Measured Silicon Benchmarks & Proofs

AAYU Silicon Performance & AI Benchmarks

Real measured results across the 3 core pillars: flat-stack zero-copy function calls, native silicon Tensor/GEMM neural operators, and Liquid Data 2.0 molecular compression.

PILLAR 1: FLAT-STACK VM

7.2x Recursive Speedup

Replaced 2KB CallFrame cloning with a preallocated flat-stack base pointer. Eliminated 28 GB of heap allocations in recursive workloads, dropping fib(32) to sub-second speed.

0.99s7.20s(7.2x faster)
PILLAR 2: SILICON AI GEMM

Native Cache-Blocked GEMM

32x32 cache-tiled matrix multiplication, batched dot-product attention, Taylor-expanded GELU/SiLU activations, and Vector RAG cosine similarity.

0.9993RAG Cosine Match (<1μs)
PILLAR 3: LIQUID DATA 2.0

91.1% - 95.2% Byte Compaction

Git CAS interning + Columnar FoR deltas + 7-layer cryptographic police inspectors with O(1) random cell access and 100% byte-flip quarantine.

100%Byte-Flip Tamper Catch Rate

Empirical Execution Speed & Architectural Ground Truth

WorkloadNative C / Rust (-O)AAYU (Silicon VM)Node.js (V8 JIT)Python 3.12Architectural Ground Truth
Cold Boot Time~2 ms13.9 ms~35 ms~50 msInstant startup for CLI tools and serverless containers
Recursion fib(32)34 ms383 ms (bytecode)173 ms780 ms2x faster than Python; frame zeroing optimization avoids 11.1B writes
50M Hot Loop94 ms2,432 ms329 ms11,442 ms4.7x faster than Python; ~25x behind native C (standard for bytecode VMs)
5M Function Calls34 ms666 ms183 ms1,908 ms2.9x faster than Python; flat stack dispatch without heap allocs
🌊 Liquid Data 2.0 vs Gzip:On 100-row JSON streaming data, Gzip achieves 960 Bytes vs Liquid Data's 1,164 Bytes. Liquid Data carries 204 bytes of in-memory database DAG metadata (blueprint schema, Git CAS string interning, columnar FoR deltas, and parity matrix). In return, Liquid Data provides O(1) random cell access (<1μs) without stream decompression, 100% byte-flip quarantine, and autonomous WAL self-healing.
🌐 Web Server & Binary Footprint:Current runtime uses a multi-threaded socket engine (thread::spawn) with Connection: close (~1,500 - 2,500 req/s in real networks; 36,000 req/s in loopback mock). Epoll + keep-alive pooling is scheduled for the next release. Windows PE binary (aayu.exe) is 2.27 MB with embedded SQLite & crypto; minimal stripped Linux ELF is ~890 KB.
⚡ Architectural Ground Truth & JIT Roadmap:As a pure bytecode interpreter, AAYU VM runs ~20-25x behind native compiled C/Rust (Rust: 94ms vs AAYU: 2,432ms on 50M loop; 34ms vs 383ms on fib(32)), while executing 3-5x faster than Python 3.12. Node.js V8 JIT sits in between (~3.5-5x behind C). To bridge the remaining gap to native silicon without compromising type safety or full-stack features, AAYU's JIT compiler roadmap (jit.rs via can_compile) selectively compiles verified pure integer loops and recursion directly into native instructions, keeping dynamic code safely in the VM.

Native Silicon Tensor & AI Engine Benchmarks

Workload / OperationDimension / InputsExpected ResultStatusVerification
Cache-Blocked GEMM (MatMul)[2, 2] x [2, 2][[19, 22], [43, 50]]VERIFIED32x32 Cache Tiled Kernel
Transformer Attention(Q, K, V)Seq=2, D=4[[0.58, 0.25, 0.15, 0.68], ...]VERIFIEDScaled Dot-Product Forward
GELU Activation[-2.0, -1.0, 0.0, 1.0, 2.0, 3.0]gelu(1.0) = 0.8412VERIFIEDTaylor Approximation Constants
Vector RAG Cosine Similarity4D Embedding Vectorssim1 = 0.9993VERIFIEDO(1) L2-Norm Normalized
Adam Optimizer StepWeights, Grads, M, V, t=1Bias-Corrected GradientsVERIFIEDstd/ai.aayu Pure Action

AAYU Core Workload Matrix (20 Programs)

Twenty distinct benchmark programs from the currently supported AAYU core subset, all run at N=10,000. Every case checks its output.

Recorded environment: Windows 10.0.26300, Intel64 Family 6 Model 142, 8 logical processors, x64, Rust VM crate v0.3.1. One warm-up, 11 measured runs per case.
WorkloadInput sizeExpected outputMedian msMin msMax ms
01 · Integer addition accumulator10,0004999500073.54670.314199.249
02 · Integer subtraction countdown10,000074.21570.876155.731
03 · Integer multiply-accumulate10,00014998500076.67974.66991.060
04 · Integer divide-accumulate10,00024995000112.86887.982163.001
05 · Float half-step accumulation10,000500075.45270.69277.032
06 · Float multiply-accumulate10,0002499750073.27970.93377.946
07 · Less-than true branch10,000500072.85171.21277.565
08 · Greater-than true branch10,000499972.78472.22777.527
09 · Equality branch10,000173.86770.72378.430
10 · Inequality branch10,000999974.74571.80578.093
11 · Less-or-equal branch10,000500172.49671.00576.236
12 · Greater-or-equal branch10,000500073.43170.29277.896
13 · Nested loop counter10,0001000076.88670.72280.723
14 · Nested loop weighted sum10,00049500073.16969.78881.360
15 · Mixed multiply and subtract10,0004999500073.87872.85999.137
16 · Multiply then divide10,00024997500074.59270.75382.061
17 · Descending sum10,0005000500075.98572.07086.114
18 · Variable copy and accumulate10,0004999500073.45069.88799.145
19 · Nested conditional branches10,0001999974.71772.38578.386
20 · Comparison loop boundary10,0005000500071.63970.37577.727

Reproduce with .\aayu.exe run tests\test_tensor_ai_math.aayu and .\aayu.exe run tests\test_compiler_security.aayu.