Xilinx wins first place in MLPerf for Resnet-50 test

Recently, the MLPerf organization announced a series of test results. NVIDIA won all tests in all six application areas for data centers and edge computing systems in the second version of MLPerf Inference. Among them, Inspur AI server NF5488A5 supports 8 third-generation NVlink fully interconnected NVIDIA A100 GPUs, setting 18 performance records in one fell swoop; Ningchang Information Industry (Beijing) Co., Ltd.'s Nettrix X640 G30 AI server equipped with NVIDIA T4 /A100 GPU cards achieved 30 world firsts in benchmarks such as ResNet, BERT, and DLRM.

Although many of these achievements are attributed to NVIDIA, FPGA vendor Xilinx also received some credit, as shown in the following from Xilinx’s official blog:

The latest MLPerf organization released the latest round of machine learning performance test results. Launched in 2018, MLPerf is an open source community of more than 23 organizations whose mission is to define a set of standardized ML benchmarks. The group's ML inference benchmark provides a commonly recognized process to measure how quickly and efficiently different types of accelerators and systems can execute trained neural networks.

This is the first time Xilinx has participated directly in this test. Although we only achieved a small score, we are very happy to have achieved leadership in the image classification category. We worked with Mipsology to rigorously undergo the standard test.

The test system uses an Alveo U250 accelerator card based on a Mipsology-optimized Domain Specific Architecture (DSA). Our custom Alveo-based DSA performs image classification tasks based on the ResNet-50 benchmark at 5011 images/second in offline mode. ResNet-50 measures image classification performance in images/second.

We achieved the highest performance/peak (TOP/s). It is a measure of performance efficiency, which essentially means that given X amount of peak hardware computation, we provide the highest throughput performance.

The MLPerf results also show that we achieved 100% of available TOP/s compared to our published datasheet performance. This impressive result demonstrates that raw peak on paper is not always real-world performance. Our device architecture delivers improved efficiency (effective TOP/s compared to peak TOP/s) for AI applications. Most vendors in the market can only deliver a fraction of their peak performance, typically reaching 40% efficiency. More importantly, ML applications are more than just AI processing. They often require pre- and post-ML processing functions that compete for system bandwidth and cause system-level bottlenecks. The power of our adaptable platforms is that they can also accelerate these critical non-AI functions and build application-level streaming pipelines to avoid system bottlenecks, thereby achieving overall application acceleration. We have achieved leading results while maintaining the programmability of the TensorFlow and PyTorch frameworks, without requiring users to have hardware expertise.

MLPerf is quickly becoming the industry standard for measuring ML performance. This is the second version of the MLPerf Inference Benchmark (v0.7), which attracted more than 1,200 peers to review. ML inference is a fast-growing market, such as autonomous driving and AI-based video surveillance, which requires computer vision tasks such as image classification and object detection. These complex computing workloads require different levels of throughput, latency, and power to run efficiently, which is where Xilinx and our adaptive computing products shine.

"The MLPerf benchmark results highlight the efficient throughput and low-latency performance our adaptive compute devices deliver for AI applications. We are excited about these initial MLPerf results and look forward to participating in the next release."