Necessity is always the mother of invention, and if you are trying to create an indigenous compute engine design and manufacturing capability, as China certainly is, then it stands to reason that the Middle Kingdom’s eight national computing centers would not only build homegrown accelerators for their supercomputers, but also create all-CPU machines that do not have an offload model and that simply use MPI to spread calculations across traditional scale out networks.

In this way, China can make a big machine that is relatively simple whose practical limits are space and power. To design weapons as well as to advance the state of the art in all of the sciences, China has the money and the power to go big instead of going dense with accelerated computer designs.

And that is precisely what China did in creating the “OceanLight” supercomputing for NSC Wuxi a few years back, which is based on the homegrown Sunway SW26010-Pro CPU and which had 41.93 million cores to deliver around 1.5 exaflops of peak theoretical performance and what it has done again to create the “LineShine” supercomputer that is now the fastest supercomputer in the world and that is installed at NSC Shenzhen.

I will be drilling down into the architecture of the LineShine machine – presumably this is a literal translation of what we would call “sunbeam” in English – in a separate story, but generally speaking LineShine is based on an Armv9-compatible server CPU designed by NSC Shenzhen in conjunction with Chinese IT giant Huawei (presumably its HiSilicon chip division). The LingKun LX2 CPU design has 304 active cores, and very likely there are more cores on the chip to increase the yield. The LineShine machine has a proprietary LingQi LQLink interconnect, which I am reasonably sure is based on a variation of InfiniBand technology but it could be a jacked-up and stripped-down version of Ethernet.

The bottom line – or in this case, the top line given LineShine’s top ranking among supercomputers that have submitted High Performance Linpack test results officially – is that this LX2 CPU delivers enough FP64 oomph with its SVE2 vector units that it only takes 13.79 million cores to deliver a peak theoretical performance of 2.74 exaflops (rounding to three significant digits). On the HPL test, LineShine delivers just a tad under 2.2 exaflops of oomph and that makes it 21.5 percent more powerful than the former top ranked machine, the “El Capitan” supercomputer based on AMD MI300A compute engines located at Lawrence Livermore National Laboratory in the United States.

China is back on top in supercomputers, and we strongly suspect that the country has several more exascale machines that it is not talking about much. We know of two, the aforementioned 1.5 exaflops OceanLight system at NSC Wuxi and the 2.05 exaflops Tianhe-3A supercomputer at NSC Guangzhou. But as Nicole and I have been reminded everyone for years, China has been ahead in the exascale race even if it did not submit official Top500 results.

I am not going to go through all of the top ten machines on the list. You can go read that for yourself, and everyone does that anyway. I want to add some value in my commentary, and I will stick to the methodology I started back with the June 2024 ranking of only looking at the new machines added during the current list and ignoring the big swatch of machines at cloud providers and telcos (mainly in China) that skew the list away from its HPC purpose. These machines are not doing real HPC work, and everyone knows it.

That said, it bears reminding that we have still fallen off the Moore’s Law wagon of doubling performance every two years, and at least in supercomputing, we are not spending enough money to get passage on the wagon. Take a look:

This is a budgetary thing much more than it is a technology thing.

It is also fun to look at which vendors have what shares of the Top500 rankings, so here is a pretty treemap that shows who has what share by aggregated capacity, which is what matters because flops, like tokens, are money.

The five official exafloppers dominate the capacity landscape, and the other five machines that have more than 400 petaflops of HPL performance also crowd out a lot of littler machines. You have to have 2.66 petaflops of HPL bang to even get on the list this time around. Which frankly is not that much given how much performance you can get out of a modern CPU or GPU.

With that, let’s take a look at the new machines on the June 2026 Top500 list. Here they are, all sorted by architecture and size within each architecture:

There are 44 new machines this time around, and one thing that is immediately obvious is that excepting the dominance of the LineShine machine, which comprises 51.6 percent of the aggregate new 5.3 exaflops capacity added to the June list, is that some HPC centers are hanging back and preferring to install “Hopper” H100 and H200 GPU in the machines that employ accelerators. This is for obvious reasons. First, Hopper GPUs are cheaper, and they also have more FP64 flops and more flops per dollar than the follow-on “Blackwell” B200 and B300 GPUs. The most powerful new machines that solely use Nvidia compute engines are based on Hopper, but there are three clusters that use Blackwell.

The other thing you will note is that there are a lot of clusters that have Intel Xeon processors married to Nvidia GPUs. This stands to reason since there are CPU preferences and prejudices in HPC as much as in the world at large. There is also the matter of price and availability in a world gone mad with GenAI. There were eleven such new machines on the June list, and there were another nine machines that had AMD Epyc CPUs paired with Nvidia GPUs. Together, these hybrid architectures comprised 15.3 percent of the installed flops capacity.

The other big new machine is the HPC7 system at Italian oil and gas giant Eni, which is based on AMD’s hybrid CPU-GPU MI300A accelerator; HPC7 is essentially a chip off the El Capitan block, and is ranked number six on the list. It is the largest commercial supercomputer with submitted results. (Don’t confuse that with the largest commercial supercomputer. We don’t know how many larger machines might be at the oil majors around the world. They don’t brag much.) These two MI300A systems comprise 16.3 percent of the new flops capacity. There are also two machines that mix discrete AMD CPUs and GPUs, as you see, and they add another 1.7 percent of the capacity.

That leaves the CPU-only crowd. There are five new HPC clusters that have AMD Epyc processors as their only compute engines, accounting for eight-tenths of a point of the aggregate new flops, and four new Intel Xeon clusters that add another 1.8 percent of the capacity. CPU-only machines are not taking over the world, but they are not going away.

Here is an interesting little table comparing the core counts, Rmax on HPL, and Rpeak performance of the new machines added over the past five lists:

As you can see, upgrades in the HPC sector come in waves, and they follow product cycles. June 2024 and November 2025 were relatively weak when it comes to new capacity installed, and June 2026 and November 2024 were particularly strong. And, I might add, dominated by the installation of exascale-class machines. Again: This is not a statement about HPC supercomputing but rather HPC systems installed that submitted Top500 HPL benchmark results. But the broader and sometimes secretive HPC market will probably reflect the official list to some degree, which is why we bother with Top500 coverage in the first place.

That brings us finally to the accelerated computing table. By the listing in the Top500 site, there are 274 machines that have some sort of acceleration, although the text on the site says there are 277 machine. I checked my math three times, and that is enough. Here is the architectural sort on those 274 machines:

Nvidia dominates by machine count, with 237 systems compared to AMD’s 32 systems. But if you look at it by peak flops, AMD has 8.18 exaflops of installed capacity compared to Nvdia’s 11 exaflops, and in terms of concurrency, AMD has sold accelerated machines with a combine CPU and GPU concurrency of 35.3 million cores compared to Nvidia’s 38.9 million cores. This is a real race, and perhaps portends the future of AI computing as well.