# Benchmark Results: maba-instant v1 ## Environment - Host: Intel Xeon Platinum 8370C @ 2.80GHz (2 vCPUs allocated) - Instructions: AVX-512F, AVX-512CD, AVX-512BW, AVX-512DQ, AVX-512VL - Compiler: GCC 13.3.0 (-O3 -march=native -ffast-math -fopenmp -std=c++20) - OS: Linux 6.6.137+ - Configuration: d = 256, L = 32 tokens, batch = 1 ## 1. Latency and Throughput Measured over 50 warmup iterations and 1,000 recorded iterations per model on identical input vectors: | Model | Paradigm | P50 (us) | P90 (us) | P99 (us) | Mean (us) | Throughput | | :--- | :--- | :--- | :--- | :--- | :--- | :--- | | maba-instant v1 (Tier-0 Exit) | Linear projection | 0.47 | 0.80 | 1.37 | 0.55 | 1,811,636 tok/s | | maba-instant v1 (Cascade) | 3-Tier JEV+DGDA+MLA | 3,341.48 | 4,777.16 | 8,742.22 | 3,826.68 | 8,362 tok/s | | Mamba-3 SSM (Step) | Selective state space | 132.92 | 210.92 | 263.92 | 149.44 | 7,298 steps/s | | RWKV-v6 (Step) | Linear recurrence | 205.20 | 239.90 | 639.93 | 226.34 | 4,422 steps/s | | LLaMA-3 AR (TTFT) | Causal attention | 732.81 | 980.92 | 2,956.80 | 877.68 | 1,139 tok/s | | LLaMA-3 AR (32 tokens) | Causal autoregressive | 25,410.97 | 32,999.95 | 58,468.28 | 27,536.60 | 1,162 tok/s | | Discrete Diffusion (32 tokens) | 16-step canvas flow | 26,848.30 | 31,771.26 | 52,105.50 | 28,227.44 | 1,133 tok/s | ## 2. Memory Usage Resident memory monitored via `/proc/self/status` during 1,000 iterations: | Model | Weights | Active State | KV Growth | Peak RSS | | :--- | :--- | :--- | :--- | :--- | | maba-instant v1 | 1.71 MB | 80 KB (static) | 0 bytes | 11.6 MB | | Mamba-3 SSM | 0.19 MB | 64 KB (static) | 0 bytes | 11.1 MB | | RWKV-v6 | 1.02 MB | 1 KB (static) | 0 bytes | 11.2 MB | | LLaMA-3 AR | 3.58 MB | 64 KB (32 tokens) | Linear with L | 16.8 MB | | Discrete Diffusion | 0.25 MB | 64 KB (static) | 0 bytes | 12.1 MB | ## 3. Order Book Matching Simulation Continuous double auction matching simulation with 60,000 order operations and simulated high-frequency taker flow: | Agent | Latency (us) | Net PnL (USD) | Maker Fills | Taker Fills | Cancels Done | Adverse Fills | Sharpe | Max Drawdown | | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- | | maba-instant v1 | 0.90 | +1,854.48 | 3,154 | 3,747 | 57,309 | 2,350 | 3.34 | $29.31 | | RWKV-v6 | 12.00 | +752.17 | 3,996 | 3,985 | 56,487 | 3,080 | 1.27 | $100.12 | | Mamba-3 SSM | 20.00 | +597.09 | 3,658 | 3,881 | 56,802 | 2,850 | 1.01 | $115.46 | | LLaMA-3 AR | 25,000.00 | -479.85 | 386 | 0 | 243 | 373 | -0.57 | $670.65 | | Discrete Diffusion | 26,000.00 | -405.60 | 394 | 0 | 223 | 361 | -0.44 | $635.62 | ## 4. Stability Test Multithreaded run executing 100,000 consecutive inference requests across 4 worker threads (`bin/maba_endurance`): - Total requests: 100,000 - NaN occurrences: 0 - Inf occurrences: 0 - Process RSS growth: 11.6 MB to 12.0 MB (delta +0.4 MB, bounded) - Exit code: 0