NVIDIA: most of the GeForce RTX 4000 specifications revealed

The next generation of graphics cards
signed NVIDIA could arrive as early as next fall.

The authors of the intrusion in the infrastructure of NVIDIA recovered various information as for the current and future technologies of the American company. Today, we are talking about the next generation of GPUs from the GeForce promoter.

Five GPU benchmarks leaked

While graphics cards based on the Ampere architecture were launched almost two years ago, everything suggests that NVIDIA will present the next generation in September. For several months, we have been talking about this architecture and its code name, Ada Lovelace.

Today, several of our colleagues evoke the diffusion of precisions as for this architecture. Details obviously linked to recent data theft at NVIDIA and which, therefore, are to be taken with hindsight. These are just rumours.

Unsurprisingly, however, it seems very clear that the GeForce RTX 4000s based on Ada Lovelace will be formidably equipped. For the time being, the leaks mention five Lovelace GPU references from the AD102 to the AD107, passing, of course, by the AD103, AD104 and AD106.

A generous L2 cache

First lesson, even if things will of course be different depending on the GPU chosen, the second level cache will progress considerably with Ada Lovelace. We are actually talking about 96 MB of L2 on the AD102 when Ampere was content with a maximum of 6 MB: a way to respond to AMD’s Infinity Cache?

This same AD102 promises to be a monster of power for what will probably equip the RTX 4090. It is about 18,432 Cuda cores, 144 stream processors (SM) and a 384-bit memory bus. The AD103 shouldn’t be bad either with still 10,752 Cuda cores, 84 SMs, 64 MB of second level cache and a 256-bit memory bus.

More modest, the AD104 still retains 7,680 Cuda cores, 60 SMs, 48 ​​MB of L2 cache and a 192-bit memory bus. While the leaks don’t mention the AD105, the AD106 has 4,608 Cuda cores, 36 SMs, 32 MB of L2 cache and a 128-bit memory bus. Finally, the AD107 brings up the rear with 3,072 Cuda cores, 24 SMs, still 32 MB of L2 cache and a memory bus that remains at 128-bit.

To make things more meaningful, our colleagues from WCCFTech have drawn up a small comparison table to see the progress of the Lovelace chips on their current generation equivalents. No doubt, the progress is impressive on paper. It remains to be seen consumption, prices and, of course, availability.

Source: WCCFTech