The tests are the product. Watch them run.
One shared run measures every substrate primitive on this box, over a bounded slice of live public data, and streams each result here the instant it lands. Everyone watching sees the same run, so a full gallery of spectators adds nothing to the load. Nothing on this page is typed in by hand, if a tile reads measuring… the number has not landed yet.
graph traversal · the record + its neighbors, over the live value-flow
598.9MTEPS
warm working-set BFS over the resident value-flow graph, the trailing average across live runs
The same figure and the same method label the scorecard, the reproducibility API and the
graph demo all now read from, so the number never disagrees with itself.
see the proof ▸
Writes
Cold
cold
57.5M/s
cold durable append · O_DIRECT · 2-disk fanned aggregate
measured
69.5 ms
replaces a bolt-on write-ahead log
Reads
Cold
cold
162.7M/s
by-id coalesced · QD64 · O_DIRECT · 2-disk fanned aggregate
measured
9.8 ms
replaces a page cache you must pre-warm
Warm
warm
322.0M/s
random by-id gather · DRAM-resident · all-cores (memory-bound)
measured
9.9 ms
replaces an in-memory cache tier
Hot
hot
3.32B/s
random by-id gather · L3-resident · all-cores (memory-bound)
measured
4.8 ms
replaces an in-memory cache tier
overall
blended
715.9M/s
effective rate for a query workload at the measured query tier mix
query mix · cold 0.0% · warm 39.1% · hot 60.9%
Traversals
Cold
cold
196.4MTEPS
cold first-touch · co-located image · all-shards/all-disks
measured
38.22 s
replaces a bolt-on native-graph store
Warm
warm
588.8MTEPS
page-cache-warm rerun · co-located image · all-shards/all-disks
measured
24.2 ms
replaces a bolt-on native-graph store
Hot
hot
1.76BTEPS
L3-local co-located image · dense seeds · all-cores
measured
1.7 ms
replaces a bolt-on native-graph store
overall
blended
990.5MTEPS
effective rate for a query workload at the measured query tier mix
query mix · cold 0.0% · warm 39.1% · hot 60.9%
Scans
Cold
cold
13.9GB/s
sequential sweep · QD64 · O_DIRECT · 2-disk fanned aggregate
measured
7.4 ms
replaces a columnar analytics engine
Warm
warm
36.62GB/s
resident SIMD sweep · in-process · all-cores (memory-bound)
measured
3.5 ms
replaces a columnar analytics engine
Hot
hot
504.25GB/s
L3-resident SIMD sweep · in-process · all-cores (memory-bound)
measured
203.0 µs
replaces a columnar analytics engine
overall
blended
84GB/s
effective rate for a query workload at the measured query tier mix
query mix · cold 0.0% · warm 39.1% · hot 60.9%
Binary-neuron kernel
Kernel
compute
6.7TSAOPS/s
all-core aggregate · 7 cores · AVX-512 VPOPCNTQ
measured
replaces a GPU inference tier
Conveyor · predictive prefetch · janitor efficiency, measured on the live handle
Conveyor · where all live reads land
89.4%served at cache speed (warm + hot)
cold 10.6%
warm 15.8%
hot 73.6%
the whole-store LIVE mix (queries plus the ingest's whole-store random dedup, uncacheable by physics over a 500x-RAM store), distinct from the query blended tile above; the cold band shrinks toward warm and hot as the conveyor lifts the working set the queries reuse
Predictive prefetch · bounded band
507.0Krecords streamed into the walk
only the slice the query touches is prefetched, the pipe stays small while the reads stay hot, never the whole graph
Janitor · co-location, epoch over epoch
co-location warming · no epochs on this handle during the re-ingest yet
Every number here is measured live on this box, on a bounded slice of public Bitcoin data you
can re-pull over Bitcoin P2P and re-run. Writes go to a throwaway scratch area, never a live
store; reads, traversals and scans read the live value-flow graph without touching it. The
honesty band on each tile tells you what class of number it is.