NVIDIA Unveils Next-Generation AI Chip 'Blackwell'

Inference Performance Over 30 Times
Cost and Energy Reduced to 1/25

NVIDIA unveiled its next-generation artificial intelligence (AI) chip architecture called 'Blackwell' along with the 'B200' chip based on it. The B200 offers over 30 times the inference performance compared to the previous generation H100, while reducing cost and energy consumption to one twenty-fifth. Inference refers to an AI model's ability to generate responses, and higher performance in this area implies more human-like thinking capabilities.

Jensen Huang, NVIDIA's Chief Executive Officer (CEO), announced the next-generation chip architecture 'Blackwell' at the 'AI Developer Conference (GTC 2024)' held on the 18th (local time) at the SAP Center in San Jose, California, USA.

The architecture is named after David Harold Blackwell, a mathematician who majored in game theory and statistics and was the first Black member inducted into the U.S. National Academy of Sciences.

NVIDIA's next-generation chip architecture 'Blackwell'.
[Photo by NVIDIA]

NVIDIA explained that Blackwell provides up to 30 times the performance improvement over the H100, which is based on the NVIDIA Hopper architecture and is currently regarded as the latest AI chip, while costing and consuming up to 25 times less energy.

The Blackwell architecture GPU is equipped with 208 billion transistors. Considering that the existing H100 consists of 80 billion transistors, this is a 2.5-fold increase.

The chip is reported to be manufactured using 4nm (1 nanometer = one billionth of a meter) technology by TSMC, the world's largest foundry (semiconductor contract manufacturer).

Blackwell also incorporates 5th generation NVLink, a communication pathway that transfers 1.8TB of data per second bidirectionally among up to 576 GPUs; an AI-based RAS engine that verifies GPU and memory reliability in long-running data center environments; and Secure AI, which encrypts sensitive data and ensures reliability.

CEO Jensen Huang said, "The Blackwell GPU is the engine driving a new industrial revolution," adding, "We will realize the potential of AI across all industries by collaborating with the world's most dynamic companies."

Many companies, including Amazon, Dell Technologies, Google, Meta, Microsoft, OpenAI, Oracle, and Tesla, are planning to adopt this chip.

The price was not disclosed. However, the market expects the price per unit to be at least $50,000, which is more than $10,000 higher than the existing H100.