There are still at least a few months before Nvidia and AMD release the next generation of GPUs, but there have been related news leaks from multiple channels recently. If you pay attention to the information about NVIDIA’s AD102 based on the Ada Lovelace architecture in the past, or AMD’s Navi 31 based on the RDNA 3 architecture, you already have a certain understanding of the general situation of the new generation of flagship products.
Twitter user
@kopite7kimi revealed that NVIDIA’s next-generation flagship graphics card is based on PG137/139-SKU30 PCB, equipped with AD102-300, equipped with 24GB of video memory, the rate is 21Gbps, and the power consumption will reach 600W. Compared with the 24Gbps video memory that has been circulated before, the memory rate in this rumor has been reduced. The overall configuration of video memory should be basically the same as the recently released GeForce RTX 3090 Ti, with a bit width of 384 bits and a memory bandwidth of about 1TB/s. As for the power consumption of 600W or above, after the introduction of the latest 12VHPWR interface, there is no problem.
As for the rival AMD, Twitter user
@greymon55 revealed that there will be seven small chips, namely two GCDs using a 5nm process, four MCDs using a 6nm process, and one interconnects control chip. The specific purpose of MCD is unclear for the time being. It is rumored to be a cache-type composite chip, which may be located on top of GDC, responsible for video memory control or including Infinity Cache. There are still too many questions.
Although Intel has just released the Alchemist (DG2) discrete graphics card of the Intel Arc brand based on the Xe-HPG architecture, there are only two products for mobile platforms, A350M and A370M. Intel Arc graphic card for the desktop platform that many gamers care about most is still a while away. However, Twitter user
@Redfire75369 has already revealed some news about the next-generation Battlemage graphics card, which Intel internally calls Elasti.
Intel is targeting the GPU at a higher level, with a release as early as 2023. It is said that Battlemage adopts a “Tiled-GPU” design method, and the number of rendering slices per tile is 10, that is, 5120 FP32 cores, which will be 25% more than the current full ACM-G10 (SoC1) chip, while Battlemage may have up to four tiles.