Nvidia Ampere performance and features leak – RTX speed boost, NVCache and more

Nvidia Ampere performance and features leak - RTX speed boost, NVCache and more

Nvidia Ampere performance and features leak – RTX speed boost, NVCache and more

Before we start talking about Ampere, remember that rumours must always be taken with a huge grain of salt. Even the most reliable of sources can be given incorrect information, and vital details about future products could always be missing, or subject to change, until their official unveiling. 

Nvidia’s Ampere architecture is due to succeed the company’s Volta and Turing designs later this year, presumably as an architecture which spans both the consumer and enterprise sections of the GPU market.

According to the YouTube channel Moore’s Law is Dead, Nvidia’s Ampere series of graphics cards will be 8/7nm lithography and feature a significant architectural shift over today’s Turing-powered graphics cards. With these changes, a 10-20% increase in graphics IPC, performance per clock, is expected over Turing. 

With Nvidia’s continuing focus on AI, it makes sense that one of Ampere’s rumoured architectural upgrades lies within its Tensor cores. Nvidia has reportedly doubled the number of Tensor cores that are present within each SM of its Ampere series graphics cards, delivering a significant boost in AI performance over Turing. On top of this, Nvidia has also reportedly delivered a 4x increase in raytracing performance, though at this time it is unclear what is meant by this. 

If Nvidia offered a 4x boost in raytracing performance over its RTX 2060, then Nvidia’s next-generation RTX 3060 would be more powerful than Nvidia’s RTX 2080 Ti in terms of raytracing performance. That said, raytracing performance relies on more than simple rays per second counts. If Nvidia can cast four times as many rays as before, they need the computational performance to make these insane ray counts useful. With Ampere, the performance hit associated with raytracing is expected to decrease, though it remains to be seen how this will impact today’s games. 

Another feature that’s reportedly coming to Ampere is NVCache, a new technology that’s designed to allow Ampere graphics cards to better utilise data in system memory and storage to speed-up memory-constrained workloads. In effect, Nvidia has created an alternative to AMD’s HBCC (High Bandwidth Cache Controller), which allows AMD to utilise system memory of fast storage to overcome the limitations of frame buffer sizes. In effect, HBCC allows AMD to use system memory and storage as more VRAM, which is something that Nvidia hopes to replicate with Turing. 

Memory-wise, Nvidia appears to be focusing on improvements in memory compression to deliver increased effective memory bandwidth with Ampere. This allows Nvidia to increase its memory performance without increasing the VRAM capacities, and build costs, of its next-generation graphics cards significantly. A new technology called Tensor Accelerated VRAM Compression is also said to be in the works.      
   

If these rumours from More’s Law is Dead are true, Nvidia users can expect significant performance upgrades from Nvidia’s next-generation Ampere graphics cards as well as a large increase in performance per watt. That said, an increase in GPU IPC and boosts in overall performance metrics should be expected from all-new graphics architectures, especially when they are combined with a large shift in lithography tech. 

Rumour has it that Nvidia also plans to create a technology called DLSS 3.0, which is designed to be usable on any game that uses TAA. TAA can be implemented in several ways, so at this time it is unclear exactly how this technology will be implemented, or what level of graphic quality that gamers can expect from DLSS 3.0. 

Right now, nothing about Nvidia’s Ampere architecture has been confirmed. Nvidia plans to discuss Ampere at its next GPU Technology Conference (GTC) keynote, which is due to take place this Thursday at 2 PM UK time. 

You can join the discussion on Nvidia’s Ampere architecture leaks on the OC3D Forums.