NVIDIA has introduced the new GeForces based on the Ampere architecture

Unfortunately, the technical specs were handled very strictly by the company, so the exact capabilities of the incoming VGAs are still unknown.

True to its promise, NVIDIA has introduced the new GeForce with Ampere architecture. Specifically, a model called the GeForce RTX 3090, 3080 and 3070 will arrive in the next few weeks. The company shared the actual data fairly modestly, so many details cannot be described yet. In principle, the two largest models use the same GA102 GPU, while the smallest uses a more restricted version codenamed GA104. These have not been confirmed by the Greens, but we have received such information from their production partners so it is probably correct.

Each chip is manufactured using Samsung’s 8nm manufacturing technology, according to NVIDIA, in a node specifically optimized for them. The largest GPU has 28 billion transistors, but its scope is not yet public.

[+]

Major Ampere developments included beam tracking and AI processing, so you’re getting new RT and tensor processors, and RTX IO will also be introduced. The latter has nothing to do with RTX, the company has a habit of putting this flag up for a while, but the bottom line is that these are special components that can relieve the processor cores by unpacking compressed stored data. According to NVIDIA, with the data transfer performance of PCI Express 4.0, it would take up to 24 processor cores to decrypt content, especially from fast SSDs, but this is solved by RTX IO inside the GPU, the overhead can be around half a processor core. Here comes the Microsoft DirectStorage API, which we’ve mostly heard about on the Xbox Series X console, as part of the Velocity architecture. Now it is almost certain that it will hit a PC at some point, and then it will need support at the hardware level, that is what RTX IO provides.

[+]

For incoming VGAs, the following table details known specifications:

What may be strange is the increase in computing power. In this regard, we learned that Ampere architecture multiprocessors can perform twice as many 32-bit floating point operations per clock as Turing. However, we did not receive an answer on whether this ability could always be used or only in certain cases. The AMD and NVIDIA GPD designs in recent years have been specifically designed to be generic, so the arithmetic and logical units inherent in them have been made available in as many cases as possible. However, in the past, there were a number of tricks at this level with certain co-broadcast or dual-broadcast processors, when many processors were added to a given multiprocessor in terms of raw numbers, but some of them could rarely be use. You don’t even have to go that far, Intel’s graphics enhancements are still that way today. If we look at the Gen11 design of the Ice Lake IGP, it has 512 shading parts, but you can only use half of that in most of the program code. However, in terms of computing power, Intel considers it at its theoretical maximum, since the hardware can do it under certain codes. Examining the Ampere here would definitely require a more detailed description that would explain exactly how the floating point processors within the multiprocessors are connected or double the capacity of registers and caches to serve twice the number of components.

[+]

According to NVIDIA, the GeForce RTX 3070 is faster than the GeForce RTX 2080 Ti, while it will cost $ 500. The GeForce RTX 3080 will offer twice the GeForce RTX 2080 for $ 800, while the GeForce RTX 3090 has no comparison data. These average advancements were measured by NVIDIA at 4K resolution in many graphics-intensive games, but the specific settings are unknown, so it is not known whether DLSS or ray tracing were active in each program.

Among the products, the GeForce RTX 3080 and 3090 will arrive in September, while the 3070 will arrive in October.