NVIDIA will hold this year's GTC conference next week from March 16 to 19. It is expected to release at least two blockbuster products, one is an LPU biased towards reasoning, and the other is the next-generation GPU architecture Feynman. Feynman (physicist Feynman) still uses the previous system named after the famous scientist. In fact, its existence was mentioned on last year's roadmap, but there was too little information. It only mentioned the news of matching the next generation HBM memory.
According to the latest news, Feynman graphics card will launch with TSMC’s A16 process.This is the world’s first 1.6nm-level process and TSMC’s first process to use SRP backside power supply., this technology can not only improve density and performance, but also improve power supply capacity, mainly for HPC high-performance computing.
However, the A16 process OEM is expensive. It has been reported that NVIDIA plans to transfer some packaging orders to Intel and use the latter's EMIB-T packaging technology instead of relying entirely on TSMC's CoWoS packaging to reduce costs and increase production capacity.
Feynman graphics cards will push AI performance to new heights, but they also bring many problems. The first is power consumption. The current Blackwell architecture is close to 1000W, and the power consumption of the dual-income Blackwell Ultra even reaches 1400W.Feynman will also reach more than 1000W, and dual-core cards may have to reach 2000W.
The surge in power consumption will naturally lead to improved performance, but there will also be heat dissipation issues. Vera Rubin has made it clear that all liquid cooling will be used for heat dissipation, and Feynman should also turn to more efficient liquid cooling.
Another change is that Feynman will integrate Groq's LPU technology, but I personally think this is not necessarily the case. The main reason is that the Feynman architecture should have completed the design, and the entire LPU is too late, and NVIDIA will obviously build the LPU into an independent product line. After all, the requirements for AI chips in training and inference scenarios are different and need to be dealt with separately.
For gamers, Feynman graphics cards will be far away no matter how good they are, because Feynman is for 2028.Game cards will still be upgraded to the Rubin architecture in the second half of next year, and Feynman game cards will not be available until at least 2029.
