InspireSemi has announced the tapeout of the Thunderbird I accelerated computing chip with the goal of uniting the benefits of the RISC-V ISA and high-performance computing architectures. This chip packs thousands of CPU cores in a single device along with a high-speed mesh network, making it more versatile than GPUs.
The first offering using the Thunderbird I chip combines four chips in a single card to give designers access to over 6,000 CPU cores. Image (modified) used courtesy of InspireSemi
With modern applications such as AI/ML, advanced image and graphics processing, and data analytics consuming more computing resources, designers often turn toward GPUs to “parallelize” their computing for higher throughput. This, however, can introduce new restrictions in the development process—something InspireSemi hopes to address with its newest chip.
Versatile Parallel Computing
Compared to traditional CPUs, GPUs provide several performance advantages for modern applications. Instead of relying on a few cores, GPUs can pack thousands of equivalent cores in a single package. While each core may not have the speed or versatility of a CPU core, the sheer number of GPU cores enables faster parallel computation.
This is particularly advantageous when massive numbers of operations must be carried out. Instead of relying on the CPU to perform these computations a few at a time, the CPU can effectively delegate the computation among the GPU’s cores to achieve the same result in a fraction of the time. As such, GPUs see considerable use in HPC applications.
Compared to a single CPU core, GPUs pack several lower-performance cores in a single device, enabling faster parallel computation and improving data throughput. Image used courtesy of Nvidia
The downside, however, is that GPUs have a much higher development cost. Compared to CPU development, which can be relatively simple when using open-source instruction sets like RISC-V, GPU development requires complex architectures that restrict versatility. InspireSemi’s Thunderbird I chip can replace GPUs by instead relying on a dense network of RISC-V CPUs.
Replicating GPU Performance With CPUs
Within the Thunderbird I chip are 1,536 64-bit RISC-V CPU cores, along with supporting memory and networking architectures. The number of cores packed into the device coupled with the high-speed mesh enables cross-core communications and more efficient deterministic computing without severely increasing latency.
The Thunderbird chip provides new performance that was previously only achievable with GPUs, increasing versatility and changing the programming paradigm for HPC applications. Image used courtesy of InspireSemi
InspireSemi claims Thunderbird I can forego the rearchitecting characteristic of GPU acceleration. Instead, designers can rely solely on the RISC-V ISA for more versatility and easier integration into existing systems. As a result, the devices can help more developers use HPC resources without demanding a dedicated GPU programming effort.
InspireSemi’s initial product leveraging the Thunderbird I chip will be a PCIe card with four Thunderbird chips (>6,000 CPU cores) supporting FP64 operations up to 24 TFLOPS. The Thunderbird I devices are expected to be delivered in Q4 2024.
Shifting Processing Back to CPUs
While it is still too early to say whether the Thunderbird I chip can significantly contend with the GPU market, designers will find a use for a vast network of CPU cores in a single chip. InspireSemi has reported a “disruptive price point” for the Thunderbird devices, and this price point will likely play a critical role in how widely adopted the devices can become.
Regardless, the tapeout of the Thunderbird I chips is a promising milestone for InspireSemi, the RISC-V ISA, and those who would ultimately benefit from a CPU-focused HPC architecture.
