NVIDIA Ada Lovelace ‘GeForce RTX’ & AMD RDNA 3 ‘Radeon RX’ Next-Generation GPUs Rumored To Be More Than Twice As Fast Than Ampere & RDNA 2
Information regarding the next-generation GPUs from NVIDIA and AMD has been rumored by the reliable leaker, Kopite7kimi. Based on his latest tweets, it looks like AMD and NVIDIA will be offering double-digit performance gains with their next-generation GeForce RTX and Radeon RX graphics cards.
AMD & NVIDIA Aiming To Offer Double-Digit Performance Gains With Next-Gen Ada Love 'GeForce RTX' & RDNA 3 'Radeon RX' GPU Powered Graphics Cards
In his tweet, Kopite7kimi reports that he has received tips from his sources and while the leaker has had an absolutely flawless track record with leaks, we will still advise treating this info with a grain of salt since these products are still more than a year away from launch and performance estimates could change dramatically once we get close to that point of time. With that said, the leaker mentions performance estimates and some specifications for both next-generation AMD and NVIDIA GPUs.
-5nm(no matter TSMC or SEC)
-AD102 in transition or GH202 in revolution?
-RDNA3, GFX11, how does AMD reach its perf goal? Double the SIMD in CU?
-How about Intel?
— kopite7kimi (@kopite7kimi) May 22, 2021
Starting with the details, it is reported that next-generation GPUs are likely to be based on a 5nm process node and will either use TSMC's or Samsung's design. The key architectures that have been discussed include NVIDIA's Ada Lovelace (AD102), Hopper (GH202), and RDNA 3 (Navi 3X). Based on the details, it seems like while Ada Lovelace will be a small transition architecturally from Ampere, it will bring a massive boost in performance while Hopper will be the revolution that NVIDIA has been waiting for, both in terms of design (power/package) and performance (raster/ray-tracing).
Kopite7kimit believes that with the next-generation cards, we are likely to get up to 100 TFLOPs of FP32 compute power in a single graphics card. That sounds a bit too optimistic but given that GPUs today rely on faster and efficient compute models, 100 TFLOPs doesn't sound that far-fetched and we could see a sparsity model for consumer-specific chips to boost their computing performance by a huge margin.
When it comes to performance numbers, the NVIDIA Ampere GA102 architecture is used as a base line. Compared to that, the NVIDIA Ada Lovelace (AD102) chip is said to offer a 2.2x increase in perf while AMD's RDNA 3 (Navi 3X) is said to offer over 2.5x perf increase.
There have been several rumors stating that the upcoming RDNA 3 GPUs are going to outperform whatever NVIDIA has to offer in terms of rasterization performance. It looks like AMD will take the lead by offering the first MCM-powered GPUs under its Radeon RX graphics card lineup. But at the same time, NVIDIA is expected to quickly transition to its Hopper MCM GPU lineup which will offer over 3x performance boost over Ampere GPUs. Following is the expected performance increase:
based on drafts, by my calculation only, the perf scale I guess:
NAVI31(4 GCDs) >2.5X
I'm optimistic. But Jensen and Lisa need time to solve problems especially bandwidth. Boring days, wait and see.
— kopite7kimi (@kopite7kimi) May 22, 2021
It is stated that one of the major bottlenecks for both AMD and NVIDIA next-generation GPUs would be memory bandwidth. GDDR6X won't be able to cut it so we might see a return to HBM or a newer GDDR memory standard that's available around 2022-2023. It is likely that GPU makers will be investing more in close-to-GPU caches such as the Infinity Cache featured on RDNA 2 graphics card to bypass bandwidth bottlenecks and conserving cost. Power consumption is also expected to go dramatically up and this generation was only a glimpse of what to expect from the next-gen line of graphics cards.
NVIDIA & AMD GPU Roadmaps:
NVIDIA / AMD / Intel GPU Roadmap (Rumor)
|AMD (RDNA 1)||AMD (RDNA 2)||N/A||AMD (RDNA 3)||N/A||AMD (RDNA 4)?|
|NVIDIA (Turing Refresh)||NVIDIA Ampere||N/A||NVIDIA Ampere Refresh|
NVIDIA Ada Lovelace
|N/A||Xe-LP||N/A||Xe-HPG (ARC Alchemist)||TBC||TBC|
Kopite7kimi also hinted at some specification details of the NVIDIA Ada Lovelace chips a while back which you can read more about here and check out the specs in the table provided below:
NVIDIA CUDA GPU (RUMORED) Preliminary:
|Flagship SKU||RTX 2080 Ti||RTX 3090 Ti||RTX 4090?|
|Process||TSMC 12nm NFF||Samsung 8nm||TSMC 4N?|
|Graphics Processing Clusters (GPC)||6||7||12|
|Texture Processing Clusters (TPC)||36||42||72|
|Streaming Multiprocessors (SM)||72||84||144|
|L2 Cache||6 MB||6 MB||96 MB|
|Theoretical TFLOPs||16 TFLOPs||40 TFLOPs||~90 TFLOPs?|
|Memory Capacity||11 GB (2080 Ti)||24 GB (3090 Ti)||24 GB (4090?)|
|Memory Speed||14 Gbps||21 Gbps||24 Gbps?|
|Memory Bandwidth||616 GB/s||1.008 GB/s||1152 GB/s?|
|PCIe Interface||PCIe Gen 3.0||PCIe Gen 4.0||PCIe Gen 4.0|
|Release||Sep. 2018||Sept. 20||2H 2022 (TBC)
As for AMD's Navi 31 GPU, it has been rumored to offer more than 2.5x performance before too by KittyYYuko who's also just as reliable as Kopite7kimi in the GPU leak scene. As for Hopper, it was said to be delayed in favor of Ada Lovelace a while ago but once again, it is possible we can see the AD10* and GH20* GPUs to be introduced close to one another however NVIDIA's own roadmap cites that their 'Ampere Next Next' GPU won't be here till 2024.