Fujitsu Monaka: News of Japan’s 144-core superchip 28 comments
Image: Satoshi Matsuoka (X)
A new supercomputer chip is being developed in Japan. This one is called Monaka, has 144 cores distributed across chipsets and will be Fujitsu’s next Arm processor for the high-performance computing (HPC) segment after the A64FX. The latter serves in what was once the world’s fastest supercomputer called Fugaku.
New details on the Fujitsu Monaka
Professor Satoshi Matsuoka, who directs the RIKEN Center for Computational Science (R-CCS) in Kobe, Japan, where the Fugaku supercomputer operates, shared a photo of the new Fujitsu Monaka chip as a “mechanical sample” on that the chip consists of several dies that have been assembled as the chiplets are clearly visible in the attached schematic.
So, in the center is the large I/O chip manufactured in 5nm, which will provide the interfaces, which also include PCIe 6.0 or the Compute Express Link (CXL) version 3.0 based on it.
Fujitsu Monaka (Image: Fujitsu)
There are four SRAM chips (also 5nm) distributed all around, which contain the last level cache. The CPU cores themselves, which must be manufactured in an even thinner 2nm process, sit above. Each of these four dies contains 36 cores, as Monaka is expected to offer a total of 144 cores. That’s not quite the 150 cores we were expecting, but it’s a big step up from the Fujitsu A64FX’s 48 cores.
Fujitsu Monaka (Image: Fujitsu)
Monaka must above all be effective
Of course, the goal is to increase the computing power with the new processor. It was announced in advance that the performance of the application would be 1.7 times higher than that of competing products in the year of their release. Monaka is now said to be twice as fast as its competitors in 2027. How we know that now, however, is another matter. This is just a rough estimate.
Energy efficiency in terms of power per watt would also be twice that of the competition. Fujitsu speaks of “Ultra Low Voltage”, while the A64FX only says “Low Voltage”. Additionally, pure air cooling should be sufficient this time instead of a combination of air and water cooling. The issue of energy efficiency is at the top of the list for the Japanese.
HBM is no longer used
While the A64FX is still equipped with fast HBM2 memory, the Monaka successor will only have DDR5 RAM. Distributed across 12 channels like AMD and Intel’s HPC processors, this will still only mean about half the throughput, but will ultimately make the platform cheaper.
The predecessor A64FX
In June 2020, Japanese supercomputer Fugaku stormed the throne of the top 500 systems with more than 400 PetaFLOPS, a big lead at the time (second place didn’t even reach 150 PetaFLOPS). Fugaku remained at the top for two years and was only replaced in June 2022 by the American Frontier system with over 1,000 PetaFLOPS. In the latest world rankings from November 2024, Fugaku still ranks 6th, which is very impressive in 4.5 years since it was included in the list.
A64FX processor of the “Fugaku” supercomputer (Image: Fujitsu)
Fugaku did not achieve its then-unrivaled computing power with the help of GPU-based computer accelerators, as is the case with Frontier and most of the top 10. Instead, an armada of 48-core processors is used. Each of the 158,976 computing nodes contains a 48-core A64FX processor. The chip, developed by Fujitsu in cooperation with Arm, used the ARMv8.2-A SVE (Scalable Vector Extension) instruction set extension for the first time. The successor will have ARMv9 and SVE2.
Top 10 Fastest Supercomputers According to Top500.org Ranking (Previous) System Location Computing Power (Rmax) Processors Top500 Record 1 (-) El Capitan USA 1,742.00 PetaFLOPS AMD Epyc 4th Gen (24C, 1.8 GHz)
AMD Instinct MI300A November 2024 2 (1) Frontier USA 1,353.00 PetaFLOPS AMD Epyc 3rd Gen (64C, 2.0 GHz)
AMD Instinct MI250X June 2022 3 (2) Aurora USA 1,012.00 PetaFLOPS Intel Xeon Max 9470 (52C, 2.4 GHz)
Intel GPU Max 1550 (?) November 2023 4 (3) Eagle USA 561.20 PetaFLOPS Intel Xeon Platinum 8480C (48C, 2.0 GHz)
Nvidia H100 November 2023 5 (-) HPC6 Italian 477.90 PetaFLOPS AMD Epyc 3rd generation (64C, 2.0 GHz)
AMD Instinct MI250X November 2024 6 (4) Fugaku Japan 442.01 PetaFLOPS Fujitsu A64FX (48C, 2.2 GHz) June 2020 7 (6) Alps Switzerland 434.90 PetaFLOPS Nvidia Grace (72C, 3.1 GHz)
Nvidia GH200 June 2024 8 (5) LUMI Finland 379.70 PetaFLOPS AMD Epyc 3rd Gen (64C, 2.0 GHz)
AMD Instinct MI250X June 2022 9 (6) Leonardo Italian 241.20 PetaFLOPS Intel Xeon Platinum 8358 (32C, 2.6 GHz)
Nvidia A100 November 2022 10 (-) Tuolumne USA 208.10 PetaFLOPS AMD Epyc 4th generation (24C, 1.8 GHz)
AMD Instinct MI300A November 2024 Changes from last listing highlighted in bold
Marc deciphers processors by testing their performance for gaming, content creation, and artificial intelligence.
