NVIDIA Grace CPU Superchip: 144 ARM cores to beat AMD EPYC
Photo Credit: NVIDIA
NVIDIA Grace CPU Superchip
At GTC 2022, NVIDIA CEO Jensen Huang unveiled the 144-core Grace CPU Superchip, the company's first CPU-only ARM chip designed for data centers.Grace CPU Superchip consists of two 72-core chips linked together via NVLink-C2C, providing a consistent 900GB / s connection that merges them into a single 144-core unit. In addition, the Superchip, based on ARM v9 Neoverse, supports Scalable Vector Extensions (SVE), SIMD instructions that increase performance and function similar to AVX.
Moving on to the benchmarks, NVIDIA claimed that the Grace Superchip CPU is 1.5 times faster in the SPECrate_2017_int_base benchmark than the two previous 64-core EPYC Rome 7742 processors it uses in its DGX A100 systems. The company based this claim on a pre-silicon simulation that features the Grace CPU with a score of 740+ (370 per chip). The current generation of EPYC Milan chips are capable of achieving scores ranging from 382 to 424 apiece, meaning the higher-end x86 chips will still be more powerful, but efficiency remains a big advantage of the NVIDIA solution. .
The two Grace CPUs communicate through the NVLink Chip-to-Chip (C2C) interface, which supports low-latency memory coherence, allowing connected devices to work on the same memory pool simultaneously. NVIDIA has claimed that NVLink-C2C can deliver up to 25 times more energy efficiency and is 90 times more regionally efficient than the PCIe 5.0 PHYs the company currently uses, supporting up to 900 GB / s of throughput or higher. Additionally, this interface supports industry standard protocols such as ARM's CXL and AMBA Coherent Hub Interface (CHI), as well as various types of connections.
Grace CPU Superchip is intended for tasks such as hyperscale computing, cloud, data analytics, HPC, and workloads TO THE. The product will support NVIDIA's entire CUDA stack and run the full range of proprietary applications, including NVIDIA RTX, NVIDIA AI, HPC and Omniverse. Shipments will begin in the first half of 2023.
Nvidia describes Arm-based Grace CPU ‘Superchip’
Did you miss a session at the Data Summit? Watch On-Demand Here.
Follow along with VentureBeat’s ongoing coverage from Nvidia’s GTC 2022 event. >>
Nvidia offered details on its Grace central processing unit (CPU) “Superchip” during CEO Jensen Huang’s keynote speech at its virtual Nvidia GTC 2022 event.
The Arm-based chip features 144 high-performance cores and a terabyte of secondary memory. Huang said the chip would double the performance and energy efficiency of Nvidia’s chips. It is on schedule to ship next year, he said, and it can be a “superchip,” or essentially two chips connected.
The chip is Nvidia’s own variant of the Arm Neoverse architecture, and it is a discrete datacenter CPU designed for AI infrastructure and high-performance computing, providing the highest performance and twice the memory bandwidth and energy-efficiency compared to today’s leading server chips, Huang said.
Huang explained that a supercomputer nearly one billion times more powerful than any that exist today is needed to predict climate change accurately. Nvidia and university researchers have developed a weather forecasting model dubbed Forecast Net — which is precisely the kind of task that CPUs like Grace will be needed for.
The Nvidia Grace CPU Superchip comprises two CPU chips connected, coherently, over NVLink-C2C, a new high-speed, low-latency, chip-to-chip interconnect.
The Grace CPU Superchip complements Nvidia’s first CPU-GPU integrated module, the Grace Hopper Superchip, announced last year, which is designed to serve giant-scale HPC and AI applications with an Nvidia Hopper architecture-based GPU. Both superchips share the same underlying CPU architecture, and the NVLink-C2C interconnect. Nvidia will be able to pair two Grace CPUs with up to eight Hopper chips.
“A new type of datacenter has emerged — AI factories that process and refine mountains of data to produce intelligence,” said Huang. “The Grace CPU Superchip offers the highest performance, memory bandwidth and Nvidia software platforms in one chip and will shine as the CPU of the world’s AI infrastructure.”
Created to provide the highest performance, Grace CPU Superchip gets an estimated performance of 740 on the SPECrate2017_int_base from its 144 cores.
This is more than 1.5 times higher compared to the dual-CPU shipping with the DGX A100 today, as estimated in Nvidia’s labs with the same class of compilers.
Grace CPU Superchip’s LPDDR5x memory subsystem offers double the bandwidth of traditional DDR5 designs at 1 terabyte per second while consuming dramatically less power with the entire CPU including the memory consuming just 500 watts.
The Grace CPU Superchip is based on the latest data center architecture, Armv9. Combining the highest single-threaded core performance with support for Arm’s new generation of vector extensions, the Grace CPU Superchip will bring immediate benefits to many applications.
Nvidia said the Grace CPU Superchip will excel at the most demanding HPC, AI, data analytics, scientific computingand hyperscale computing applications with its highest performance, memory bandwidth, energy efficiency and configurability.
Meanwhile, rival Advanced Micro Devices announced the general availability of the AMD Instinct MI200 family of accelerators and ROCm 5 software. Previously introduced in November, AMD Instinct MI200 series accelerators, which includes the MI210, MI250 and MI250X, are built on AMD CDNA 2 architecture.
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact.
