Nvidia’s Blackwell GPUs: A Deep Dive into the Latest Leaks and Rumors In 2024

Nvidia’s Blackwell GPUs are the successors to the Ada Lovelace series. They are expected to hit the market in late 2024, idk or early 2025. The Blackwell family is rumored to include five different GPUs: GB202, GB203, GB205, GB206, and GB207.

This is my second article in regards to blackwell leaks. Things are getting interesting day by day.

Nvidia Blackwell Memory Specifications Leaks

The memory interface configurations of the Blackwell family are not expected to be much different from the Ada Lovelace series. However, rumors exist that the flagship GB202 GPU might feature a 512-bit memory interface. This could be a game-changer for the ultra-enthusiast segment as it not only offers higher memory bandwidth but it will also increase the memory capacity by a good amount. A 512-bit interface with 16 GB (2 GB) modules will lead to 32 GB capacities with 1792.00 GB/s of total bandwidth.

Market Segmentation

The Blackwell family of graphics processors is expected to address various market segments. The GB202 will likely address the highest-end of the market (e.g., GeForce RTX 5090, GeForce RTX 5090 Ti), the GB203 will address high-end and performance-mainstream segments (e.g., GeForce RTX 5080, GeForce RTX 5070 Ti, and GeForce RTX 5070), while the GB205 will address the mainstream part of the market (GeForce RTX 5060 Ti, GeForce RTX 5060).

Performance

Nvidia’s Blackwell GPUs are expected to substantially boost inference for the GPT-3 model, featuring 178 billion parameters. This improvement is potentially double or more than what the existing H200 architecture already provides. However, it’s important to clarify that while the graph explicitly highlights Large Language Model (LLM) performance, this doesn’t necessarily translate to an equivalent level of raw compute power unless there’s a substantial increase in cores and power efficiency.

Manufacturing Process

The second rumor is that Nvidia will utilize TSMC’s 3nm process node to manufacture the Blackwell GPUs. So far, all recent GPU launches by Nvidia have been a node shrink starting with the Pascal GPUs that were made on TSMC’s 16nm, Turing on TSMC’s 12nm, Ampere on Samsung 8nm, and the current Ada GPUs utilize the TSMC 4N (5nm Optimized) node.

Conclusion

As always, these are leaks and rumors, so they should be taken with a grain of salt. The actual specifications and performance of the Blackwell GPUs will only be confirmed once Nvidia officially announces them.

Source: @kopite7kimi