DeepCube Unveils Deep Learning Suite
DeepCube is launching a new suite of products and services, including a software-based deep learning accelerator which runs on a range of current and next-generation hardware.
The suite targets neural network training and inference while the accelerator is designed to boost of the performance of existing hardware. The suite includes an automated training framework dubbed CubeIQ; an inference engine for running new deep learning models; and a CubeAdvisor service designed to assist customers in deploying deep learning models. The service addresses the difficulty of moving models from development to production.
“The product suite collectively serves as an inference accelerator that enables a 10-fold boost to CPU’s, GPU’s and other hardware, AMD’s second-generation Epyc processors being one example,” DeepCube said.
In one example, CubeIQ and CubeEngine ran on AMD’s second-generation Epyc-based cloud instances and are coupled to improvements in current and future generations of the AMD CPU family. The combination is said to yield improved inference performance on neural networks ranging from ResNet-50 image recognition to BERT-Large.
Tel Aviv-based DeepCube said it would release those AI benchmarks in an upcoming blog post.
Noting that many early adopters remain unable to transition their AI models to production, DeepCube said Wednesday (Feb. 24) it is aiming its deep learning suite at technical challenges associated with training and inference—”which is no easy feat,” added DeepCube CEO Michael Zimmerman.
The commercial versions of DeepCube’s proprietary technology illustrates how DeepCube is “taking steps toward democratizing deep learning across industries,” Zimmerman added.
CubeIQ trains models using “pre-knowledge” of the goal and operating environment, the company said. A reduction in model size results in performance increases and reduced computing requirements while creating new edge deployment options without sacrificing model accuracy.
DeepCube’s acceleration platform is designed to help deploy deep learning models in data centers as well as within smart edge devices.
Meanwhile, the inference engine targets new deep learning models, dynamically assigning kernels best suited to specific hardware and deep learning models. That composable approach differs from current monolithic inference engines, the company claims.
The CubeAdvisor service provides guidance from deep learning experts and data scientists. The service is geared toward deploying models in production based on cost, performance, power and latency requirements.
DeepCube worked with AMD (NASDAQ: AMD) to test the deep learning suite on cloud instances using second-generation Epyc processors, the partners said. The suite, “coupled with optimizations specific to current and future generation AMD Epyc CPUs,” will help boost inference throughput and reduce latency, said Kumaran Siva, head of AMD’s Epyc cloud unit.
DeepCube notes that its technology is hardware agnostic; it also ran benchmarking tests with other chip makers besides AMD.
DeepCube said it is offering a free trial license for CubeIQ and CubeEngine, with CubeAdvisor as an option.
-Editor’s note: A previous version of this story incorrectly stated the DeepCube suite was powered by AMD processors. Rather, the company’s technology boosts efficiency for a range of hardware.