2020/2021 trends

“It’s a renaissance in semiconductors”

2021 trends:

• CPU - including server, mobile, and edge, where currently one can see big change moving from Intel processors/architectures to "self-made" ARMs and RISC-V. Starting with: Apple and its new M1 chip - specialized SoC with CPU, GPU, and NPU on-board; nVidia acquiring ARM; Amazon, Microsoft, Google - designing their own chip. Not to mention existing chips producers or academic initiatives like one that takes part in H2020 - Manticore, a general-purpose, ultra-efficient, RISC-V, chiplet-based architecture for data-parallel floating-point workloads. Intel new line of CPUs could be "politically correctly" called "not impressive".

• GPU - GPUs arena is dominated by nVidia however, last November AMD launched HIP - where you can compile HIP C++ code to CUDA or to AMD Radeon platform. There are quite a few interesting solutions especially in GPUs designed especially for ML purposes - like Intel "Ponte Vecchio" high-performance, highly flexible discrete general-purpose GPU architected for HPC modeling and simulation workloads and AI training; or extreme one - like Cerberas wafer = - 56x the size of the largest GPU. The Cerebras Wafer Scale Engine is 46,225 mm2 with 1.2 Trillion transistors and 400,000 AI-optimized cores.

• NPU / TPU / DPU / IPU / xPU - all ASIC specialized hardware, starting with Google TPU (benchmarks for currently unknown TPU .v4 was just released); Intels Gaudi - training, Goya - inference processors; through very specialized Marvell, Fungible DPU solutions. Competition here is hard and quite selective - in the end those are still ASICs - you cannot use them for other tasks than their definition, but they are trying to win in their fields.

• FPGA - re-programmable hardware. This space generally is occupied by Xilinx and Intel (80%) with Lattice in third place. This area (together with SoC - which sometimes is hard to distinguish) is definitely booming, and this is not only looking for technology trends and advantages that FPGA is giving but also growing development software, interfaces with high-level languages, or looking at the big business movements like the acquisition of Xilinx by AMD.

• SoC - CPUs, that contains not only GPUs, but also parts of FPGAs/xPUs - and I am not focusing again on Apple M1 chips, or big ARM ecosystem, there is exponentially growing specialized RISC-V solution: Celerity, Manticore, SiFive, including so specialized designs like Seagate or Western Digital (storage solution providers) used in real-time processing.

• Quantum - this is quite futuristic as of today - but it grows, and has some hype, so worth having some initial insight, and could be something to follow up.

Building custom chips is becoming practical due to the evolution of electronic design tools and the semiconductor manufacturing industry, which has lowered market entry.

Custom chips can complement existing designs or processors while adding product differentiation. Extract from Gartner recommendations:

• Assess the priority of adopting custom ICs by conducting a cost analysis on your current product vis-à-vis future products and the impacts of technology migration.

• Develop a flexible strategy about sourcing custom ICs by carefully navigating through the semiconductor design and manufacturing market, your in-house skills, R&D expertise.

• Develop product strategies to be competitive by identifying specific technical requirements that can be addressed with custom ICs not covered by existing suppliers.

The recent success of Apple M1 just proved that by investing in own designs/solutions you can get best possible performance, power, cost ... Generally - this is rennesaince of semiconductors, and everyone currently is trying to get some piece of that cake.

2020/2021 bullet points:

• AMD - big acquisitions over last years Mellonx, ARM,

• ARM at the same time introduced a new architecture ARM v9.

• nVidia - acquired Xilinx.

• RISC-V is flourishing - big ecosystem, thriving new projects...

• Xilinx ACAP - sci-fi comes to life - "Terminator" style chips - SoC system that can rebuild themselves - a fully software-programmable, heterogeneous compute platform that combines Scalar Engines, Adaptable Engines, and intelligent AI and DSP Engines

• Everyone starts doing their own SoCs - Apple, Amazon, Microsoft, Google, ...