NVIDIA Ampere @ 8nm: more CUDA Cores, more Ray Tracing, more IPC and goodbye to GTX

It seems that today will be quite an interesting day in the hardware world. While NVIDIA is scheduled to present the Ampere architecture at GTC 2020 digital next week, details have already been leaked. It will be a completely renovated architecture, far superior to Turing architecture. The latest leaked slides would indicate that the gaming versions of the Ampere graphics will be an eyesore.

NVIDIA within the GTC will not give data on gaming graphics based on Ampere, but we can get an idea. They will focus on their star sectors: AI, Deep Learning, Machine Learning and autonomous driving. Rumor has it even about x4 times more performance in Ray Tracing, which would be a crazy jump.

NVIDIA Ampere will be a monster

The first aspect that stands out about the filtered slides is that we will have more in less. Basically the RT Cores are increased at the same time as the DIE size is reduced. This has been achieved by jumping from 12nm lithography to 8nm lithography. Additionally, these GPUs have been developed from the ground up for high-performance computing and gaming.

It seems that it will not be an evolution, as happened with Turing with respect to Pascal, but a multipurpose solution. This involves a risk, as we saw with AMD GCN, a hybrid architecture between gaming and high-performance computing. The AMD GCN architecture was characterized as highly inefficient and unable to compete with NVIDIA's top of the line. We assume that NVIDIA will not comment on the same errors as its rival.

Eventually NVIDIA will use Samsung's 8nm lithography, as AMD has bought the excess production of TSMC's 7n node. A limited number of GPUs will be based on TSMC's 7nm node, possibly only the GA100 and GA102 chips.

But the biggest news in Ampere will be DLSS 3.0 (when two newscasts presented DLSS 2.0). DLSS technology will have more presence, it will even be complemented by TAA filters. We will see how these technologies will complement each other and what they will contribute to the world of video games.

New NVCache, NVEC codec and NVIDIA CUDA 11

One of the new features is something called NVCache, which seems to be an alternative to AMD HBCC. This solution allows to load or offer textures directly from the memory of the system, even being able to load in the SSD. Theoretically, it should improve VRAM load times.

The seventh generation of NVENC encoding will be added offering 8K @ 60Hz resolutions under H264 and H265 decoding. For some this is just science fiction, as not many have reached 4K resolution. NVIDIA will also have CUDA 11 ready, although it is not clear if it will implement it in Ampere directly.

More IPC, more L2 cache and more frequency

This new Ampere architecture from NVIDIA will offer a significant leap from Turing. Based on the data on the slides, we would see up to a 20% performance improvement, far from the rumors of up to 50%, which seemed insane. Additionally, the amount of L2 cache will double, which should positively impact performance.

According to the leaked slide, the base frequency of these graphs could be around 1900MHz and in Boost mode it could go above 2200MHz. The number of CUDA Cores and SM will be doubled, with x4 times more performance from each GPC. Come on, a hypothetical RTX 3060 could render as an RTX 2080 Ti. Finally, the RTX 3080 could have 10GB GDDR6 and the RTX 3080 Ti could have 12GB GDDR6.

Goodbye NVIDIA GTX, nice to meet you

The Ampere architecture is the end of the NVIDIA GTX graphics family that has been with us for so many years. All graphics cards will be renamed RTX, so Ray Tracing will be democratized and accessible to everyone. So it would not be unusual to see an RTX 3030, for example, with the same performance as the RTX 2060 for 200 euros.

Finally, NVIDIA seems to be revamping the GeForce Experience interface and adding important improvements. This tool has not undergone changes since its launch and has remained a short and lacking solution.

By the way, the gaming graphics will arrive in September

Source: MD