Nvidia
Earlier this week at COMPUTEX, Nvidia announced that the new GH200 Grace Hopper “Superchip”, a CPU and GPU combination specially crafted for large-scale AI applications, has entered full production. it’s a beast It features 528 GPU tensor cores, supports up to 480 GB CPU RAM and 96 GB GPU RAM, and boasts up to 4 TB/s GPU memory bandwidth.
We previously covered the Nvidia H100 Hopper chip. This is currently Nvidia’s most powerful data center GPU. This is more than his A100 chip in 2020, which powers AI models such as his ChatGPT at OpenAI, and the first round of training runs for many of the news-generating AI chatbots and image generators being talked about today. also marked a significant upgrade.
Faster GPUs can almost translate into more powerful generative AI models, as they can perform more matrix multiplications in parallel (and faster) that today’s artificial neural networks need to function. increase.
The GH200 takes its “hopper” foundation and combines it with Nvidia’s “Grace” CPU platform (both named after computer pioneer Grace Hopper) and Nvidia’s NVLink chip-to-chip (C2C) interoperability. Integrate into one chip through connection technology. Nvidia expects this combination to dramatically speed up AI and machine learning applications for both training (model creation) and inference (model execution).
“Generative AI is rapidly transforming businesses, opening up new opportunities and accelerating discovery in healthcare, finance, business services and many other industries,” said Ian, Vice President of Nvidia Accelerated Computing. Buck said in a press release. “With the Grace Hopper superchip in full production, manufacturers around the world will soon be providing enterprises with the accelerated infrastructure they need to build and deploy generative AI applications that leverage their own proprietary data. Become.”
Key features of the GH200 include a new 900GB/s coherent (shared) memory interface, which is 7x faster than PCIe Gen5, according to the company. The GH200 also delivers 30x higher total system memory bandwidth to the GPU compared to his aforementioned Nvidia DGX A100. Additionally, the GH200 can run all his Nvidia software platforms, including Nvidia HPC SDK, Nvidia AI, Nvidia Omniverse.
Notably, Nvidia has also announced that it will be incorporating this CPU/GPU combo chip into its new supercomputer called the DGX GH200. This supercomputer harnesses the combined power of 256 GH200 chips that can be run as a single GPU, delivering 1 exaflops of performance and 144 terabytes. That’s almost 500 times more memory than his Nvidia DGX A100 of the previous generation.
DGX GH200 will be able to train huge next-generation AI models (such as GPT-6) for generative language applications, recommender systems, and data analytics. Nvidia hasn’t announced pricing for his GH200, but Anandtech says the price of one of his DGX GH200 computers “easily hits his low eight-figure price.”
Overall, thanks to continued hardware advancements by vendors like Nvidia and Cerebras, high-end cloud AI models will continue to get smarter over time, processing more data and performing more than ever before. It’s fair to say that you’ll be able to do things much faster. Let’s hope they don’t argue with tech journalists.
