The race for supercomputing Not only can countries boast of being included in the World Top 500 list published every year, but they constitute very useful equipment for the advancement of technology in all fields, be it computer science, medicine, meteorology And much more. . And now there is good news from Germany, because they have reached a deal to build two new hierarchical superiorsfirst Hunter which will arrive in 2025 and after Shepherdwhich will be ready in 2027.
This is all the result of a partnership between AMD, HPC and the University of Stuttgart with the aim of improving the development of HPC and AI in the region. The agreement currently concerns two supercomputers, Hunter, whose design has already started and should be operational in 2025, and Herder, a system exascale whose deployment is planned for 2027.
Discover the two new German supercomputers
Hunter and Herder’s goal is to improve High Performance Computing (HPC) and artificial intelligence capabilities, as well as high-performance data analytics (HDPA). The cost of these two supercomputers amounts to approximately 115 million dollars (estimates), and will be funded by several government agencies as part of a new effort by Germany to become a technology superpower (remember, it also subsidizes Intel factories in the country).
Starting with the first of these two supercomputers, named Hunter, it will have an HPE Cray EX4000 rack and inside it will accommodate 136 interconnected nodes with four high-performance HPE Slingshot interconnects. The system will use accelerators AMD Instinct MI300A, which combine both the CPU, GPU and HBM memory on a single chip. In total, it is estimated that this should provide a computing capacity of approximately 39 petaFLOPSwhich represents a generous increase in performance compared to Hawk, the current supercomputer at the University of Stuttgart, which provides around 26 PetaFLOPS.
As for Herder, for now there is no detailed information on the hardware he will mount, but since it should be ready by 2027, it is assumed that they will wait until they have the latest hardware available until then. Recall that in this case, Herder is an exascale supercomputer, and this means that it must be capable of performing a minimum of 1ExaFLOP or floating point operations per second. Typically, this type of supercomputers does not consist of a single rack, but rather multiple interconnected racks, as you can see in the image above. Big words.
In any case, we must also mention the Jupiter supercomputer, currently under development and which will be the first exascale computer in the German country. Of course, Germany’s investment in technological development is monumental. It will therefore be interesting to see whether these capabilities will guide the country’s (and Europe’s) overall development process, particularly in the field of AI.
AMD takes center stage
As we mentioned previously, the Germans chose AMD and its Instinct MI300A accelerators to assemble the first of these supercomputers, Hunter, highlighting the signature of Lisa Su.
Exascale #Supercomputing comes to Stuttgart: Hunter will be based on the AMD Instinct MI300A accelerated processing unit (APU). This will reduce the energy required to run Hunter by approximately 80% at peak performance and increase peak performance to 39 petaFLOPS. https://t.co/o6mv4xYdh5
December 21, 2023 • 9:32 p.m.
These APUs (because they are actually APUs with CPU, GPU and HBM memory) have 8 cores and 16 Zen 4 architecture process threads for each CCD, and since they have 3, this means that each of them provides 24 cores and 48 process threads. It has 1 MB of L2 cache per core (i.e. 24 MB) and a separate cache pool (32 MB per CCD).
These accelerators integrate 153 billion transistors, feature Zen 4 architecture for the CPU and CDNA 3 for the GPU, support up to 192 GB of HBM3 memory and have up to 8 chiplets + 8 memory stacks.
