NVIDIA Ada Lovelace AD102 GPU: leaker says over 75 billion

NVIDIA’s next-gen Ada Lovelace GPU architecture and its big reveal at GTC 2022 next week is just so close, we can smell the silicon… which the flagship AD102 GPU silicon will have over 75 billion transistors.

The new AD102 GPU with 75 billion transistors means that Ada Lovelace is skipping up real close to Hopper H100, with NVIDIA cramming in 5 billion more transistors for a total of 80 billion transistors. We don’t know how the GPU architectures compare just yet, but both of them have huge L2 cache (Ada Lovelace has up to 96MB of cache, and Hopper with up to 50MB of cache).

All that cache takes up space, so the huge 75 billion transistor count on AD102 could be because of that, and why it has close to the same transistor count as NVIDIA’s monster Hopper H100 GPU. The news is coming from leaker “kopite7kimi” who posted a vote on Twitter asking followers how many transistors they thought NVIDIA’s flagship AD102 GPU had.

The vote ended with 39.3% voting for 70-80 billion transistors, 29.6% for 60-70 billion transistors, 18.9% for 50-60 billion transistors, and lastly 40-50 billion transistors had 12.2% of the vote. There were 639 votes in total, with one of them being from me where I personally voted for 70-80 billion transistors.

Leaker “kopite7klmi” just followed up on that initial tweet, with another saying “remember this” with a laughing emoji. The leaker added that “most people made the right choice” in that there are over 75 billion transistors, careful not to say over 80 billion transistors like the options in the vote (it didn’t go over 80 billion transistors).

Inside, NVIDIA’s next-gen AD102 GPU is expected to have 18342 CUDA cores at its max — but that’s the full-fat AD102 — we won’t see that at launch. NVIDIA will be using a slightly cut-down AD102 GPU for its GeForce RTX 4090, with that expected to have 16384 CUDA cores, and 24GB of GDDR6X memory.

NVIDIA’s next-gen AD102 GPU will be made on TSMC’s new 4nm process node, with NVIDIA tapping the 4nm process node — which was optimized and designed exclusively for NVIDIA’s monster Hopper H100 GPU — so we should expect the same optimizations and design exclusivity for the GeForce RTX side of things with Ada Lovelace and AD102.

Hopper H100 uses 80GB of HBM3 memory, so it’s very different from Ada Lovelace AD102 which uses GDDR6X memory (and the rest of the stack of AD10x GPUs that use GDDR6X and GDDR6 memory).

NVIDIA should be announcing its next-gen Ada Lovelace GPU architecture at its upcoming GTC 2022 event, where the company has announced its new “GeForce Beyond” special broadcast, where we’re expecting to see the new GeForce RTX 40 series graphics cards announced (at least one: the new GeForce RTX 4090).