US’s Nvidia has introduced the H200, a specialized graphics processor unit (GPU) for training artificial intelligence (AI) models driving generative AI advancements.
The new chip is an enhanced version of the H100, the GPU responsible for training OpenAI’s GPT-4, its top large language model (LLM).
A limited supply of the chips has interested major firms, startups, and government agencies seeking to secure them engaged in a fierce battle.
Analysts estimated that the H100 chips cost between $25,000 and $40,000. Thousands of these chips are required to develop leading models in a process known as “training.”
H200’s significant improvement includes 141GB of next-generation “HMB3” memory to enable the chip to conduct “inference,” using an LLM to create text, images, or predictions after training it.
Nvidia said the H200 can achieve almost double the output speed of H100. The study was based on a test with Meta Platform’s Llama 2.
The H200 will roll out in the second quarter of next year and is expected to rival Advanced Micro Devices’ (AMD) MI300X GPU. Like the H200, the MI300X boasts extra memory compared to its predecessors, enhancing its capacity to accommodate large models for inference.
Nvidia also assured seamless compatibility, stating that the H200 can be used with existing H100 systems and software, eliminating the need for changes.
Shares of Nvidia Rise After Unveiling H200
Nvidia’s shares climbed for two days after the California-based tech company introduced its newest H200 chip on November 13.
Nvidia’s stock closed Monday’s session 0.59% higher at $486.20, and it rose by 0.29% to $487.60 in after-hours trading.
With the company’s newest AI GPUs, its shares have gained over 230% this year. Nvidia’s stock has more than tripled in value year-to-date.
The tech giant expects to bring in about $16 billion in revenue for its fiscal third quarter, 170% higher than a year earlier.
Analysts also believed enterprise generative AI could significantly impact Nvidia’s profitability ahead of its earnings report on November 21.
