Hardware Performance Profiling

Hardware performance profiling runs a series of micro-benchmarks to capture timing characteristics unique to the physical CPU. Even two identical processor models differ due to silicon lottery, thermal state, and power management.

How it works

Neoprint executes 6 distinct workloads and measures their execution time in milliseconds:

1. Float arithmetic

Stresses the floating-point unit (FPU) pipeline with chained multiply-add-sqrt operations.

let x = 1.0000001
for (let i = 0; i < 50000; i++) {
  x = x * 1.0000001 + 0.0000001
  x = Math.sqrt(x * x + 1)
}

2. Trigonometry

Tests transcendental math unit performance with sin/cos/tan.

for (let i = 0; i < 30000; i++) {
  sum += Math.sin(i * 0.001) * Math.cos(i * 0.002) + Math.tan(i * 0.0001)
}

3. Array sort

Exercises branch prediction and memory access patterns.

Array.from(new Float64Array(5000)).sort((a, b) => a - b)

4. Object allocation

Measures heap allocation speed and garbage collector pressure.

for (let i = 0; i < 10000; i++) {
  arr.push({ x: i, y: i * 2, z: String(i) })
}

5. String hashing

Tests string interning and memory bandwidth with repeated character processing.

for (let i = 0; i < 20000; i++) {
  for (let j = 0; j < str.length; j++) {
    hash = ((hash << 5) - hash + str.charCodeAt(j)) | 0
  }
}

6. Matrix multiplication

Stresses ALU and cache lines with 64x64 matrix multiply using Float64Arrays.

Why identical CPUs produce different results

Factor	Effect
Silicon lottery	Manufacturing variance means each chip has slightly different transistor characteristics
Thermal throttling	CPU frequency varies with temperature
Power management	Battery vs plugged in, power plan settings
Background load	Other processes affect available compute
Cache state	L1/L2/L3 cache warming differs per run
CPU microcode	Different microcode versions affect scheduling

Entropy and stability

Property	Value
Entropy	~4 bits
Stability	0.50
Typical duration	8-15ms

Stability is lower than other collectors because timing varies with system load. However, the relative ratios between benchmarks are more stable than absolute values.

Role in risk scoring

The neoprint.attestDevice() function checks that hardware perf timings are realistic:

• All benchmarks should complete in > 0ms (instant = emulated environment)
• All benchmarks should complete in < 500ms (artificially slow = suspicious)

Web Worker offloading

Hardware perf is one of the collectors that runs in a Web Worker by default, keeping the main thread free during benchmarking.

Usage

const fp = await neoprint.get({ collectors: ['hardwarePerf'] })
const perf = fp.components.hardwarePerf.value

console.log(perf.floatArith)    // e.g. 2.1ms
console.log(perf.trigonometry)  // e.g. 1.8ms
console.log(perf.arraySort)    // e.g. 3.2ms
console.log(perf.objectAlloc)  // e.g. 1.5ms
console.log(perf.stringHash)   // e.g. 2.0ms
console.log(perf.matrixMul)    // e.g. 1.2ms