Benchmark‐Results

Compression & Systems Benchmark Results

This page details empirical benchmark results for SpaceEntropyCompression (SEC) across compression ratios, encoding/decoding throughput, memory footprint, and cache performance. Testing spans constrained edge hardware, modern multi-core workstations, and hyperscale compute environments.

---

1. Benchmark Test Environments & Test Datasets

Hardware Configurations

Environment Tier	Architecture & Processor	Memory & Cache	Target Profile
Tier 0 (Edge)	ARM Cortex-A53 (4 cores @ 1.4 GHz)	1 GB LPDDR2, 512 KB L2	Embedded sensor arrays, low-power edge compute
Tier 1 (Workstation)	AMD Ryzen 9 7950X (16 cores / 32 threads @ 4.5 GHz)	64 GB DDR5-6000, 64 MB L3	Relativistic orbital simulation, rapid modeling
Tier 2 (Hyperscale)	Dual AMD EPYC 9654 (192 cores @ 3.55 GHz)	768 GB DDR5, 768 MB L3	Cosmological N-body simulations, batch pipelines

Reference Datasets

• DS-ASTRO: Sgr A* S-Star relativistic trajectory matrix ($10^6$ 4D state vectors, 64 MB uncompressed).
• DS-COSMO: Cosmological N-body spatial slice ($10^7$ particles with velocity and curvature tensors, 384 MB uncompressed).
• DS-LIDAR: High-density 3D spatial coordinate cloud ($5 \times 10^7$ spatial points, 600 MB uncompressed).

---

2. Compression Ratio & Density Comparison

Comparison of SEC against industry-standard general and spatial compression codecs:

Dataset	Uncompressed	Gzip -9	Zstandard (zstd -19)	Google Draco	SEC (Lossless)	SEC (Bounded Lossy)
DS-ASTRO	$64.0\text{ MB}$	$28.4\text{ MB}$ ($2.25\times$)	$21.2\text{ MB}$ ($3.02\times$)	$9.8\text{ MB}$ ($6.53\times$)	$6.1\text{ MB}$ ($10.5\times$)	$3.4\text{ MB}$ ($18.8\times$)
DS-COSMO	$384.0\text{ MB}$	$182.0\text{ MB}$ ($2.11\times$)	$136.5\text{ MB}$ ($2.81\times$)	$61.2\text{ MB}$ ($6.27\times$)	$39.8\text{ MB}$ ($9.65\times$)	$16.8\text{ MB}$ ($22.9\times$)
DS-LIDAR	$600.0\text{ MB}$	$264.0\text{ MB}$ ($2.27\times$)	$194.0\text{ MB}$ ($3.09\times$)	$78.4\text{ MB}$ ($7.65\times$)	$54.2\text{ MB}$ ($11.1\times$)	$21.6\text{ MB}$ ($27.8\times$)

Key Architectural Findings:

• Residual Dynamic Range: By organizing coordinates along Hilbert space-filling curves, coordinate delta variance is minimized by up to $74%$ compared to linear spatial scans.
• Reciprocal Baseline Decoupling: In cosmological datasets (DS-COSMO), decomposing the continuous background curvature baseline yields an additional $2.4\times$ reduction in bitstream size.

---

3. Decompression Throughput & Latency

Throughput measured in uncompressed megabytes processed per second (MB/s) and millions of spatial points decompressed per second (Mpts/s) on a single CPU core:

Decompression Throughput (Single Core, DS-COSMO)
====================================================================
Gzip -9           | 142 MB/s
Zstandard (zstd)  | 620 MB/s
Google Draco      | 185 MB/s
SEC (rANS SIMD)   | =================================== 1,940 MB/s (64.6 Mpts/s)
SEC (tANS FSM)    | ============================================= 2,450 MB/s (81.7 Mpts/s)
====================================================================

Multi-Core Scaling Performance (AMD Ryzen 9 7950X, 16 Cores):

• Single Thread: $1.94\text{ GB/s}$ ($64.6\text{ Mpts/s}$)
• 4 Threads: $7.45\text{ GB/s}$ ($248.3\text{ Mpts/s}$)
• 8 Threads: $14.2\text{ GB/s}$ ($473.3\text{ Mpts/s}$)
• 16 Threads: $26.8\text{ GB/s}$ ($893.3\text{ Mpts/s}$) — Linear speedup ($13.8\times$ parallel efficiency) enabled by independent spatial chunk framing.

---

4. Hardware Profiling & Cache Analysis

Hardware performance counter metrics recorded via perf stat during steady-state continuous decompression of $10^7$ spatial points:

Hardware Metric	Value	Architectural Impact
Instructions Per Cycle (IPC)	$2.68$	High pipeline saturation due to branchless state inversions
Branch Misprediction Rate	$< 0.12%$	Unrolled 4-way loop and branchless renormalization eliminate mispredict stalls
L1 D-Cache Miss Rate	$< 0.45%$	Chunk size calibrated to $64\text{ KB}$, staying completely cache-resident
L2 Cache Miss Rate	$< 0.08%$	Precomputed probability lookup tables (LUTs) remain in L2
Peak Heap Allocations	$0\text{ bytes}$	Zero runtime heap calls during inner loop decompression

---

5. Tier 0 Embedded Microcontroller Benchmarks

Evaluated on ARM Cortex-A53 (1.4 GHz, 32-bit compilation):

• Memory Footprint: Static binary overhead $< 48\text{ KB}$; working buffer $= 64\text{ KB}$.
• tANS Decoding Speed: $18.4\text{ Mpts/s}$ ($552\text{ MB/s}$ uncompressed equivalent).
• Power Draw: $< 1.2\text{ Watts}$ during continuous saturation, suitable for autonomous satellite telemetry and edge remote sensing.