NVIDIA GeForce RTX 4000 – Lots of information about possible specifications and performance of Ada graphics cards

Last week, the first concrete information about the upcoming NVIDIA GeForce RTX 4000 graphics cards appeared on the web, and this information appeared as a result of a hacker attack that occurred less than two weeks ago. Then we found out that at least five graphics chips were planned for desktop and laptop computers: AD102, AD103, AD104, AD106 and AD107. In some cases, NVIDIA plans to offer a narrower memory bus, but at the same time increase the amount of L2 cache several times. Even the smallest AD107 chip offers more L2 cache than the currently higher GA102 core. In the meantime, Moore’s Law is Dead has published a number of additional information about the possible specifications and performance of the NVIDIA GeForce RTX 4000 graphics cards. Of course, we’ve been working on rumors, but as the saying goes – there’s an iota of truth in every rumor.

On the network appeared a bunch of unofficial information about the specifications and performance of NVIDIA GeForce RTX 4000 graphics cards from the Ada family.

The NVIDIA GeForce RTX 4000 could receive a much larger L2 cache. We got acquainted with the circuit diagram AD102

Moore’s Law is Dead asserts that the initially groundbreaking NVIDIA AD102 chip, based on the Ada architecture, was built with a power consumption of 400-450W in mind, but the need to catch up with AMD RDNA 3’s leading drip cards means that actual power consumption can reach Ultimately to 600 Watts. Sources cited by MLID do not confirm, however, that the best GeForce RTX 4000 cards should have a consumption of up to 800 Watts. Unofficially, the premiere of the new geometry will take place before the end of the third quarter (i.e. towards the end of September), although it is more about the first paper presentation, not the actual availability of cards.

NVIDIA GeForce RTX 4000 – Ada graphics systems specifications and DLSS source code posted online

The MLID also provides some additional information about the specifications and performance of individual Ada graphics chips compared to the Ampere. It turns out that there is a high probability that at least the top-of-the-line AD102 chip will support not only GDDR6X memory, but also the new generation of bone – GDDR7 – that it is running on, among other things Samsung. It is supposed to offer an effective speed of 32 Gbps, but its introduction to GeForce RTX 4000 systems will not take place at the time of release, but may be updated (for example in the form of GeForce RTX 4000 SUPER cards). The AD102 chip, at best, will double the performance of Ray Tracing and deliver at least an 80% increase in classic raster performance – compared to the GeForce RTX 3090 with the GA102.

The AD103 chip will be equipped for cards such as the GeForce RTX 4070 and GeForce RTX 4060 Ti, which offer, among other things, a 256-bit memory bus and up to 10752 CUDA FP32 cores. Drip performance will be 10-30% higher compared to the GeForce RTX 3090. A significant performance increase is also expected in the RT. The core of the AD104 (including the GeForce RTX 4060) is to offer a 192-bit memory bus and up to 7,680 CUDA FP32 cores. Performance in the drip is up to a GeForce RTX 3080, while the performance in the RT offers a level similar to a GeForce RTX 3090 Ti with a full GA102.

Then there’s the NVIDIA AD106 chip, using a 128-bit memory bus and up to 4608 CUDA FP32 cores. This core could become the central part of the NVIDIA GeForce RTX 4050 Ti card, delivering performance that pips between the GeForce RTX 2080 SUPER and GeForce RTX 2080 Ti, while offering almost RT performance with the GeForce RTX 3070. The AD107 chip remains with the last 3072. Cores and a 128-bit bus. The card that will benefit from the AD107 will most likely be a GeForce RTX 4050, and its performance will be similar to the GeForce RTX 3060.

Source: Moore’s Law