Mohit Madan@icyzh

High Energy Physics (HEP) experiments produce massive amounts of detector data, generating around 15 PB per year. Storage capacity already constrains what can be retained long-term. Current pipelines rely on standard lossless compressors (like LZMA) that treat IEEE-754 float32 byte streams as unstructured data, ignoring the internal structural heterogeneity of floating-point formats. To solve this, we introduce Vortex Codec, featuring a novel architectural prior called TDTEmbedding (Type-Decomposed Token Embedding). Instead of a single shared lookup table, TDTEmbedding maintains four independent tables indexed by byte position modulo 4. This allows the neural sequence model to learn specialized representations for low-entropy exponent bytes and high-entropy mantissa bytes from scratch. We integrated this with a Compressive Transformer using Learnable Token Eviction (LTE) and an arithmetic range coder. Our 87M parameter model was trained and profiled using an MI300X GPU on the AMD Developer Cloud. The results establish state of the art archival efficiency: Vortex Codec achieves 3.397 and 5.063 bits-per-byte (BPD) on the ATLAS FTAG and CamelFloat32 benchmarks, representing a 28.0% and 19.5% improvement over LZMA-9.

Grab your ideas, spin up compute, and start building what’s next! Start with $100 in AMD Developer Cloud credits and begin building right away. Get access to hands-on tutorials, workshops, and curated learning resources to help you build faster and go further. To unlock credits and join the hackathon, sign up for the AMD AI Developer Program. 🤝 Go solo or team up — ship something real. 💰 $21,500+ total prize pool & 1 × AMD Radeon™ AI PRO R9700 GPU up for grabs. 📍 On-site Venue (May 9-10): MindsDB SF AI Collective 3154 17th St, San Francisco, California, USA 📍On-site participation is by invitation only. Travel and accommodation expenses will not be covered. 🧑‍💻 Developers, builders, and innovators - bring your boldest ideas and let’s turn them into real AI systems on AMD.