Fujitsu, AMD Lay Groundwork To Pair Monaka CPUs With Instinct GPUs

Fujitsu and AMD announced plans on Friday to develop a new, more energy-efficient AI and HPC compute platform that will pair the Japanese tech vendor's next-gen CPUs with the House of Zen's Instinct accelerators.

Fujitsu's Arm-based A64FX processors powered the number-ranked Fugaku supercomputer until AMD's larger Epyc and Instinct-based Frontier system unseated the beast in early 2022.

"By combining AMD's innovative GPU technology with Fujitsu's low-power high-performance processor Fujitsu-Monaka, we seek to create an environment in which more companies will be able to utilize AI while reducing the power consumed by datacenters," Vivek Mahajan, the Chief Technology Officer at Fujitsu said in a statement.

While we still don't know much about how the two companies plan to meld these parts, a slide deck released [PDF] alongside the announcement suggests Fujitsu is targeting CPU-based HPC and AI inference, and GPU-based training.

We do, however, know a fair bit about Fujitsu's Monaka processor [PDF]. Slated to be built on a combination of 2nm and 5nm process tech, the chip closely resembles AMD's own HPC-centric Epyc-X chips and will employ a chiplet architecture with distinct, compute, SRAM, and I/O dies stitched together on a silicon interposer. Similar to AMD's Genoa-X, the compute and SRAM tiles will be stacked and connected using through-silicon vias (TSVs).

Fujitsu's Monaka CPU looks a heck of a lot like Arm-take on AMD's cache-stacked X-chips. - Click to enlarge

Monaka will feature up to 144 ArmV9 cores with support for SVE2 vector extensions and confidential computing capabilities. The latter makes sense as Fujitsu has made a habit of deploying its chips not just in HPC clusters, but in the cloud as well.

Compared to the 48-core A64FX, Monaka will also support dual socket configurations, allowing for denser nodes.

In terms of memory and I/O, the part is expected to feature 12 DDR5 memory channels, PCIe 6.0, and CXL 3.0. This is another notable departure from the speedy HBM2 memory used on the A64FX.

Fujitsu is setting some rather lofty goals for the chip, predicting it will offer twice the application performance and performance per watt of competing CPUs when it begins shipping in 2027.

The fact that Monaka is still years off means we know even less about the Instinct accelerators planned for the collaboration. Depending on when the two begin rolling out systems in 2027, AMD's MI400-series platform due out in 2026 may be a candidate. But, beyond the fact that it'll use AMD's "next-gen architecture," we don't know anything about the part.

• xAI picked Ethernet over InfiniBand for its H100 Colossus training cluster
• ParTec expands supercomputer patent fight from Microsoft to Nvidia
• RISC-V reaches milestone with RVA23 profile ratification
• AMD pumps Epyc core count to 192, clocks up to 5 GHz with Turin debut

AMD told The Register they couldn't elaborate on the collaboration beyond what was announced. We also reached out to Fujitsu for comment regarding the timeline and integration plans, but had not heard back at the time of publication.

With that said, the idea of pairing GPUs with Arm-based chips is by no means new. For instance, prior to launching its Grace-Hopper Superchips, Nvidia worked with Ampere on an Arm-based GPU server. ®