NVIDIA shipped more than 1 billion RISC-V cores in a year

As part of the RISC-V 2024 summit, NVIDIA shared interesting statistics – over the year, it shipped more than 1 billion RISC-V cores of its own design as part of GPUs and other accelerators, TechPowerUp reports. They are part of the GSP (GPU System Processor), which is responsible for managing the GPU and other service functions. A GSP has anywhere from 10 to 40 cores, depending on the complexity of the chip it drives.

The company switched to developing a new RISC-V-based GSP in 2016. Prior to this, for more than ten years, NVIDIA used a proprietary Falcon processor for GSP, the capabilities and flexibility of which became insufficient over time. Instead of purchasing and modifying off-the-shelf cores from third-party vendors (Arm, MIPS, Synopsys ARC) or trying to improve Falcon, the company decided to turn to the open RISC-V architecture and has so far developed three types of cores.

Image source: NVIDIA

Thus, NV-RISCV32 is the simplest 32-bit core without out-of-order execution, operating at a frequency of up to 1.8 GHz and having a performance of up to 1.8 CoreMark/MHz. NV-RVV is NV-RISCV32 with 1024-bit vector extensions. And NV-RISCV64 is a more serious 64-bit core with support for out-of-order execution and SMP, a frequency of 2 GHz and a performance of 5 CoreMark/MHz. In addition, NVIDIA has developed more than 20 ISA extensions for specific tasks.

RISC-V based GSP first appeared in the Turing generation in 2018. Now GSP is responsible for accelerator power and resource management, security, inter-chip communication, video codec management, video output, working with NVDLA, etc. Google has chosen a similar path – its TPU AI accelerators use SiFive cores. However, now there are already “large” AI accelerators built exclusively on RISC-V cores: Meta✴ MTIA of the first and second generations, InspireSemi Thunderbird, Tenstorrent Wormhole and Grayskull, as well as Esperanto ET-SoC-1 and ET-SoC- 2.