It is anticipated that NVIDIA’s upcoming GeForce RTX 50 series graphics cards will make use of the new Blackwell GPU architecture. And there have been many rumours floating around in the market. TechTubers, RedGamingTech, and Moore’s Law is Dead, who recently released videos discussing NVIDIA’s next-gen architecture, are the sources of the most recent rumours.
Although this information isn’t confirmed and rumours should be taken with a grain of salt because more well-known leakers have previously made similar claims, Blackwell is still a few years away from release.
There have also been rumours that the Blackwell GPU architecture will debut in 2024, so GTC 2024 may mark the architecture’s first server-level debut before it is implemented in GeForce cards.
NVIDIA GeForce RTX 50 series will continue to use a monolithic die design and according to Moore’s Law is Dead, given what AMD was able to release during the first consumer-chipset generation in the form of Navi 31 “RDNA 3,” NVIDIA isn’t concerned by AMD’s chiplet strategy. The green is anticipated to continue using a monolithic die, which still has a number of benefits, the most important of which is that the product will be produced on schedule.
Maybe Blackwell.
— kopite7kimi (@kopite7kimi) July 19, 2022
Another monolithic chip, Ada-Blackwell, Next’s may have been revealed by Kopite7kimi last year. NVIDIA still uses monolithic GPUs in both its consumer and server offerings, suggesting that the company has so far fallen behind in chiplet designs. Although AMD is producing chiplet GPUs for its server offerings, the chipmaker may take the full chiplet route in the RDNA 4 generation.
The second rumour claims that NVIDIA will produce the Blackwell GeForce RTX 50 GPUs using TSMC’s 3nm process node.
Since Pascal GPUs were produced on TSMC’s 16nm node, Turing GPUs were produced on TSMC’s 12nm node, Ampere GPUs were produced on Samsung’s 8nm node, and the current Ada GPUs are produced on TSMC’s 4N (5nm Optimized) node.
According to rumours, the 3nm node from TSMC, which started volume production in Q4 2022, is better than anticipated. However, as amazing as the node may sound, it will also be very expensive. Insiders claim that TSMC 3nm wafers are 25% more expensive than TSMC 5nm wafers, which could lead to higher prices for next-generation GPUs like the NVIDIA GeForce RTX 50 “Blackwell” series. According to reports, the CEO of NVIDIA also travelled to Taiwan to meet with the CEO of TSMC to discuss securing 3nm wafers in advance for their upcoming GPU lineup.
It's too early to discuss Blackwell now. I won't be surprised if it has a 512-bit memory interface.
— kopite7kimi (@kopite7kimi) April 7, 2023
Although this isn’t a completely new design, the details begin by stating that Blackwell GPUs will feature a significant overhaul to the CUDA architecture. The denoising and Ray/Path Tracing hardware units will receive additional improvements, and the Blackwell SM units are anticipated to receive a “New Structure.”
The Blackwell GPUs are said to support GDDR7 memory; while GDDR6X still has some room to spare, GDDR7 would unquestionably be a better option if it offered greater efficiency. Though it’s been a while since NVIDIA went back to the Sammy team on the consumer-end chips since utilising Micron’s technology, Samsung is currently working on its brand-new GDDR6W DRAM which should provide even better performance & higher capacities.
Kopite7kimi also suggests that Blackwell might have a 512-bit SKU, even though NVIDIA hasn’t taken that path in a while. The vendor’s GTX 285 was its final 512-bit card, which was released back in 2009.
The PCIe Gen 5 interface and clock speeds of 3 GHz+ are also rumoured to be features of the NVIDIA Blackwell GeForce RTX 50 series. With 144 SMs and 96 MB of L2 cache, the GB102 GPU is expected to be the gaming lineup’s top model. The Ada AD102 GPU has the same number of SMs and cache units, but it appears that the fundamental changes to the architecture will result in a significant performance increase.
Also Read:
