Supercomputer Installations Drop in Second Half of 2026

Supercomputer Installations Drop in Second Half of 2026
3 min read

The November 2026 Top500 ranking shows a sharp slowdown in new supercomputer installations. Only 27 machines made their debut, a figure significantly below usual expectations and the lowest turnover recorded since the first ranking in 1993.

The Top500 is published twice a year, in June during the ISC in Europe, and then in November at the SC conference in the United States. To be listed, a machine must have been tested with the HPL (High Performance LINPACK) benchmark, which measures its computing power. According to the column’s author, this decline may reflect a shift in priorities, with some organizations preferring to rent capacity from clouds rather than buying a machine every four or five years and building a data center to house it. However, he offers another explanation: some organizations may not yet have been ready to test equipment currently being installed, or that will be installed in early 2027.

The product lines announced for the coming year are central to this analysis. Nvidia is preparing its “Rubin” GR200 GPUs and its “Vera” CV100 CPUs, while AMD is deploying its “Altair” MI400 GPUs and its “Venice” Epyc 9006 processors. Intel does not have a scientific computing-oriented GPU on the market, according to the source. The author also notes that the “Diamond Rapids” Xeon 7 processor, with 256 performance cores, could arrive a year late if launched in the second half of 2027, as rumors suggest. During its third-quarter earnings presentation, Intel indicated that the 16-memory-channel version of this chip was being sampled. However, he believes the slowdown is mainly due to roadmap timing, particularly for networking, which is stuck on 800 Gb/s ports while 1.6 Tb/s ports are expected.

A List Dominated by Accelerators

The three compilers of the ranking, Erich Strohmaier of Lawrence Berkeley National Lab, Jack Dongarra of the University of Tennessee and Oak Ridge National Laboratory, and Horst Simon of the ADIA Lab in Abu Dhabi, write that “the result does not indicate that performance has stopped growing.” They add: “Rather, it shows that a larger share of the progress comes from additions and upgrades below the top, rather than from complete replacement of the most advanced installations.”

Among the 27 new machines, the most powerful is “Hangang,” installed at the KISTI supercomputing center in South Korea. It is based on Nvidia Grace-Hopper superchips, HPE Cray EX254n server nodes, and the Slingshot 400 interconnect, and achieves a peak power of 588.3 petaflops. The second system is “Lux,” from Oak Ridge National Laboratory, with 323.9 petaflops. In total, the 27 machines represent 1.32 exaflops of peak power. Accelerated machines, equipped with GPUs or other high-performance computing processors, now number 290 out of the 500 in the ranking, accounting for 58% of systems, 70.4% of cores, and 79.9% of total peak power.

One final point strikes the author: a single cloud, the eSurfing Cloud of China Telecom in Shanghai, accounts for 93.3% of the cores of the 27 new machines, but only 2.9% of their peak power. He finally notes that the Top500 only lists systems that have voluntarily submitted their results and does not constitute an exhaustive inventory of equipment worldwide.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x