TIVRA तीव्र

Overview

TIVRA is the hardware acceleration layer that lets yakmesh exploit every silicon feature on modern hardware — AVX-512, SHA-NI, NVIDIA CUDA, AMD NPU (XDNA), and ONNX Runtime — for post-quantum cryptography operations and ML-based network intelligence.

Every call to SHA3-256, ML-DSA-65, ML-KEM-768, or model inference routes through TIVRA. If native acceleration is available, it is used automatically. If not, pure-JS fallbacks from @noble/hashes and @noble/post-quantum provide correct results on every platform.

Three-Tier Architecture

Hardware Detection

probe() runs once at startup and populates the global HW flags object:

import { initialize, HW } from 'yakmesh/utils/accel';

await initialize();

console.log(HW.cpuModel);    // 'AMD Ryzen 7 8700F'
console.log(HW.avx512);      // true  (Zen 4)
console.log(HW.nativeSha3);  // true  (OpenSSL 3.x)
console.log(HW.amdNpu);      // true  (XDNA NPU)
console.log(HW.amdNpuTops);  // 16    (TOPS rating)

SHA3-256 Acceleration

Every hash in yakmesh — content addresses, attestation fingerprints, seed expansion — uses SHA3-256. On Zen 4 hardware with native OpenSSL, TIVRA delivers 4.6× faster hashing vs pure JavaScript.

import { sha3_256, sha3_256hex } from 'yakmesh/utils/accel';

// Returns Uint8Array (32 bytes)
const hash = sha3_256(data);

// Returns hex string
const hex = sha3_256hex(data);

// Automatic path selection:
// HW.nativeSha3 = true  → crypto.createHash('sha3-256')  [4.6× faster]
// HW.nativeSha3 = false → @noble/hashes sha3_256          [always correct]

Post-Quantum Cryptography

ML-DSA-65 (Digital Signatures)

FIPS 204 lattice signatures used for DOKO certificates, MANTRA gossip, and gateway attestation. Native liboqs acceleration delivers ~10× faster signing and verification.

ML-KEM-768 (Key Encapsulation)

FIPS 203 lattice KEM used by ANNEX for session establishment. Seeds are sourced from PRAHARI quantum entropy when available.

import { mlKem768Keygen, mlKem768Encapsulate, mlKem768Decapsulate } from 'yakmesh/utils/accel';

// Key generation (with optional PRAHARI seed)
const { publicKey, secretKey } = mlKem768Keygen(hybridSeed);

// Encapsulation (sender)
const { cipherText, sharedSecret } = mlKem768Encapsulate(publicKey);

// Decapsulation (receiver)
const sharedSecret2 = mlKem768Decapsulate(cipherText, secretKey);

GPU Batch Verification

The BatchVerifyQueue collects ML-DSA-65 verification requests and processes them in batches for higher throughput. When an NVIDIA GPU with CUDA is available, the queue targets GPU parallel processing; otherwise it uses parallel CPU verification.

import { batchVerify } from 'yakmesh/utils/accel';

// Enqueue verification — returns Promise<boolean>
const valid = await batchVerify.enqueue(signature, message, publicKey);

// Queue automatically flushes when:
// - batch reaches minBatchSize (8), or
// - flushInterval (5ms) elapses

NPU Inference Engine

The InferenceEngine manages ONNX model sessions and routes inference to the best available accelerator. Provider priority:

import { inference } from 'yakmesh/utils/accel';

// Load ONNX model
await inference.loadModel('sakshi-anomaly', './models/sakshi-anomaly.onnx');

// Run inference — automatically routes to NPU/GPU/CPU
const result = await inference.infer('sakshi-anomaly', {
  features: new Float32Array([0.8, 0.2, 0.5, ...]),
});

// Check provider
console.log(inference.provider);      // 'DmlExecutionProvider'
console.log(inference.isAccelerated); // true

Telemetry

TIVRA tracks every call and native-path hit rate for performance monitoring:

import { getTelemetry, getStatus } from 'yakmesh/utils/accel';

const t = getTelemetry();
// {
//   sha3Calls: 14200,  sha3NativeHits: 14200,  sha3NativeRate: '100.0%',
//   signCalls: 380,    signNativeHits: 380,     signNativeRate: '100.0%',
//   verifyCalls: 1240, verifyNativeHits: 1240,  verifyNativeRate: '100.0%',
//   inferCalls: 45,    inferNpuHits: 45,        inferAccelRate: '100.0%',
// }

const status = getStatus();
// { hardware: { cpu, simd, gpu, npu }, acceleration: { sha3, pqCrypto, batchVerify, inference } }

Integration Points

TIVRA is wired into 12+ hot-path files across the codebase:

API Reference

Version History