Americas

  • United States

Intel stumbles in supercomputer rankings, delays next-gen CPU

Opinion
Jul 05, 20214 mins
Data Center

Intel-based supercomputers make up fewer of the TOP500 fastest than previously as the company puts the brakes on its Xeon Scalable CPUs, codenamed Sapphire Rapids, until next April at the earliest.

Data center corridor of servers with abstract overlay of digital connections.
Credit: Sdecoret / Getty Images

This week the TOP500 list of the world’s fastest supercomuters found that, once again, Fugaku is number one, benchmarking at 442 Pflop/sec, making it three times faster than the second place machine. Impressive, but also indicative that it might also be the first to break the exaflop barrier if it’s working on the right kind of problem.

TOP500 pointed out that Fugaku’s score (and everyone else’s) is based on double-precision benchmarks, the most accurate floating point math calculation you can do. But much of AI and machine learning is single-precision, which can be less than half the compute power of double precision.

So when running in single-precision mode, Fugaku’s peak performance might actually be above an exaflop. When running the new HPL-AI benchmark, Fugaku clocks in with 2 Eflop/sec. But it won’t be official until they are doing exaflops of double-precision math.

And get ready for the Fugaku clones. The Arm-based A64FX processor designed by Fujitsu for Fugaku is available for sale. Servers built on the chip will be sold by HPE’s Cray subsidiary, and there were four other A64FX systems on the list (at #14, #25, #63, and #288), all in Japan for now, but given Fugaku’s impressive performance and HPE’s reach, there will likely be more.

AMD’s big leap

AMD’s gains continue to impress. It has crossed the 10% barrier of server sales, according to Mercury Research, and now it’s products account for darn near 10% of the TOP500. The company’s EPYC processors are found in 49 of the world’s top supercomputers, up from 21 on the most recent previous list from November 2020 list and more than four times the 11 supercomputers on last June’s list.

Out of the 49 supercomputers using EPYC CPUs, 29 of them are new entries and three are in the top 10, includes Perlmutter, a new entry, and Nvidia’s Selene (#6) and JUWELS Booster Module (#8), a Bull Sequana design in Germany.

Intel setbacks

AMD’s rise has come at the expense of Intel, which saw its share of the top 500 supercomputers shrink from 470 last June to 431 this year.

And Intel has suffered a setback on its attempt to right the ship. It announced earlier this week it has delayed production of its next-generation Xeon Scalable CPUs, codenamed “Sapphire Rapids,” to the first quarter of 2022 and said it will start ramping shipments by April of next year at the earliest.

Lisa Spelman, the newly anointed head of Intel’s Xeon and Memory Group, said in a blog post announcing the delay that the new CPUs would come with two new features: the next generation of Deep Learning Boost and an acceleration engine called Intel Data Streaming Accelerator. Spelman said Intel is delaying the chips because it needs extra time to validate the CPU.

“Given the breadth of enhancements in Sapphire Rapids, we are incorporating additional validation time prior to the production release, which will streamline the deployment process for our customers and partners. Based on this, we now expect Sapphire Rapids to be in production in the first quarter of 2022, with ramp beginning in the second quarter of 2022,” Spelman wrote.

Supermicro is the greenest (for now)

Fugaku wasn’t just the fastest supercomputer, for a while it was the greenest, meaning the most energy efficient. That was in part to the fact it has no DRAM DIMMs. The memory is on the CPU silicon die. When you have supercomputers with thousands of nodes, DRAM memory consumption adds up fast.

But Fugaku’s green title taken by a Japanese computer, the MN-3 from Supermicro and Preferred Network. The MN-3 is comprised of Intel Xeon CPUs and MN-Core boards developed by Preferred Networks. The MN-3 delivers 29.7 Gigaflops of performance per watt on a benchmark run that showed a total of 1.62 Petaflops performance. It wasn’t exactly a top performer, coming in at #335 on the list, though.

Right behind it, at 29.52 gigaflops per watt, was the creatively-named HiPerGator AI, an EPYC/Nvidia DGX A100 powered computer at the University of Florida (HiPerGator, get it? From the place that gave us Gatorade). HiPerGator came in at #22 on the list.