AMD nearly doubles Top500 supercomputer hardware share
Intel loses out as Instinct GPUs power the world’s fastest big-iron system
Analysis In a sign of how meteoric AMD's resurgence in high performance computing has become, the latest list of the world's 500 fastest publicly known supercomputers shows the chip designer has become a darling among organizations deploying x86-based HPC clusters.
The most eye-catching bit of AMD news among the supercomputing set is that the announcement of the Frontier supercomputer at the US Department of Energy's Oak Ridge National Laboratory, which displaced Japan's Arm-based Fugaku cluster for the No. 1 spot on the Top500 list of the world's most-powerful publicly known systems.
Top500 updates its list twice a year and published its most recent update on Monday.
Beyond securing the top spot, Frontier is also a big deal for AMD because it's the world's first publicly benchmarked exascale supercomputer, achieving a peak performance of 1.1 exaflops, based on the standard Linpack benchmark used to measure the world's top systems.
It was only a few years ago that Intel and the DOE proclaimed that the Intel-powered Aurora supercomputer would become the first exascale system in the US, but delays have pushed back the timeline to sometime later this year. These delays apparently prompted Intel sometime last year to revise a 2019 press release to change the originally stated delivery timeline from 2021 to 2022 and to remove the mention of Aurora being the first US exascale supercomputer.
As a fun side note, The Register noticed Intel edited its 2019 press release about Aurora to remove the mention of it being "the first exascale supercomputer" and to change the delivery date from 2021 to 2022. You don't often see companies editing old press releases like this. pic.twitter.com/4HqipenMPD
— Semiconductor News by Dylan Martin (@DylanOnChips) May 31, 2022
The important caveat to AMD's exascale win is that Frontier may not actually be the world's fastest supercomputer when considering systems that don't have publicly submitted benchmark results. That's because China apparently has not one but two systems that have reached a peak performance of 1.3 exaflops, but the systems' operators have yet to submit results to Top500.
AMD: From 6 to 93 supercomputers in five years
When AMD was launching its first-generation Epyc chips in 2017, the company's CPUs only accounted for six of the world's fastest 500 supercomputers.
Now AMD's CPUs are in 93 of the top 500, according to the spring 2022 update that was just released. That's nearly a fifth of the list, which is almost double the share AMD had in spring of last year.
Many of these AMD-powered systems are among the fastest on the list, with the chip designer's CPUs present in five of the top 10, 10 of the top 20, 26 of the top 50, and 41 of the top 100.
By contrast, Intel's CPU share of the Top500 has declined to 388 systems from 464 five years ago, with the list's spring 2022 update bringing the x86 giant's CPU share below four-fifths of total systems for the first time in nearly 10 years.
Intel's representation in the fastest supercomputers in the Top500 has diminished too, with the company's CPUs present in one of the top 10, five of the top 20, 15 of the top 50, and 46 of the top 100.
One of the things that has helped AMD gain traction over the last few years is the fact that its Epyc server CPUs have had higher core counts than Intel's Xeon CPUs, which makes Epyc well-suited for HPC applications that scale well with cores.
This is reflected in the latest Top500 list, with AMD Epyc cores accounting for 27 percent of the total cores across all systems. Meanwhile, Intel cores represent 45 percent of total cores, which makes sense give that its CPUs are still present in most systems.
It is important to remember that the HPC world isn't made up entirely of x86 chips. There are still a total of 19 supercomputers with chips that weren't designed by Intel or AMD.
Seven use IBM's Power chips, which power systems in the No. 4 and 5 spots and account for 5.5 percent of all cores. Five use Fujitsu's Arm-compatible A64FX chips, which power Japan's Fugaku system in the No. 2 spot and account for nine percent of all cores.
Four use the NEC Vector Engine chips, which represent a very low percentage of cores. The Sunway TaihuLight system in the No. 6 spot uses China's homegrown ShenWei chip, which represents 11.6 percent of all cores. Another in China uses the Hygon Dhyana chip, which uses AMD's first-generation Zen architecture as part of a joint venture.
AMD finally makes gain in Nvidia-dominated accelerator territory
As of this Top 500 list, 168 supercomputers in the Top500 are using GPUs. This reflects the fact that while HPC applications have increasingly taken advantage of such components, they are still not used by most of the world's fastest systems. For those using accelerators, 157 are from Nvidia. None of the other accelerator vendors come even close to Nvidia's share, even AMD, which has only had one system using its GPUs for the last few years.
However, AMD did finally see a slight uptick in GPU share with the Top500's spring 2022 update, thanks to seven new systems that combine AMD's third-generation Epyc chips with its new Instinct MI250X GPUs, which are more competitive against Nvidia's GPUs than previous products.
Among these seven systems with AMD CPUs and GPUs are Frontier and two other supercomputers in the top 10 in the list. According to an AMD spokesperson, the Epyc chips used in these systems are code-named Trento and have optimizations in their I/O die to create cache coherency with the MI250X, which essentially allows the two processor types to share memory more easily.
AMD clearly has a way to go before it can take any meaningful GPU share from Nvidia, especially considering the fact that the overall number of systems with accelerators continues to increase, which gives Nvidia an opportunity to defend its footprint.
Outside of AMD and Nvidia, there are a few curiosities in the accelerator realm of the Top500. Two systems, for instance, still use Intel's discontinued Xeon Phi accelerators. One system in China uses the Matrix-2000 accelerator developed by the country's National University of Defense Technology. Another in China uses a homegrown accelerator vaguely called the "Deep Computing Processor."
Then there's Japan, which has two systems using homegrown accelerators. One is called the PEZY-SC3, which was developed by the country's PEZY Computing semiconductor company. The other is called the MN-Core, which was developed by Japan's Preferred Networks.
The latest update from Top500 shows momentum for AMD in multiple ways, but we should remember that Intel, with a new comeback plan from CEO Pat Gelsinger, is hungry to make up for the mistakes it has made in the past several years and create more competitive chips again.
There's also the potential for more systems in the future using chips based on Arm and other alternative architectures, especially given that Europe and countries like China are increasingly looking at designing their own. Thus, AMD shouldn't get complacent. ®