GeForce RTX 4000: Up to 144 SM for NVIDIA’s Ada Lovelace GPUs?

In the past few hours NVIDIA confirmed the theft of some information proprietary and hour by hour information on the company’s future plans is sprouting on the net based on a first tranche of data released by the hackers of Lapsus $.

After the codename Blackwell, probably an architecture following the Hopper project for the datacenter world that will be announced this year, the alleged “Ada (Lovalace)” chip configurations which should constitute the offer GeForce RTX 4000 coming this year.

In the documentation we talk about five called GPUs AD102, AD103, AD104, AD106 e AD107, “acronyms”, however, already leaked in the past few hours. Based on the table you can see above, the top of the range GPU AD102 should have a maximum of 144 Streaming multiprocessorwhich means – as long as NVIDIA maintains the same number of CUDA cores per SM, equal to 128 – a maximum of 18432 CUDA core. The number in question was already circulating in December 2020.

Recall that the chip GA102current leader of the Ampere range and in the predicate of peeping out on the RTX 3090 Ti (announced and not yet released …) makes available 10752 CUDA core. Ada Lovelace could then offer 71% more CUDA core than Ampere.

GeForce RTX 4000 well over 450W for the top of the range? We are talking about 800-850W

As for the other GPUs, AD103 would have a maximum of 84 MS (10752 CUDA core), AD104 would stop at 60 SM (7680 CUDA core)while AD106 e AD107 would integrate 36 and 24 SM, ie 4608 and 3072 CUDA core respectively. NVIDIA’s GeForce RTX 4000 lineup GPUs are expected to debut at September this year, produced by TSMC at 5 nanometers.