NVIDIA's Flagship AD102 Ada Lovelace GPU Allegedly Packs A Whopping 75B Transistors
The next-gen flagship GPU from NVIDIA is rumored to pack in more than 75 billion transistors. A typically reliable tech Twitter source pitched a multiple-choice survey to followers earlier in the year and was today pleased to report that the public had been ‘right’, choosing 70-80B as the most likely transistor count for AD102.
For the uninitiated, the AD102 GPU from NVIDIA is the flagship Ada Lovelace chip which is expected to be used to create the GeForce RTX 4090 for consumers, and possibly other high-end models later down the line—e.g. RTX 4080 Ti and RTX 4090 Ti. There have been rumors that NVIDIA is going to go big with its consumer AD102, so this generation it will be quite a bit closer to the server/enterprise model (Hopper H100) than with the previous Ampere generation.
To put the rumored transistor count of the AD102 into some perspective, let us look at Ada Lovelace vs Ampere numbers. According to the latest reports, the chip going into the RTX 4090, the AD102 GPU, has 75 billion transistors in a 611mm2 die manufactured by TSMC on its N5 process. The GPU behind the existing RTX 3090, the GA102 GPU, has 28.3 billion transistors in a 628mm2 die manufactured by Samsung on its 8nm process.
If the Tweet shared by Kopite7kimi is correct, the new Ada Lovelace GPU has 2.65X more transistors than the Ampere predecessor but thanks to TSMC’s N5 process it has a smaller physical die. The extra transistors will be used for a significantly boosted mix of traditional 3D accelerator (CUDA) cores, as well as ray tracing (RT) and AI acceleration (Tensor) cores. Cache sizes are also expected to be larger than ever, and we can expect other enhancements to video acceleration and processing hardware.
NVIDIA has scheduled its Project Beyond special broadcast to kick off the GTC 2022 event a week from today. It is widely expected that the GeForce RTX 40 series will become official at this event, starting with the RTX 4090 (powered by AD102). Thus, we should get some solid tech specs shared by NVIDIA during the course of GTC, perhaps some details will be outlined by CEO Jensen Huang in his keynote immediately after the special broadcast (timed for 8am PT on Sept 20).
