Double Performance(1)

Other facts (71)

Timeline

References (38)

See also

Keep researching

Rdf:typein dispute`rdf:type`

`runsOnHardware`Runs on Hardware(4)

`hardwarePlatform`Hardware Platform(2)

`isFasterThan`Is Faster Than(2)

`measuredOnHardware`Measured on Hardware(2)

`mentionsHardware`Mentions Hardware(2)

`occursOnHardware`Occurs on Hardware(2)

`performedOnHardware`Performed on Hardware(2)

`achievableIn20Hours`Achievable In20 Hours(1)

`achievableIn24Minutes`Achievable In24 Minutes(1)

`achievableIn4Hours`Achievable In4 Hours(1)

`achievedOnHardware`Achieved on Hardware(1)

`achievesEffectiveBandwidthOn`Achieves Effective Bandwidth on(1)

`addressesAudienceWith`Addresses Audience With(1)

`areSignificant`Are Significant(1)

`cannotPinCores`Cannot Pin Cores(1)

`cannotUseAllGpuCores`Cannot Use All Gpu Cores(1)

`compares`Compares(1)

`degradesAtLargeN`Degrades at Large N(1)

`executedOn`Executed on(1)

`executedOnHardware`Executed on Hardware(1)

`fasterPerAmxUnitThan`Faster Per Amx Unit Than(1)

`fitsComfortablyOnM3Ultra`Fits Comfortably on M3 Ultra(1)

`hardwareTarget`Hardware Target(1)

`hasNoMemoryConcernOn`Has No Memory Concern on(1)

`hasZeroDegradationWithScale`Has Zero Degradation With Scale(1)

`hostsExperiments`Hosts Experiments(1)

`improvesThroughputOn`Improves Throughput on(1)

`includesBandwidthWallFinding`Includes Bandwidth Wall Finding(1)

`isFasterPerAmxUnitThan`Is Faster Per Amx Unit Than(1)

`isLessDramatic`Is Less Dramatic(1)

`isNegligible`Is Negligible(1)

`isOutperformedByM3Ultra`Is Outperformed by M3 Ultra(1)

`isTheBottleneck`Is the Bottleneck(1)

`isUntappedOpportunity`Is Untapped Opportunity(1)

`keepsFullParallelism`Keeps Full Parallelism(1)

`killsPerformance`Kills Performance(1)

`limitsCorePinning`Limits Core Pinning(1)

`locatedOn`Located on(1)

`location`Location(1)

`maintainsMultiCoreParallelism`Maintains Multi Core Parallelism(1)

`measuredOn`Measured on(1)

`nonIssue`Non Issue(1)

`nonIssueOnM3Ultra`Non Issue on M3 Ultra(1)

`projectsHardware`Projects Hardware(1)

`recommendCpuOnly`Recommend Cpu Only(1)

`reducesDispatchOverhead`Reduces Dispatch Overhead(1)

`sharesBenchmarks`Shares Benchmarks(1)

`slowerThanM3UltraAtSameTask`Slower Than M3 Ultra at Same Task(1)

`superiorInPerUnitSpeed`Superior in Per Unit Speed(1)

`supportsParallelUpdatesAcrossPcores`Supports Parallel Updates Across Pcores(1)

`targetsHardware`Targets Hardware(1)

`trainedOnHardware`Trained on Hardware(1)

`worksPerfectly`Works Perfectly(1)

`wouldDoublePerformance`Would Double Performance(1)