MELBOURNE BEACH, Fla.--(BUSINESS WIRE)--Gigantor Technologies Inc., an accelerator of deep neural network models for artificial intelligence, announced today the U.S. Patent and Trademark Office issued Patent # 11,099,854. The ‘Pipeline Operations in Neural Networks’ patent was issued on August 24, 2021, and was filed only four months prior without any conflicting claims. The newly published patent describes the next evolution of neural network acceleration for real-world applications that require real-time results.
"A truly exciting moment for Gigantor Technologies as we further strengthen our intellectual property portfolio," was the first reaction from Don Gaspar, CEO. Don goes on to explain, "GigaMACS™ is the solution to the problems which prevents AI's adoption for real-world scenarios. As neural network model complexity increases, GPUs and other solutions slow down proportionally and even lose data. GigaMACS™ operates at the input rate regardless of model complexity and maintains the same high performance with near-zero latency without any mathematical precision loss. No one else can make that claim."
Gigantor's patented technology easily processes high-definition 4K images at 240 FPS and faster. The secret to GigaMACS™ super-fast inference throughput is the customized parallel pipeline circuit and an innovative mass multiplier dedicated to each input channel. By accepting frames directly from the source and eliminating the RAM, outputs are returned while inputs are still being received with a recorded latency of less than half a millisecond. Neural network models on GigaMACS™ run identically to the original model on GPUs or TPUs, just much faster.
Mark Mathews, CTO, said, "I've never had a Patent move so quickly from Application to Issued, and with all Claims Allowed to boot. GigaMACS™ is a conceptually divergent direct inference implementation that takes full advantage of the previous investment in trained weights and model structure. Unlike GPUs and TPUs, a GigaMACS™ circuit avoids contention for shared resources like multipliers and RAM. A GigaMACS™ pipeline architecture accepts data at the incoming pixel frequency and need not drop frames to keep up while producing the same results as the original model."