Nvidia Corp. unveiled an improved AI processor that boosts the chip’s capacity and performance, aiming to solidify the company’s supremacy in a growing field. NVIDIA GH200 Grace Hopper will be available in the second quarter of 2024.
Nvidia said Tuesday at the Siggraph conference in Los Angeles that the Grace Hopper Superchip, a hybrid graphics chip, and CPU, will benefit from a new form of memory. The device is based on high-bandwidth memory 3, or HBM3e, which can access data at a blistering 5 gigabytes per second.
Nvidia’s latest announcement follows the previously announced GH200 with HBM3, which is presently in production and will be available later this year. This implies that NVIDIA will produce two versions of the same device, the first using HBM3 and the second incorporating HBM3e. Nvidia also says that the next-generation GH200 will be able to execute AI models 3.5 times quicker than the present model thanks to HBM3e memory.
Nvidia GH200 Grace Hopper: Built on a 72-core Grace CPU
The new GH200 Grace Hopper Superchip is based on a 72-core Grace CPU with 480 GB of ECC LPDDR5X memory and a GH100 computational GPU with 141 GB of HBM3E memory in six 24 GB stacks and a memory interface with a 6,144-bit width. NVIDIA physically installs 144 GB of memory, however, only 141 GB is accessible for increased yields.
NVIDIA’s current GH200 Grace Hopper Superchip features 96 GB of HBM3 memory with a bandwidth of less than 4 TB/s. In comparison, the new model increases memory capacity by around half and bandwidth by more than a quarter. Such major enhancements enable the new platform to run larger AI models than the previous iteration, resulting in considerable performance increases.
Microsoft and NVIDIA sign a deal to bring Game Pass to Nvidia GeForce Now
Nvidia claims this is due to differing configurations of the same component, similar to the 40GB and 80GB versions of the A100 from a few years ago. In truth, the RAM loadout is the main distinction here, at least for the time being. Instead of the 96GB of HBM3 vRAM and 480GB of LPDDR5x DRAM found in the current-generation GH200, the “next-generation” GH200 has 141GB of HBM3e and 500GB of slower 400GB/sec LPDDR5x. 512GB/sec LPDDR5x DRAM was used in the previous iteration.
Nvidia has been working on Grace Hopper Superchips for a while now
Since at least 2021, when the firm disclosed preliminary plans, the Grace Hopper Superchips have been a hot issue for Nvidia’s CEO.
The Superchip is built on the Arm architecture, which is extensively used in mobile devices and competes with Intel and AMD’s x86-based technology. It is referred to as a “superchip” by Nvidia since it combines the Arm-based Nvidia Grace CPU with the Hopper GPU architecture.
The Grace Hopper Superchip receives a boost from the world’s fastest memory, HBM3e, in the latest iteration of the Nvidia GH200 Grace Hopper. According to Nvidia, the HBM3e memory is up to 50% quicker than the HBM3 technology included in the GH200’s current iteration.
Nvidia Computex 2023 brings the future of gaming
“We’re very excited about this new GH200. It’ll feature 141 gigabytes of HBM3e memory. HBM3e not only increases the capacity and amount of memory attached to our GPUs, but also is much faster,” Ian Buck, VP and general manager of hyperscale and HPC at Nvidia, said during a meeting with press and analysts, according to Venture Beat.
Nvidia is not just developing faster hardware, but also scaling it in a new server design.
Buck stated that Nvidia is working on a new dual-GH200-based Nvidia MGX server system that would have two Grace Hopper Superchips of the next generation. He went on to say that the new GH200 will be linked using NVLink, Nvidia’s interconnect technology.
Both CPUs and GPUs in the new dual-GH200 server will be coupled with a completely coherent memory connection thanks to NVLink.
Nvidia GH200 Grace Hopper release date
According to NVIDIA, the GH200 Grace Hopper platform with HBM3 is now in production and will be commercially available next month. In contrast, the GH200 Grace Hopper platform with HBM3e is now sampling and is expected to be ready in the second quarter of 2024.
The new GH200 Grace Hopper, according to NVIDIA, employs the same Grace CPU and GH100 GPU technology as the prior iteration, thus, no further modifications or steppings are necessary.
Featured image credit: Nvidia
