#United States #Los Altos #LLM #TurboQuant #Verkor

Verkor Revolutionizes AI with TurboQuant Accelerator Chip

In a remarkable stride for artificial intelligence and semiconductor technology, Verkor, Inc. has launched the first TurboQuant silicon IP, namely VerTQ. This innovative accelerator chip is designed specifically for enhancing the performance of Large Language Models (LLMs) by employing Google's cutting-edge TurboQuant algorithm. The technological innovation brought forth by VerTQ allows for an impressive 4.3 times reduction in KV cache memory usage, making it a game-changer in a landscape where memory resources are increasingly scarce.

The TurboQuant algorithm was mathematically announced by Google on March 24, 2026, and VerTQ stands as the first hardware implementation of this significant advancement. The announcement had immediate repercussions on the stock prices of major memory chip manufacturers, highlighting the algorithm's impact on the industry. VerTQ effectively manages the compression of kernel vector (KV) data while simultaneously accelerating essential computational processes, such as the Attention mechanism, performed on-chip.

One of the key features of VerTQ lies in its ability to perform Flash Attention operations, including online SoftMax computations, all without needing to decompress KV-cache data. This feature not only significantly enhances computational efficiency but also conserves valuable memory bandwidth essential for robust AI processing.

VerTQ's scalable architecture can support between one and 32 attention decoders, offering remarkable flexibility for different applications. Mapped onto a Xilinx XCVU29P-3 FPGA running at 125 MHz, the accelerator utilizes 500,619 LUTs, 247,022 FF, and 748 DSP48E2 for a single attention decoder. The remaining resources include 4 RAMB36 and 9 RAMB18 blocks, demonstrating a well-optimized design aimed at expanding inference capabilities in edge AI applications, including automobiles, drones, and robotics.

Built using Verkor's innovative Conductor 2.0 platform, VerTQ was autonomously developed within approximately 80 hours, showcasing an astounding reduction in the chip development cycle from years to mere weeks.

“Conductor 2.0 drastically compresses the chip development timeline while continuously evolving to handle increasingly complex designs,” stated Suresh Krishna, CEO of Verkor. This autonomous design automation platform has enabled Verkor to produce progressively larger and more intricate silicon IPs derived from impactful algorithms, ensuring practical and timely solutions for industry needs.

The deliverable package for VerTQ includes comprehensive resources such as product specifications, microarchitecture details, a test plan, verification IP, and hierarchical RTL documentation, making it ready for immediate deployment.

As of now, Verkor, guided by top-tier AI researchers and semiconductor veterans, emphasizes the potential of VerTQ and its Conductor 2.0 platform in various sectors, including automotive, edge computing, security, and more. This strategic focus highlights Verkor's ambition to lead in the realm of enterprise agentic AI and semiconductor design automation.

If you are interested in exploring more about VerTQ and what Verkor's technology can offer to your business or research, don't hesitate to reach out to the company for further information.

Discover how Verkor is redefining the landscape of AI technology with innovative solutions such as the TurboQuant accelerator chip.