2026年下半年超級電腦裝機量驟降,Top500 11月榜單僅新增27台
Supercomputer Installs Take A Dive In The Second Half Of 2026
Top500 2026年11月榜單只新增27台超級電腦,汰換量是自1993年以來68份榜單中最低。作者認為,裝機停滯更可能與路線圖時序有關,尤其網絡仍使用800 Gb/sec端口,而市場期待1.6 Tb/sec端口;Top500維護者指出,性能提升較多來自榜單頂端以下系統的新增和升級,而非領先系統整體換代。
Given that compute engines and networks are constantly being updated and that HPC centers the world over have their own budget and upgrade cycles, there has been a fairly constant level of new machines being installed at any given time.
This has always been reflected in the Top500 supercomputer rankings that come out twice a year – first in June with the International Supercomputing (ISC) event in Europe and then again in November with the Super Computing (SC) event held in the United States. But on the November 2026 Top 500 list, new machine installations – and to be precise, new machines that have been tested using the High Performance LINPACK (HPL) benchmark test to gauge their relative performance – stalled and the number of machines is about half what we normally expect.
NEXTPLATFORM AD
This is not a good sign, and it might show perhaps how priorities are shifting to renting capacity on clouds instead of trying to acquire a new machine every four or five years and building a datacenter to support ever-hotter and ever-more-expensive HPC systems.
But don’t freak out just yet. It could also just be that companies were not ready to test iron that is being installed now or early in 2027 as Nvidia rolls out its “Rubin” GR200 GPUs and “Vera” CV100 CPUs or AMD rolls out its “Altair” MI400 GPUs and its “Venice” Epyc 9006 processors. Intel does not have an HPC-focused GPU in the market (although it does have ones aimed at inference), but there are also probably some customers who have acquired future “Diamond Rapids” Xeon 7 CPU with 256 P-cores, which will be about a year late to market if it debuts in the second half of 2027 using Intel’s 18A-P process (1.8 nanometers) as the rumor mill is suggesting. Intel said on its call with Wall Street analysts going over its third quarter financial results (when I was having major surgery, so I missed it) that the Diamond Rapids chip with 16 memory channels is sampling now. The mainstream Diamond Rapids with only eight memory channels has been killed off and the existing “Granite Rapids” Xeon 6+ will fill in the gap until Intel does a Xeon 8 refresh sometime in the future.
“The result does not indicate that performance has stopped growing,” write the caretakers of the Top500 list. (That would be Erich Strohmaier, of the Lawrence Berkeley National Lab, Jack Dongarra, of the University of Tennessee and Oak Ridge National Lab, and Horst Simon, of the ADIA Lab in Abu Dhabi.) “Instead, it shows that a larger share of improvement is coming from additions and upgrades below the very top rather than from wholesale replacement of leading installations.”
All that said, I still believe the stall is more about roadmap timing – particularly with networking, which is stalled at 800 Gb/sec ports when the world is expecting 1.6 Tb/sec ports – than it is moving workloads to the cloud at the same time there are much better compute engines ramping next year.
NEXTPLATFORM AD
We will see as the June 2027, November 2027, and June 2028 Top500 supercomputer rankings come out. To balance things out, I would expect about two dozen extra machines to be installed to cover the drop experienced in the November 2026 rankings. If the June 2027 list is a little light, then the following two lists will have to be above average for things to balance out.
The thing to also remember whenever you look at the Top500 rankings is that the list is comprised of machines that have submitted HPL benchmark results formally for bragging rights. It is not intended to be an exhaustive list of all of the HPC systems that are installed in the world. Many organizations with fairly large or truly monstrous machines that are doing HPC work on them do not submit results; and many smaller organizations do submit results. The result is that we get a skewed view of the HPC systems market. Moreover, it is not a complete list even of those machines that have been tested running HPL and that are still in use. There are always only 500 unique machines on the list, even if there are probably somewhere on the order of 3,000 proper HPC sites worldwide, many with multiple machines – and I am not including AI training and inference systems in that number. I mean machines that do ModSim as their day job, which is only what should count and which is my major gripe about the middle and end of the Top500 list. There are a lot of clusters from clouds and telecom providers and software companies that most certainly do not do HPC as their day job.
All that said, the Top500 list is one of the few ways we get to see different architectural choices that HPC centers are making over time. And as such, I think this is still very valuable.
There are only 27 new machines that were added in the November 2026 Top500 list, which also means only 27 less powerful machines were knocked off the list. This is the lowest amount of churn across the 68 Top500 lists that have been compiled since 1993.
