NEUCHIPS' RecAccel N3000 system delivered 1.7x better perf-per-watt for inference DLRM while maintaining 99.9% accuracy leveraging its INT8 calibrator.
NEUCHIPS Inc. participated in MLPerf v3.0 with its RecAccel N3000 and demonstrated industry leading performance and power efficiency. The RecAccel N3000 system delivered 1.7x better perf-per-watt for inference DLRM while maintaining 99.9% accuracy leveraging its INT8 calibrator.
The RecAccel Quad-N3000 PCIe card is expected to increase perf-per-watt 2.2x while also delivering the lowest total cost of ownership (TCO). These results give cloud service providers confidence to choose a new solution on existing and future datacenter platforms.
MLPerf v3.0 testing was performed on a GIGABYTE G482-Z54 server with AMD EPYC 7452 32-core CPU and contained eight RecAccel-N3000-32G-PCIe cards. During the system testing, the RecAccel N3000 performance resulted in nearly 100% scaling across each card.
“We deliver high performance computing products that help intelligent systems effectively manage complex data sets,” said Kumaran Siva, corporate vice president, Strategic Business Development, AMD. “We are extremely proud of the joint work we have done with NEUCHIPS to achieve leadership performance and power efficiency for DLRM inference in MLPerf v3.0 using AMD EPYC processors. We look forward to continue working with NEUCHIPS to deliver industry leading AI solutions.”
“We were thrilled to participate in MLPerf v3.0 and achieve our goal to deliver the world’s most energy efficient DLRM inference platform with the first domain specific architecture,” said Youn-Long Lin, the CEO and Chairman of NEUCHIPS. “We look forward to working with the cloud ecosystem to support industry sustainability initiatives.”