Nvidia Announces Launch of New H200 Chipset
Nvidia (NVDA-US) made waves on Monday (13th) with the launch of its latest H200 chipset. This graphics processing unit (GPU) is set to revolutionize the AI industry, serving as a key component in training and building various AI models that are driving the generative AI boom.
Upgrade from H100
The H200 is an upgraded version of the H100, the chip that OpenAI currently utilizes to train its most advanced large language model (LLM), GPT-4. With demand for these AI GPUs on the rise, large enterprises, start-ups, and government agencies are all vying for a limited supply of chips. According to Raymond James, the cost of H100 chips ranges from $25,000 to $40,000, and a large number of these chips are required for the process of “training” a model.
Revenue and Stock Performance
The soaring demand for Nvidia’s AI GPUs has led to a substantial increase in the company’s stock price, which has risen over 230% in 2023. With an expected revenue of around US$16 billion in the fiscal third quarter, Nvidia anticipates a staggering 170% increase from the same period last year.
H200 Features
The H200 comes equipped with 141GB of next-generation “HBM3” memory, allowing the chip to excel in “inference,” a crucial process that utilizes large models to generate text, images, or predictions after training. Notably, Nvidia claims that the output speed of the H200 is almost double that of the H100, based on data from Meta (META-US) Llama 2 LLM post-tests.
Competition and Compatibility
Nvidia’s H200 is expected to be shipped in the second quarter of 2024, presenting competition to AMD’s MI300X GPU (AMD-US). The H200 will be compatible with the H100, ensuring a seamless transition for AI companies currently using the previous model for training.
New Developments
Furthermore, Nvidia announced plans to provide 4-GPU or 8-GPU server configurations on the company’s HGX complete system. This new chip, referred to as the GH200, pairs the H200 GPU with an Arm architecture processor. However, Nvidia has hinted that the H200 may not retain its crown as the fastest AI chip for long, with the promise of the B100 chips based on the upcoming Blackwell architecture in 2024.
Nvidia’s decision to move to one-year architecture upgrades, due to the strong demand for its GPUs, signals exciting developments on the horizon for the AI industry. With the H200 set to make its mark in 2024, the future of AI technology looks promising and dynamic.
Nvidia (NVDA-US) launched its latest H200 chipset on Monday (13th), a graphics processing unit (GPU) used to train and build various AI models that drive the generative AI boom.
This latest GPU is an upgraded version of the H100. The H100 is the chip that OpenAI uses to train its most advanced large language model (LLM), GPT-4. Large enterprises, start-ups and government agencies are all competing for a limited supply of chips.
Raymond James estimates that H100 chips cost between $25,000 and $40,000, and require thousands of chips working together to create the largest model in a process called “training.”
Demand for Huida’s AI GPUs has boosted the company’s stock price, which has risen more than 230% so far in 2023. Huida expects revenue in the fiscal third quarter to be around US$16 billion, a 170% increase from the same period last year.
A key improvement to the H200 is that it includes 141GB of next-generation “HBM3” memory, which will help the chip perform “inference,” which uses large models to generate text, images, or predictions after training .
Huida said the output speed of the H200 is almost twice that of the H100. This is based on using Meta
(META-US) Llama 2 LLM post-test data.
The H200 is expected to ship in the second quarter of 2024 and will compete with AMD’s MI300X GPU (AMD-US). The AMD chip, similar to the H200, has additional memory over its predecessor, which helps fit large models on the hardware to run a collection.
Huida said that H200 will be compatible with H100, meaning that AI companies that have already used previous models for training will not need to change their server systems or software to use the new version.
Huida said it will provide 4-GPU or 8-GPU server configurations on the company’s HGX complete system, a chip called the GH200. The chipset pairs the H200 GPU with an Arm architecture processor.
However, the H200 may not retain Huida’s crown as the fastest AI chip for long.
Although companies like Huida offer many different chip configurations, the latest semiconductors usually take a big step forward every two years when manufacturers move to a different architecture that is less than adding memory or other optimization can bring more significant performance improvements. H100 and H200 are based on Huida’s Hopper architecture.
Huida told investors in October that it would move from two years of architecture upgrades to one year due to strong demand for its GPUs. The company showed a slide suggesting that it would announce and release B100 chips based on the upcoming Blackwell architecture in 2024.
#Huida #launches #H200 #fastest #chip #date #training #models #Anue #JuhengUS #Stock #Radar
