Verkor Introduces VerTQ, TurboQuant Accelerator for LLM Inference

Verkor has introduced VerTQ, described as the industry's first TurboQuant accelerator silicon IP, announced in a press release. The technology implements the TurboQuant algorithm developed by Google Research, which cuts key-value cache memory usage in large language models by a factor of 4.3 while maintaining or improving performance.

VerTQ compresses KV data and accelerates attention operations, including Flash Attention and online SoftMax, directly on-chip without decompressing data. This approach reduces memory bandwidth requirements and increases inference efficiency, especially for applications where memory is limited.

The chip was built autonomously by Verkor's Conductor 2.0 AI platform using standard electronic design automation tools. The process took about 80 hours from algorithm to a verified FPGA implementation. Mapped to a Xilinx FPGA running at 125 MHz, VerTQ supports between one and thirty-two attention decoders.

VerTQ is designed for edge AI systems such as autonomous vehicles, drones, and robots, where compact design, low power use, and cost efficiency are key. The VerTQ customer package includes specifications, verification IP, testbenches, and FPGA images, and the product is available now.