Fujitsu and AMD introduced plans on Friday to develop a brand new, extra energy-efficient AI and HPC compute platform that may pair the Japanese tech vendor’s next-gen CPUs with the Home of Zen’s Intuition accelerators.
Fujitsu’s Arm-based A64FX processors powered the number-ranked Fugaku supercomputer till AMD’s bigger Epyc and Intuition-based Frontier system unseated the beast in early 2022.
“By combining AMD’s progressive GPU expertise with Fujitsu’s low-power high-performance processor Fujitsu-Monaka, we search to create an surroundings wherein extra corporations will be capable of make the most of AI whereas decreasing the facility consumed by datacenters,” Vivek Mahajan, the Chief Know-how Officer at Fujitsu stated in a press release.
Whereas we nonetheless do not know a lot about how the 2 corporations plan to meld these components, a slide deck launched [PDF] alongside the announcement suggests Fujitsu is focusing on CPU-based HPC and AI inference, and GPU-based coaching.
We do, nevertheless, know a good bit about Fujitsu’s Monaka processor [PDF]. Slated to be constructed on a mix of 2nm and 5nm course of tech, the chip intently resembles AMD’s personal HPC-centric Epyc-X chips and can make use of a chiplet structure with distinct, compute, SRAM, and I/O dies stitched collectively on a silicon interposer. Just like AMD’s Genoa-X, the compute and SRAM tiles will likely be stacked and related utilizing through-silicon vias (TSVs).
Fujitsu’s Monaka CPU seems to be a heck of loads like Arm-take on AMD’s cache-stacked X-chips. – Click on to enlarge
Monaka will characteristic as much as 144 ArmV9 cores with assist for SVE2 vector extensions and confidential computing capabilities. The latter is sensible as Fujitsu has made a behavior of deploying its chips not simply in HPC clusters, however within the cloud as nicely.
In comparison with the 48-core A64FX, Monaka may even assist twin socket configurations, permitting for denser nodes.
By way of reminiscence and I/O, the half is anticipated to characteristic 12 DDR5 reminiscence channels, PCIe 6.0, and CXL 3.0. That is one other notable departure from the speedy HBM2 reminiscence used on the A64FX.
Fujitsu is setting some fairly lofty objectives for the chip, predicting it can provide twice the appliance efficiency and efficiency per watt of competing CPUs when it begins transport in 2027.
The truth that Monaka continues to be years off means we all know even much less in regards to the Intuition accelerators deliberate for the collaboration. Relying on when the 2 start rolling out methods in 2027, AMD’s MI400-series platform due out in 2026 could also be a candidate. However, past the truth that it’s going to use AMD’s “next-gen structure,” we do not know something in regards to the half.
AMD informed The Register they could not elaborate on the collaboration past what was introduced. We additionally reached out to Fujitsu for remark relating to the timeline and integration plans, however had not heard again on the time of publication.
With that stated, the thought of pairing GPUs with Arm-based chips is certainly not new. As an illustration, previous to launching its Grace-Hopper Superchips, Nvidia labored with Ampere on an Arm-based GPU server. ®
