The Occamy processor, which makes use of a chiplet structure, packs 432 RISC-V and AI accelerators and comes with 32GB of HBM2E reminiscence, has taped out. The chip is backed by the European Area Company and developed by engineers from ETH Zürich and the College of Bologna, stories HPC Wire.
The ESA-backed Occamy processor makes use of two chiplets with 216 32-bit RISC-V cores, an unknown variety of 64-bit FPUs for matrix calculations, and carries two 16GB HBM2E reminiscence packages from Micron. The cores are interconnected utilizing a silicon interposer, and the dual-tile CPU can ship 0.75 FP64 TFLOPS of efficiency and 6 FP8 TFLOPS of compute functionality.
Neither ESA nor its improvement companions have disclosed the Occamy CPUs’ energy consumption, however it’s stated that the chip will be passively cooled, which means it could be a low-power processor.
Every Occamy chiplet has 216 RISC-V cores and matrix FPUs, totaling round a billion transistors unfold over 73mm^2 of silicon. The tiles are made by GlobalFoundries utilizing its 14LPP fabrication course of.
The 73mm^2 chiplet is not a very massive die. For instance, Intel’s Alder Lake (with six high-performance cores) has a die measurement of 163 mm^2. So far as efficiency is anxious, Nvidia’s A30 GPU with 24GB of HBM2 reminiscence delivers 5.2 FP64/10.3 FP64 Tensor TFLOPS in addition to 330/660 (with sparsity) INT8 TOPS.
In the meantime, one of many benefits of chiplet designs is that ESA and its companions from ETH Zürich and the College of Bologna can add different chiplets to the bundle to speed up sure workloads if wanted.
The Occamy CPU is developed as part of the EuPilot program, and it’s one in every of many chips that the ESA is contemplating for spaceflight computing. Nevertheless, there aren’t any ensures that the method will certainly be used onboard spaceships.
The Occamy design goals to help high-performance and AI workloads by means of a bare-metal runtime, however it’s not but clear whether or not the runtime will likely be at a container degree or on the bare-metal degree. The Occamy processor will be emulated on FPGAs. The implementation has been examined on two AMD Xilinx Virtex UltraScale+ HBM FPGAs and the Virtex UltraScale+ VCU1525 FPGA.