Military startup aims large with neural processor chip

June 07, 2016 // By Peter Clarke
KnuPath
KnuEdge Inc. (San Diego, CA), a company that has raised more than $100 million since it was founded in 2005, has announced the KnuPath processor, a neuromorphic processor intended to accelerate voice recognition and other machine learning applications.

Dan Goldin, who led the National Aeronautics and Space Administration (NASA) throughout its renaissance in the 1990s, is the founder and CEO.

The KnuPath processor is already in production and in use at customer sites, KnuEdge said. At the same time the company launched KnuVerse, its voice-recognition and authentication software. The company said it has been supplying the software for military applications for five years.

KnuEdge has also divided itself into multiple operations to better drive sales forward. These are: hardware operation KnuPath, software operation KnuVerse and service operation Knurld. Knurld delivers the same neuromorphic computation technologies to a broader market via Knurld.io, a public cloud API service that allows developers to incorporate voice authentication into their proprietary products.

KnuPath
The KnuPath processor, codenamed Hermosa, is described as a Lambda Fabric processor and is generally intended to be deployed in datacenters where its ability to scale can be used. The KnuPath Lambda Fabric is desined to scale up to 512k devices and offers a rack-to-rack latency of 400ns, approaching half that of existing high-speed interconnects, the company said.

Each KnuPath processor includes 256 tDSP cores, 64 programmable DMA engines, an integrated L1 router to provide a maximum performance of 256-Gflops/s and 3,702Gbytes/s of memory bandwidth. There are 16 bi-directional I/O paths per processor chip delivering 320 Gbps-per processor and the chip consumes 34W.

Despite this hefty power consumption KnuEdge claims that when compared to the alternative processors available today its first-generation Lambda Fabric processors demonstrate an efficiency performance per watt advantage of between a factor of 2 and 6.


256 tDSP cores in one Hermosa processor. Source: KnuPath.

Put another way on its website KnuPath claims: "Hermosa outperforms the newest competing GPUs by 2.7 to 8.1x, as measured by key benchmarks like FFT, Sparse Matrix-Vector Multiply and k-means Clustering."

"Many of today's CPUs, GPUs and FPGAs force system designers to either create workarounds with last-generation chip sets or reduce their requirements for advanced-compute projects," said Goldin, in a statement. "After ten years of stealth development and rigorous testing, Lambda Fabric enters the market as mature technology which enables system designers to meet the most demanding requirements now, and also helps them rethink what is possible with neural computing in the future."

As can be seen from the package photograph ARM is involved in the KnuPath. However it is not clear from the sparse technical description at the KnuPath how it is involved. The KnuPath subsidiary of KnuEdge is based in Austin, Texas, which is a major site for ARM in North America.