AI and machine studying programs are working with knowledge units within the billions of entries, which implies speeds and feeds are extra vital than ever. Two new bulletins reinforce that time with a aim to hurry knowledge motion for AI.
For starters, Nvidia simply revealed new efficiency numbers for its H100 compute Hopper GPU in MLPerf 3.0, a outstanding benchmark for deep studying workloads. Naturally, Hopper surpassed its predecessor, the A100 Ampere product, in time-to-train measurements, and it’s additionally seeing improved efficiency because of software program optimizations.
MLPerf runs 1000’s of fashions and workloads designed to simulate actual world use. These workloads embody picture classification (ResNet 50 v1.5), pure language processing (BERT Massive), speech recognition (RNN-T), medical imaging (3D U-Internet), object detection (RetinaNet), and advice (DLRM).
Nvidia first revealed H100 check outcomes utilizing the MLPerf 2.1 benchmark again in September 2022. It confirmed the H100 was 4.5 occasions quicker than the A100 in numerous inference workloads. Utilizing the newer MLPerf 3.0 benchmark, the corporate’s H100 logged enhancements starting from 7% to 54% with MLPerf 3.0 vs MLPerf 2.1. Nvidia additionally stated the medical imaging mannequin was 30% quicker beneath MLPerf 3.0.
It needs to be famous that Nvidia ran the benchmarks, not an unbiased third-party. And Nvidia isn’t the one vendor working benchmarks. Dozens of others, together with Intel, ran their very own benchmarks and can probably see efficiency positive aspects as properly.
Community chip for AI
The second announcement is from Enfabrica Corp., which has emerged from stealth mode to announce a category of chips known as Accelerated Compute Material (ACF) processors. Enfabrica stated the chips are particularly designed for AI, machine studying, HPC, and in-memory databases to enhance scalability, efficiency and complete value of possession.
Enfabrica was based in 2020 by engineers from Broadcom, Google, Cisco, AWS and Intel. Its ACF answer was developed from the bottom as much as tackle the scaling problems with accelerated computing, which grows extra knowledge intensive by the minute.
The corporate claims that these units ship scalable, streaming, multi-terabit-per-second knowledge motion between GPUs, CPUs, accelerators, reminiscence and networking units. The processor eliminates tiers of latency and optimizes bottlenecks in top-of-rack community switches, server NICs, PCIe switches and CPU-controlled DRAM, in accordance with Enfabrica.
ACF will provide 50 occasions the DRAM enlargement over current GPU networks by way of Compute Specific Hyperlink (CXL), the high-speed community for sharing bodily reminiscence between servers.
Enfabrica has not set a launch date as of but however says an replace will likely be coming within the close to future.
Copyright © 2023 IDG Communications, Inc.
