FRACTAL-BLT // Zero-Allocation .NET 10 MoE Runtime

CLI SIMULATOR // fractal-core.exe ARCH: x64-native // THREADS: 1 // GPU: RTX 5070 Ti

FRACTAL-CORE v2.0.0 [NativeAOT Release Build]
Initializing zero-allocation memory pools... [OK]
Type 'help' to inspect available runtime commands.

>

root@fractal:~#

REAL-TIME TELEMETRY [ ILLUSTRATIVE TARGET ]

Component	Status / Metric
Garbage Collector	DORMANT
Hot-Path Allocations	0 BYTES
Memory Footprint	10.06 MB (STATIC)
BLT Entropy Encoder	45.05 MB/s
GNN Adjacency Router	541.3 µs / Matrix
PCIe DMA Transfer	Direct O_DIRECT

ZERO-ALLOCATION INFERENCE SHOWCASE

Live SSE Telemetry captured via Windows Terminal and Run-Demo.ps1.

FRACTAL-BLT vs. PYTHON/PYTORCH

STARTUP MEMORY OVERHEAD

PyTorch: ~2.4 GB

Fractal-BLT: 10.06 MB

GIL / THREADING

[!] PYTHON GIL BLOCKED

[✓] FRACTAL: LOCK-FREE / TRUE CONCURRENCY

Traceback (most recent call last):
File "train.py", line 42, in <module>
RuntimeError: CUDA out of memory. Tried to allocate 1.24 GiB. Python interpreter overhead detected.

PIPELINE ARCHITECTURE (INTERACTIVE)

Click on a component in the data path to inspect its implementation details.

NVMe Drive

→

BLT Encoder

→

GNN Router

→

PCIe DMA

→

RTX GPU

Select a module above.

Information will appear here.

ARCHITECTURE Q&A

HOW DOES FRACTAL-BLT ELIMINATE HEAP ALLOCATIONS?

By aggressively utilizing `ReadOnlySpan`, `stackalloc` unmanaged stack buffers, and C# 12 `[InlineArray]` semantics. We do not use the `new` keyword on the hot path. The Garbage Collector remains completely asleep during inference.

WHY RAW PTX INSTEAD OF CUBLAS/TORCH?

Embedding JIT-compiled PTX assembly as string literals within the C# binary guarantees a zero-dependency, self-contained standalone executable. We interact exclusively with the CUDA Driver API (`cuLaunchKernel`), skipping massive C++ abstraction bloat.

PATCHING THROUGHPUT (MB/s)

0 MB/s

Python (BPE)

0 MB/s

Fractal-BLT (Entropy)

SYSTEM PROFILES

Metric	Fractal-BLT	PyTorch Baseline
Startup Memory	10.06 MB	2.4 GB+
JIT Time	~20ms (NativeAOT)	Seconds (PT2 compile)
Thread Concurrency	Lock-Free	GIL Blocked
Routing Strategy	1-Hop Stackalloc GCN	VRAM-bound MLP
Binary Size	Self-Contained (4 MB)	Massive Wheel (GBs)

OPENAI-COMPATIBLE SSE ENDPOINT

Simulate a streaming response from the /v1/chat/completions endpoint.

📄 FULL API REF 🚀 GETTING STARTED 💻 CODE EXAMPLES

// Waiting for request...

CURL REQUEST

curl -N -X POST http://localhost:5000/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "model": "fractal-moe-64x", "messages": [ {"role": "user", "content": "Compute"} ], "stream": true }'