The new set of MLPerf results proves it
The days and sometimes weeks it took to train AIs only a few years ago was a big reason behind the launch of billions of dollars-worth of new computing startups over the last few years—including Cerebras Systems, Graphcore, Habana Labs, and SambaNova Systems. In addition, Google, Intel, Nvidia and other established companies made their own similar amounts of internal investment (and sometimes acquisition). With the newest edition of the MLPerf training benchmark results, there’s clear evidence that the money was worth it.
The gains to AI training performance since MLPerf benchmarks began “managed to dramatically outstrip Moore’s Law,” says David Kanter, executive director of the MLPerf parent organization MLCommons. The increase in transistor density would account for a little more than doubling of performance between the early version of the MLPerf benchmarks and those from June 2021. But improvements to software as well as processor and computer architecture produced a 6.8-11-fold speedup for the best benchmark results. In the newest tests, called version 1.1, the best results improved by up to 2.3 times over those from June. Read More
Tag Archives: MLPerf
TSMC and Graphcore Prepare for AI Acceleration on 3nm
One of the side announcements made during TSMC’s Technology Symposium was that it already has customers on hand with product development progressing for its future 3nm process node technology. As we’ve reported on previously, TSMC is developing its 3nm for risk production next year, and high volume manufacturing in the second half of 2022, so at this time TSMC’s lead partners are already developing their future silicon on the initial versions of the 3nm PDKs.
One company highlighted during TSMC’s presentations was Graphcore. Graphcore is an AI silicon company that makes the IPU, an ‘Intelligence Processing Unit’, to accelerate ‘machine intelligence’. It recently announced its second generation Colossus Mk2 IPU, built on TSMC’s N7 manufacturing process, and featuring 59.2 billion transistors. The Mk2 has an effective core count of 1472 cores, that can run ~9000 threads for 250 Teraflops of FP16 AI training workloads. The company puts four of these chips together in a single 1U to enable 1 Petaflop, along with 450 GB of memory and a custom low-latency fabric design between the IPUs. Read More
A Benchmark for Machine Learning from an Academic/Industry Cooperative
MLPerf is a consortium involving more than 40 leading companies and university researchers, which has released several rounds of results. MLPerf’s goals are:
Accelerate progress in ML via fair and useful measurement
Encourage innovation across state-of-the-art ML systems
Serve both industrial and research communities
Enforce replicability to ensure reliable results
Keep benchmark effort affordable so all can play Read More
The Vision Behind MLPerf
A broad ML benchmark suite for measuring the performance of ML software frameworks, ML hardware accelerators, and ML cloud and edge platforms.
… since 2012 the amount of compute used in the largest AI training runs has been increasing exponentially with a 3.5 month-doubling time (by comparison, Moore’s Law had an 18-month doubling period). Since 2012, this metric has grown by more than 300,000x (an 18-month doubling period would yield only a 12x increase). Improvements in compute have been a key component of AI progress, so as long as this trend continues, it’s worth preparing for the implications of systems far outside today’s capabilities.” Read More
Rick's Cafe AI 