NVIDIA Unveils New Grace CPU “Superchip” for AI and HPC
At this year's GTC Conference, NVIDIA released a variety of impressive new computing technology. One stand out was its Grace "Superchip" CPU (central processing unit).
When it comes to the computing field, it's natural that technology is expected to improve by leaps and bounds each year.
Fueled by economics, competition, and Moore's Law, this has been relatively true for decades. Despite fears of Moore's Law slowing down, 2022 has already seen impressive new technologies reach the market.
Hoping to create the next most innovative compute technology, NVIDIA made waves in the industry by releasing a plethora of new computing technologies that take its performance to the next level.
Amongst these technologies is its new Grace CPU "Superchip," a new AI chip meant for improved performance and efficiency.
The Grace CPU Superchip. Image used courtesy of NVIDIA
This article will take a closer look at the new Grace CPU, the technology that underlies it, and the performance it offers.
The Grace CPU Superchip
What could be seen as one of the most impressive announcements from this year's NVIDIA GTC Conference was the introduction of NVIDIA's new Grace CPU "Superchip."
Designed to run all of NVIDIA's computing software stacks and optimized for data center compute, the Grace CPU claims to be a workhorse of a processor.
The Superchip integrates 144 Arm cores into a single socket, basing the design on the latest Arm v9 data center architecture. Interestingly, this design consists of two individual, 72-core CPU chips connected together through NVIDIA's new NVLink-C2C technology.
Grace Superchip CPU architecture configurations, which NVIDIA claims are possible thanks to the NVLink-C2C. Screenshot used courtesy of NVIDIA [video]
To support this, Grace also states to feature a unique and efficient memory subsystem consisting of LPDDR5x memory with Error Correction Code (ECC).
According to NVIDIA, this memory subsystem offers a bandwidth of 1 TB/s, almost double the bandwidth of traditional DDR5 designs, and is said to consume a maximum of 500 W. All of this comes alongside a 396 MB on-chip cache.
Altogether, the system claims to achieve 2x the performance per watt of today's leading CPUs. Further, NVIDIA claims Grace to achieve 1.5x more performance on the SPECrate® 2017 benchmark than some of AMD's most recent 64-core EPYC.
NVIDIA eventually hopes to use the Grace CPU to support AI development and deployment and hyper-scale computing tasks in their servers.
At the heart of the Grace CPU is NVIDIA's newly released NVLink-C2C technology.
Following the trend of chiplet design, NVIDIA's NVLink-C2C technology is a new, ultra-fast chip-to-chip and die-to-die interconnect technology designed to facilitate communications between NVIDIA:
- Graphic processing units (GPUs)
- Data processing units (DPUs)
- Network interface cards (NICs)
- System on chips (SoCs)
The new technology is built on top of NVIDIA's existing SERDES and LINK design technologies, making NVLink-C2C capable of tasks like PCB-level integration, silicon interposer design, and wafer-level connections.
NVLink-C2C is meant for ultra high-bandwidth chip-to-chip communications. Image used courtesy of NVIDIA
Most importantly, thanks to innovative new packaging technology, NVIDIA claims that NVLink-C2C interconnects can deliver up to 25x more energy efficiency and 90x more area efficiency than PCIe Gen 5 on existing NVIDIA chips.
With all things considered, the technology could support up to 900 Gbps or more of interconnect bandwidth.
Excitingly for the industry, NVIDIA is making NVLink-C2C open for semi-custom silicon-level integration with NVIDIA technology. Additionally, NVLink-C2C will support Arm AMBA CHI protocol to increase design flexibility for end-users.
Keeping NVIDIA's Momentum Rolling
In the hardware world, it is expected that each year a company will improve its technology significantly. This year, NVIDIA has seemingly done that with its new Grace CPU, which significantly improves performance over existing offerings thanks to some exciting new technology.
With that in mind, NVIDIA's GTC Conference still has even more releases to delve into to see how else they are raising the bar for themselves and the industry.