Jensen Huang has just completed his GTC Spring 2022 keynote and so the embargo has been lifted on several exciting announcements from NVIDIA. NVIDIA has a lot of exciting hardware and software to discuss at this “AI Developer Conference #1,” from the Hopper H100 to next year’s Grace CPU superchips to the Jetson Orin.
NVIDIA announced the Hopper H100 as their most advanced GPU ever at 80 billion transistors. The H100 is for AI infrastructure and is built on TSMC’s 4N process. The H100 is rated for 4000 TFLOPS FP8, 2000 TFLOPS FP16, 1000 TFLOPS TF32, and 60 TFLOPS for FP64 performance. The HBM3 memory in the H100 can deliver 3TB/s bandwidth and supports PCI Express 5.0 and 4th Gen NVLink connectivity. The TDP of the H100 SXM module is 700 Watts.
NVIDIA is also introducing the DGX H100 consisting of eight H100 accelerators and the DGX SuperPOD with 32 DGX H100s for 1 Exaflops of AI performance.
NVIDIA Eos, meanwhile, is NVIDIA’s new supercomputer that will serve as a model for OEM and cloud partners. The NVIDIA Eos design is comprised of 576 DGX H100 systems for performance of 18 Exaflops at FP8, 9 EFLOPS at FP16, and 275 PFLOPS at FP64.
The 144-core Grace CPU superchip
In the first half of 2023, the Grace Hopper superchip as CPU and GPU is designed for large-scale AI and HPC workloads. Also coming in the first half of 2023 is the Grace CPU Superchip, which offers 144 cores, uses LPDDR5X system memory, and is designed to deliver 2x performance per watt compared to traditional servers.
The Grace CPU Superchip is based on Arm Neoverse processor cores and aims to deliver “twice as much memory bandwidth and power efficiency” compared to today’s leading server processors. The Grace CPU Superchip has two CPU chips connected via NVLink-C2C while the Grace Hopper Superchip is their first CPU-GPU integrated module.
NVIDIA Jetson Orin Development Kit
The Jetson Orin Development Kit for robotics for advanced AI purposes was also announced today. The Jetson AGX Orin aims to deliver 8 times the processing power of the Jetson AGX Xavier. The Jetson AGX Orin features an NVIDIA Ampere GPU, Arm Cortex A78AE processor cores, and the latest accelerator IP. The Jetson AGX Orin development kit is available at $1999 while production models will start in Q4 at $399.
On the software front, they showcased NVIDIA Triton as an open source hyperscale model inference solution. They also updated existing NVIDIA software packages including Riva 2.0, NeMo Megatron 0.9, Merlin 1.0 and updated NVIDIA Maxine as audio/video quality enhancement SDK.
Those are the main highlights, stay tuned for more GTC Spring news this week on Phoronix.
If you enjoyed this article, consider joining Phoronix Premium to view this site without ads, multi-page articles on one page, and other benefits. PayPal tips are also accepted free of charge. Thanks for your help.