Leaks have emerged detailing next-generation technology that should change the way we see graphics processing units (GPUs) from Nvidia and AMD.
This innovation is seen to improve raw power and power consumption features, leaker @greymon55 posted in a tweet cited by Extreme Tech. The tweet underscored that upcoming flagship GPUs from Nvidia and AMD will have more than twice the power of current top offerings.
The leaker indicated that the forthcoming Ada Lovelace AD102 GPU will be built on a five-nanometer process from TSMC, which is seen as a huge leap from its present Ampere GPU line that was manufactured using Samsung's 8-nm process. This 5-nm process is behind the exceptional Apple systems-on-a-chip (SoCs) on the iPhone 13 and the newly unveiled M1 Pro and Max silicon.
Next-gen Nvidia GPU: Twice the Power of Current Flagship RTX 3090
Nvidia's Ada Lovelace AD102 will pack a stupendous 144 streaming multiprocessors, which is nearly twice the present 82 multiprocessors in the RTX 3090. The upcoming GPU may also carry 18,432 CUDA scores, comprising a 54-percent increase.
The new GPU is seen to keep its 384-bit memory bus width and still use its GDDR6X memory. Core clock speeds are also rumored to be around the low 2GHz-plus range.
This latest leak indicates that Nvidia is set to increase clock speed to 2.3 GHz that would result in an astounding FP32 performance, reaching between 85 and 92 Teraflops. Currently, the RTX 3090 is just theoretically achieve 35 Teraflops--certainly a massive power upgrade.
All this processing might means the new GPU is an energy-chomping monster. Nvidia's Ada Lovelace AD102 may consume around 450 to 650 watts at peak, which confirms other rumors claiming the GPU will hover around 500w. As such, the AD102 is a big leap from the current 350w in the RTX 3090. Previous generations of the GPU "only" have a maximum of 250w.
AMD GPU to Feature Multi-Chip Module Design Similar to Ryzen, Epyc CPUs
A separate leak revealed that AMD would shift from a monolithic design to a Multi-Chip Module (MCM) for its next-generation GPU, the RX 7900 XT. The MCM layout is akin to the one used on AMD's CPU flagships, Ryzen and Epyc.
AMD is rumored to discard Compute Units (CUs) and replace them with Work Processor Groups (WGP), a report from WCCFTech said. The MCM will also carry a Graphics Core Die (GCD) and a Multi-Cache Die. The GCD comes from TSMC's 5-nm process, while the Multi-Cache Die is based on the chipmaker's 6-nm process. Having two GCDs per card, the RX 7900 XT will feature 7,680 cores per GCD, thus 15,360 cores in total--an enormous rise from the 5,120 "streaming processors" in the present RX 6900 XT.
The AMD 7900 XT is also set to carry 32GB of GDDR6 on a 256-bit interface, 256MB of Infinity Cache per module adding up to a total of 512MB of on-die "3D" cache. It will also have clock speeds reaching the mid-2GHz range. These push the GPU to reach a theoretical FP32 performance of nearly 75 Teraflops, a that is lower than the supposed power of Nvidia's Ada Lovelace AD102, but such spec does not indicate superior gaming performance. Board power will be around 350 to 550w, the leak added.