NEXTPLATFORM AD
Six lists ago, I decided to focus in on the new machines in each Top500 ranking and try to organize them by architecture to get a better sense of what HPC people are buying and installing as time goes by. I think this is a much better use of this list than just looking at the list as a whole, which has some truly old machinery on it at any given time. I think a rolling list with machines that are not older than four years or maybe five years – probably five given that HPC centers are stretching how long machines are in the field, just like the hyperscalers and cloud builders are – might be a better indicator of what is going on out there in HPC Land, and sometime next year I will start doing such a sort. (If I get some extra time, I can back cast it into the 2022 and 2023 Top500 lists and do it sooner.) The point is, the fact that a very large machine that was very expensive and is very old is still in use is interesting, of course, but it doesn’t tell you what new things people are choosing. I love that we use these machines for a long time, but when this machines hang around on the official Top500 list, that means another newer machine that might make the list can’t.
Here is how the number of new machines added to the past six Top500 rankings, which comprises 264 machines from June 2024 through November 2026. The table below shows the number of cores across those machines, the Rmax throughput on the HPL test, and the Rpeak peak theoretical performance inherent in the system. Take a gander:
You will note, no doubt, that the number of cores – which is actually CPU cores and GPU streaming multiprocessors (SMs) added up across the compute engines in the machines – across those 27 new machines is quite large, at 71.4 million cores. But it gets weirder.
There is a China Telecom cloud in Shanghai called the eSurfing Cloud that has an infrastructure as a service (IaaS) offering based on China’s 56-core Hygon C86 X86-based server CPU, which is itself derived from a Zen1 core licensed many years ago from AMD, as well as the C588 GPU from Chinese chip maker MetaX. That machine has 66.63 million cores and SMs across its compute complexes, which are built from nodes with two CPUs and eight GPUs. None of them is particularly powerful compared to what is available in the United States, however, and this cloud cluster only had 38.66 petaflops of peak performance and only rated 18.36 petaflops sustained on the HPL benchmark. This one “machine” on the cloud represented 93.3 percent of all of the cores across the 27 new machines on the November 2026 list, but only represented 2.9 percent of the Rpeak performance aggregated across those 27 machines.
NEXTPLATFORM AD
You can see the feeds and speeds of those new “HPC” machines in the table below:
The most powerful new machine on the list is the “Hangang” system at the KISTI supercomputing center in South Korea, which is based on Nvidia Grace-Hopper superchips and is made by Hewlett Packard Enterprise Cray EX254n server nodes and using HPE’s Slingshot 400 interconnect. The Hangang machine has a peak performance of 588.3 petaflops, and was one of five Grace-Hopper or Grace-Blackwell systems across the newbie systems, which as a group comprised 51.8 percent of all new capacity. (682.5 petaflops of aggregate compute.)
The second largest new machine in the November 2026 Top500 is the “Lux” supercomputer at Oak Ridge National Laboratory, which is made from HPE ProLiant XD685 nodes and uses a combination of “Turin” Epyc 9575F CPUs and “Antares” MI355X GPU accelerators. This machine is rated at 323.9 petaflops.
The rest of the machines are miniscule by comparison, and if you add them all up, Rpeak is only 1.32 exaflops aggregate across all of them.
One of the other things I like to do with the Top500 twice a year is look at how accelerated machines are being used as opposed to CPU-only machines. In some cases, the CPU is actually augmented with a huge amount of vector and sometimes matrix math, so it is not strictly speaking much like a general purpose CPU (think of the A64FX processor used in the “Fugaku” system at RIKEN Lab in Japan, for instance.)
This time around, there are 290 machines that have GPU or other kinds of accelerators to goose their floating point performance out of the 500 on the list:
The accelerated systems comprise 58 percent of machines, but 70.4 percent of total cores on the list and 79.9 percent of total peak exaflops. To be precise, the accelerated machines have an aggregate of 22,64 exaflops of FP64 oomph, compared to 5.7 exaflops of CPU-only peak theoretical performance. Nvidia has the lion’s share of the machines, but AMD GPUs are in a lot of very large machines and has 37.7 percent of FP64 flops compared to 52.9 percent for machines using Nvidia GPUs.
It will be interesting to see how AMD’s MI430X, which is aimed explicitly at the HPC crowd and which has very high real FP64 performance, does in the coming year and a half as shipments of this accelerator ramp and take on Nvidia GPUs with emulated, software-based FP64 computing atop matrix math engines instead of true FP64 performance.
來源:The Next Platform · nextplatform.com