NVIDIA has provided more details on its next-gen GeForce RTX 40 series GPUs and their die sizes, transistor counts, and more for the upcoming AD102, AD103, and AD104 GPUs.
The company provided some more details on its flagship AD102 GPU, but some of the pieces of the AD103 and AD104 GPU puzzles weren't there. AnandTech's Ryan Smith reported the detailed specifications of the GPUs, where we now know that NVIDIA's next-gen AD102 GPU has a 608 mm² die, 76.3B transistors, AD103 has a 378.6 mm² die, 45.9B transistors, while AD104 has a 294.5 mm² die, 35.8B transistors.
We also know that the ROP count on Ada Lovelace is significantly higher, with AD102 rolling out with a huge 192 ROPs, AD103 with 112 ROPs, and AD104 with 80 ROPs. For comparison's sake, the current-gen GA102 "Ampere" GPU has 112 ROPs (so the same as AD103). Increased ROP count = higher rasterization performance in games.
The room has to be made on the architecture for all the new CUDA cores, ROPs, and more... with NVIDIA confirming it's removed NVLink from the Ada Lovelace GPU architecture. It's a huge disappointment, as I'm a huge fan of multi-GPU technology. It's not the best these days, but the glory days of SLI used to be so awesome.
- Read more: NVIDIA's new GeForce RTX 4090: 2x faster than RTX 3090 Ti for $1599
- Read more: NVIDIA's new GeForce RTX 4080 revealed: 16GB model starts at $1199
- Read more: NVIDIA GeForce RTX 4090 FE: 2850MHz GPU at stock in Cyberpunk 2077
- Read more: NVIDIA teases Ray Tracing: Overdrive, RTXDI, DLSS 3 for Cyberpunk 2077
- Read more: NVIDIA's next-gen AD102 GPU: 76.3B transistors, up to 18432 CUDA cores
- Read more: NVIDIA details its new GeForce RTX 4090 Founders Edition graphics card
NVLink wasn't the only thing to get axed from Ada Lovelace, with the room used for logical blocks, and more. L2 cache has been increased, with NVIDIA confirming AD102 has 96MB of L2 cache, AD103 has 64MB of L3 cache, and AD104 with 48MB of L2 cache. This means we'll see RTX 4080 16GB graphics cards with 64MB cache, while RTX 4080 12GB cards have 48MB of cache.
Ryan points out that 96MB of L2 cache on AD102 adds almost 5 billion transistors, and that's just for the L2 cache. The ROP count increase is 75% gen-over-gen, which is why we're seeing gigantic leaps in performance, but it's also being helped with the highest GPU clocks that NVIDIA has ever shipped in a graphics card, ever.