TriBinary is a reproducible, open-source scientific simulation laboratory and gate-level benchmark suite designed to answer a fundamental computer architecture research question:
Research Question: Under identical microarchitectural templates and physical silicon budget assumptions, how do binary, balanced-ternary, and quaternary GPU architectures compare in speed, dynamic energy, silicon area, and noise resilience across diverse compute workloads?
All findings are pre-registered and locked under git tag prereg-v1. Negative results are reported directly without bias.
flowchart TD
subgraph MultiSM ["Multi-SM GPU Architecture"]
GM["Global Memory (DRAM/HBM Macro)"] --> SM0["SM 0 (8 Lanes)"]
GM --> SM1["SM 1 (8 Lanes)"]
GM --> SM2["SM 2 (8 Lanes)"]
GM --> SM3["SM 3 (8 Lanes)"]
end
subgraph SM ["Streaming Multiprocessor (SM)"]
Warp["SIMT Warp Sequencer & PC"] --> Dec["Instruction Decoder"]
Dec --> RF["Register File Array (Macro: true)"]
RF --> ALU["Gate-Level Netlist ALU Lanes (0..7)"]
ALU --> SMEM["Shared Memory Scratchpad (Macro: true)"]
end
subgraph Systems ["Logic Systems Evaluated"]
GPUB["GPU-B: Binary Reference"]
GPUT["GPU-T: Balanced Ternary {-1, 0, +1}"]
GPUBT["GPU-BT: Control (Ternary Math on Binary Cells)"]
GPUQ1["GPU-Q-1b: Quaternary Logic, 1-bit Memory"]
GPUQ2["GPU-Q-2b: Quaternary Logic, 2-bit MLC Memory"]
end
-
Sign Detection vs Full Comparison (Finding 2 / H2):
-
Standalone Sign Detection: Balanced ternary is 79% to 96% slower and requires 125% to 150% more transistors than binary. In binary two's complement, sign extraction is an
$O(1)$ zero-gate operation (tapping the MSB wire), whereas balanced ternary must dynamically evaluate the most significant non-zero trit through an$O(\log n)$ priority tree. (H2 Sign-Detect REFUTED). -
Relational Comparator (
$a < b$ ): Balanced ternary is 20% to 24% faster and uses 33% to 41% fewer transistors. The native ternaryTCMPgate computes the full trichotomy ($<, =, >$ ) in a single logic level without a borrow/subtract carry-chain. (H2 Comparator SUPPORTED).
-
Standalone Sign Detection: Balanced ternary is 79% to 96% slower and requires 125% to 150% more transistors than binary. In binary two's complement, sign extraction is an
-
Multi-SM Effect Decomposition (H2 & H5):
- Math vs Hardware: Simulating ternary arithmetic on binary hardware (GPU-BT) consumes ~1.58x MORE energy than native binary due to 2-bit-per-trit encoding overhead. However, native ternary hardware (GPU-T) achieves an energy ratio of 0.36x to 0.44x relative to GPU-BT. The net energy saving (42.6% on GEMM, 34.9% on BitNet) is strictly a native hardware and interconnect advantage.
-
Noise Margin Penalty (H4):
- Gaussian noise sweeps confirm that reducing half-margins (
$0.50V$ in binary vs$0.25V$ in ternary vs$0.167V$ in quaternary) causes non-binary systems to experience symbol errors at substantially lower noise levels (crossover at$\sigma/V \approx 0.07$ ). (H4 SUPPORTED).
- Gaussian noise sweeps confirm that reducing half-margins (
- Python 3.10+
- Rust 1.75+ and Cargo
# Clone repository
git clone https://github.com/AybarsBarut/TriBinaryGPU.git
cd TriBinaryGPU
# Set up virtual environment
python -m venv .venv
.\.venv\Scripts\activate
# Install dependencies and build Rust simulator core
pip install --upgrade pip maturin pytest numpy scipy pyyaml httpx
maturin develop --releasetribinary verify --suite m1# Run 32-lane multi-SM benchmark across all 5 designs
tribinary run experiments/m3_full_gpu.yaml
# Run noise sensitivity sweep
tribinary run experiments/h4_noise.yaml
# Run parameter sensitivity analysis (1024 Monte Carlo samples)
tribinary run experiments/h1_sensitivity.yamltribinary report
# View report in docs/report.mdTriBinary integrates with NVIDIA AI Foundation endpoints (build.nvidia.com) via ai/nim_client.py:
- Red Team Auditor (
ai/agents/red_team.py): Usesnvidia/nemotron-3-ultra-550b-a55bto audit assumptions for bias. - Report Drafter (
ai/agents/report_drafter.py): Drafts report narratives enforcing the zero-hallucinated-number template rule. - Cell Designer (
ai/agents/cell_designer.py): Drafts candidate gate netlists with mandatory quarantine and verification gates.
To run the Red Team agent:
$env:NVIDIA_API_KEY="nvapi-..."
tribinary reviewLicensed under Apache-2.0. See CITATION.cff for citations in academic publications.