Novel Mathematics. Real Quantum Hardware. Results That Speak for Themselves.
Novel mathematical algorithms and frameworks, verified on
real quantum hardware — and delivering measurable advantages
on classical systems. All results independently verifiable.
Novel Mathematics. Multiple Breakthroughs.
Phoenix Quantum Labs develops novel mathematical algorithms and frameworks that produce verified results on both quantum and classical hardware. The same core mathematics drives our quantum state-preparation and quantum RAM results, and our classical computation tools.
We discovered mathematical structures that unlock computational advantages on any hardware they're applied to — from consumer GPUs to quantum processors. Not simulations. Not incremental improvements. Fundamentally new mathematics.
Phoenix Attention — Long-Context KV-Cache Compression
Drop-in long-context serving for the NVIDIA Nemotron-H family (8B–120B). Compresses the KV cache while preserving retrieval: 244× KV-cache compression at 256K context, 2.27× more concurrent long-context requests per node, full needle retrieval (RULER niah 100, multikey 100, multivalue 97.5). Bit-exact prefill; quality matches the unmodified model; no retraining. Verified on 8× H100. Support for Llama and other transformer families is in active development.
VIEW BENCHMARKS →Aleph — Constant-Memory Long Context
Gives an LLM constant-memory long context: per-request memory stays flat as context grows, instead of a KV cache that grows with every token. Trained into NVIDIA Nemotron-H-4B (a natively 8K-context model), it extends long-context retrieval to 128K tokens — full RULER suite ~100% at 128K, where the base model scores 0% — at flat ~1.6 GB per-request memory, with the base model frozen so its reasoning and generation are unchanged. Transformer-hybrid support in active development.
VIEW BENCHMARKS →Phoenix SVD — Weight Decomposition
CUDA SVD engine for weight decomposition, model compression, and LoRA / PiSSA initialization. On real production model weight matrices at their real sizes: 9–17× on attention/FFN weights (16.7× on Nemotron q_proj, 19.1× on Llama 3.3 70B q_proj), near-lossless — singular values match torch to 10−9–10−14. Genomics single-cell PCA on a 1M-cell atlas: 180× vs CPU scikit-learn, machine-exact, where exact CPU SVD is intractable. Drop-in torch.linalg.svd-compatible API.
Quantum State Preparation
Constant-depth, noise-resilient target-state preparation on real quantum hardware. 0.97 fidelity vs 0.41 for a standard preparation at a short coherence time (T1 = 5 µs). Validated on hardware.
VIEW RESULTS →QRAM — Coherent Superposition Query
Coherent superposition QRAM query at 0.84 fidelity vs 0.29 for bucket-brigade, using ~95% fewer entangling gates. Verified on IBM quantum hardware.
VIEW RESULTS →VOIS — O(1) Search
GPU-native similarity search with constant-time retrieval. 2.5–4.2x faster than Meta's FAISS at every recall level. 197,000 queries/second on a consumer RTX 4060. Same math, classical hardware.
VIEW BENCHMARKS →PX Compute — GPU + CPU Memory Pool
Systems-level GPU + CPU memory pool allocator for workloads that recycle large buffers in tight loops — LLM KV caches, attention scratch, gradient buffers, embedding tables. 53–77× over cudaMalloc on large-block reuse (single H100), matching PyTorch's own caching allocator; 22.6× CPU vs libc malloc on 1 GB reuse.
Aegis-EC — Training-Integrity Codec
Catches silent data corruption (SDC) in long training runs the moment it happens — across weights, gradients, and checkpoints — so a corrupted run restarts from the last clean checkpoint instead of being found broken at the end. 100% single- and double-bit detection (0.0000% miss over 20M trials), word-level correction at 1.17% overhead vs SECDED's 12.5%. Fully GPU-accelerated on H100 (detect ~2.25 TB/s, correct ~1 TB/s) — a full-model scan costs a small fraction of one training step.
VIEW BENCHMARKS →Seeking Research Partners & Funding
We are actively seeking SBIR/STTR funding, research partnerships, and strategic investment. U.S. patents filed. Select results available under NDA.
CONTACT US →