host dict | lattice dict | hops dict | fail_closed dict | gemm listlengths 4 4 | note stringclasses 1
value |
|---|---|---|---|---|---|
{
"system": "Windows",
"machine": "AMD64",
"processor": "Intel64 Family 6 Model 183 Stepping 1, GenuineIntel",
"python": "3.12.10"
} | {
"N": 6328,
"degree": 6,
"dtype": "uint16",
"lut_bytes": 75936,
"lut_kib": 74.15625,
"construction": "circulant offsets {±1,±2,±3} mod N"
} | {
"count": 50000000,
"median_s": 0.28765409998595715,
"ns_per_hop": 5.753081999719143,
"hops_per_s": 173819876.03319725,
"samples_s": [
0.2777026000258047,
0.2792536999913864,
0.28757450002012774,
0.28765409998595715,
0.29040580001310445,
0.2965820000099484,
0.2991461000056006
]
... | {
"poison_after_hops": 100,
"median_detect_s": 8.00006091594696e-7,
"detect_ns": 800.006091594696
} | [
{
"dim": 64,
"dtype": "float32",
"bytes_ab": 32768,
"median_s": 0.00000580001506023109,
"ns": 5800.01506023109,
"gflops": 90.3942480416785,
"hops_in_one_gemm": 1008.1578987600451
},
{
"dim": 256,
"dtype": "float32",
"bytes_ab": 524288,
"median_s": 0.000219899986404925... | Hops are integer LUT walks, not language-model tokens. GEMM is dense float32 matmul of the stated order. |
{
"system": "Windows",
"machine": "AMD64",
"processor": "Intel64 Family 6 Model 158 Stepping 9, GenuineIntel",
"python": "3.11.9"
} | {
"N": 6328,
"degree": 6,
"dtype": "uint16",
"lut_bytes": 75936,
"lut_kib": 74.15625,
"construction": "circulant offsets {±1,±2,±3} mod N"
} | {
"count": 50000000,
"median_s": 0.4879461000673473,
"ns_per_hop": 9.758922001346946,
"hops_per_s": 102470334.31171784,
"samples_s": [
0.4850631000008434,
0.486387399956584,
0.48783789994195104,
0.4879461000673473,
0.48885229998268187,
0.4927473000716418,
0.49634680000599474
]
} | {
"poison_after_hops": 100,
"median_detect_s": 0.000001500011421740055,
"detect_ns": 1500.011421740055
} | [
{
"dim": 64,
"dtype": "float32",
"bytes_ab": 32768,
"median_s": 0.00001009996049106121,
"ns": 10099.96049106121,
"gflops": 51.90990603022772,
"hops_in_one_gemm": 1034.946328054184
},
{
"dim": 256,
"dtype": "float32",
"bytes_ab": 524288,
"median_s": 0.00023520004469901... | Hops are integer LUT walks, not language-model tokens. GEMM is dense float32 matmul of the stated order. |
T112 L2 Lattice
Not a language model. Not weights. A 74.16 KiB adjacency table.
uint16[6328][6] = 75,936 bytes. Degree-6 circulant graph on (N = T_{112} = 6328). Fits in a 256 KiB Kaby Lake L2 (29%).
- Paper / DOI: https://doi.org/10.5281/zenodo.22406010
- Code: https://github.com/ultranetcommand-neo/t112-l2-lattice
- Author: Matthew Scott Gibson · https://orcid.org/0009-0001-4167-201X
Files
| File | What |
|---|---|
lut.npy |
(6328, 6) uint16 adjacency |
lut.csv |
node,n0..n5 |
telemetry_i7-7700.json |
Neo, 9.76 ns/hop |
telemetry_i7-14700F.json |
Local, 5.75 ns/hop |
Offsets: {±1, ±2, ±3} mod 6328. Golden 50M-hop SHA-256 (seed 7): 9d9bc34b4df51e88d71d280bb0547e8b9a117c5873ab3582792dd6a7059bec68
Hops are not tokens. GEMM in the telemetry is vendor BLAS, not this table pretending to be a transformer.
Load
import numpy as np
lut = np.load("lut.npy") # shape (6328, 6), dtype uint16
node = 0
node = int(lut[node, 0]) # one hop
Apache-2.0. Matthew Scott Gibson / Crimson OS.
ORCID: https://orcid.org/0009-0001-4167-201X
- Downloads last month
- -