During a recent presentation, the speaker highlighted the rapid advancement of computing, noting that computation has increased 1000 times in the last eight years, surpassing the previous progress of Moore's Law. The speaker introduced a new platform called Blackwell, which features a highly advanced GPU named Hopper. Blackwell is a platform that goes into two types of systems, one being form-fit and function-compatible with Hopper.
Blackwell has 208 billion transistors and incorporates a unique design that allows two dies to function as one chip, with no memory locality or cache issues. The speaker emphasized the efficiency of Blackwell, as it can seamlessly integrate into existing infrastructure and software. The chip offers significant computational power and boasts a new transformer engine with a fifth-generation MV link that enables fast communication and synchronization between GPUs.
Blackwell also introduces new formats, such as FP4 and FP6, which increase the amount of parameters that can be stored in memory and effectively double the throughput for inference tasks. The speaker emphasized the importance of these features for generative AI and content token generation.
To further expand GPU capabilities, the speaker introduced another chip called the MV link switch, which allows every GPU to communicate with each other at full speed simultaneously. This innovation enables the creation of an AI system called DGX, which delivers up to 720 petaflops of training performance and can fit into a single rack. The DGX system utilizes liquid cooling and achieves high bandwidth through the MV link spine.
The speaker demonstrated the impact of Blackwell and DGX by comparing the time and energy required to train a GPT model. Blackwell significantly reduces the number of GPUs and power consumption while maintaining the same training duration.
The speaker also highlighted advancements in the field of robotics, with an emphasis on end-to-end systems. The AGX system, specifically designed for autonomous robotics, demonstrates the commitment of the company to develop AI systems for physical applications. The speaker hinted at the possibility of a chatbot-like moment for robotics, suggesting that the integration of large language models into robotics may be on the horizon.
In conclusion, the speaker showcased the impressive advancements in GPU technology through Blackwell and DGX, emphasizing their impact on computation, efficiency, and scalability. The presentation also highlighted the company's commitment to developing AI systems for robotics, presenting the AGX system as a significant step forward in the field.
Glyph AI is a voice optimization platform designed to transcribe and repurpose voice data effectively. Whether it's for podcasts, meetings, lectures, or interviews, Glyph AI can generate structured notes suitable for each scenario, powered by extensive templates.
Q: How much has computation increased in the last eight years?
A: Over the course of the last eight years, we've increased computation by 1000 times.
Q: What is Blackwell?
A: Blackwell is the name of a platform, not a chip. It is the most advanced GPU in the world in production today.
Q: How many transistors does Hopper have?
A: Hopper has 208,000,000,000 transistors.
Q: What are the two types of systems that use Blackwell chips?
A: The two types of systems are form-fit, function compatible to Hopper, and a prototype board with two Blackwell chips and four Blackwell dies connected to a grace CPU.
Q: What is the purpose of the new transformer engine?
A: The new transformer engine allows for faster links and computation in the network, which amplifies performance and allows for better synchronization and updates in systems with multiple GPUs.
Q: How does Blackwell compare to Hopper in terms of performance?
A: Blackwell is two and a half times the FP eight performance for training per chip compared to Hopper. It also introduces a new format called FP six and doubles the throughput with FP four, which is important for inference.
Q: What is the purpose of the MV link switch chip?
A: The MV link switch chip allows every single GPU to communicate with every other GPU at full speed simultaneously.
Q: How powerful is the DGX AI system in terms of petaflops?
A: The DGX AI system is 720 petaflops, almost an exaflop for training.
Q: What is the bandwidth of the DGX MV link spine?
A: The DGX MV link spine has a bandwidth of 130 terabytes per second, which is greater than the aggregate bandwidth of the Internet.
Q: How does Blackwell compare to Hopper in terms of power consumption?
A: Blackwell only requires 4 power compared to Hopper's 15 power, making it much more energy-efficient.
Q: What is the next wave of AI?
A: The next wave of AI is focused on physical AI and robotics, where data from multiple sources are processed and compressed into large language models to enable autonomous systems.
.png)
.png)
.png)
.png)
.png)