Nvidia touts MLPerf 3.0 tests; Enfabrica details network chip for AI

AI and machine learning systems are working with data sets in the billions of entries, which means speeds and feeds are more important than ever. Two new announcements reinforce that point with a goal to speed data movement for AI.

For starters, Nvidia just published new performance numbers for its H100 compute Hopper GPU in MLPerf 3.0, a prominent benchmark for deep learning workloads. Naturally, Hopper surpassed its predecessor, the A100 Ampere product, in time-to-train measurements, and it’s also seeing improved performance thanks to software optimizations.

MLPerf runs thousands of models and workloads designed to simulate real world use. These workloads include image classification (ResNet 50 v1.5), natural language processing (BERT Large), speech recognition (RNN-T), medical imaging (3D U-Net), object detection (RetinaNet), and recommendation (DLRM).

Nvidia first published H100 test results using the MLPerf 2.1 benchmark back in September 2022. It showed the H100 was 4.5 times faster than the A100 in various inference workloads. Using the newer MLPerf 3.0 benchmark, the company’s H100 logged improvements ranging from 7% to 54% with MLPerf 3.0 vs MLPerf 2.1. Nvidia also said the medical imaging model was 30% faster under MLPerf 3.0.

It should be noted that Nvidia ran the benchmarks, not an independent third-party. And Nvidia isn’t the only vendor running benchmarks. Dozens of others, including Intel, ran their own benchmarks and will likely see performance gains as well.

Network chip for AI

The second announcement is from Enfabrica Corp., which has emerged from stealth mode to announce a class of chips called Accelerated Compute Fabric (ACF) processors. Enfabrica said the chips are specifically designed for AI, machine learning, HPC, and in-memory databases to improve scalability, performance and total cost of ownership.

Enfabrica was founded in 2020 by engineers from Broadcom, Google, Cisco, AWS and Intel. Its ACF solution was developed from the ground up to address the scaling issues of accelerated computing, which grows more data intensive by the minute.

The company claims that these devices deliver scalable, streaming, multi-terabit-per-second data movement between GPUs, CPUs, accelerators, memory and networking devices. The processor eliminates tiers of latency and optimizes bottlenecks in top-of-rack network switches, server NICs, PCIe switches and CPU-controlled DRAM, according to Enfabrica.

ACF will offer 50 times the DRAM expansion over existing GPU networks via Compute Express Link (CXL), the high-speed network for sharing physical memory between servers.

Enfabrica has not set a release date as of yet but says an update will be coming in the near future.

Source