©
As they say, video cards NVIDIA GeForce RTX 50 the next generation will use the new Blackwell architecture. Recently, there have been many rumors about the chip, so everything that is known has been impoverished together.
So far there has been news from TechTubers, RedGamingTech and Moore’s Law is Dead, who have posted their videos on NVIDIA’s next generation architecture. Since the Blackwell launch is still a couple of years away, this information is not reliable, and the rumors should be taken with a grain of salt, because they are interesting because they converge with the previous ones.
Blackwell is the code name for NVIDIA’s next-generation GPU architecture, which is supposed to replace Hopper. Blackwell is built on the GPU
GB102 in a consumer variant in addition to the leaked industrial GB100 GPU.
The Blackwell GPU architecture will be introduced in 2024, meaning it could be the first server-level implementation before the architecture hits GeForce cards.
So, the NVIDIA GeForce RTX 50 series, which uses Blackwell GPUs, will keep monolithic approach to the crystal. Moore’s Law is Dead says NVIDIA isn’t worried about AMD’s chip strategy given what they were able to release during the first generation of GPUs in the form of Navi 31 “RDNA 3”. NVIDIA is expected to continue to use a monolithic die that still retains more of the benefits, the biggest of which is timely production of graphics cards.
A year ago, insider Kopite7kimi also reported that Blackwell could be another monolithic chip. It appears that NVIDIA has been at the bottom of chiplet design so far, as its consumer and server offerings still use monolithic GPUs. AMD may be going down the path of a full chiplet in the RDNA 4 generation, but the chipmaker makes graphics chiplets for its server offerings.
It is mentioned that NVIDIA will use 3nm TSMC process for the production of Blackwell GPUs. So far, all recent NVIDIA GPU releases have been node reductions, starting with Pascal GPUs that were made on 16nm TSMC, Turing on 12nm TSMC, Ampere on 8nm Samsung, and current Ada GPUs use 5 -nm optimized TSMC 4N node.
Rumor has it that TSMC’s 3nm wafers are over $20,000, pushing up the cost of next-generation Intel, AMD, and NVIDIA CPUs and GPUs.
TSMC began mass production of its 3nm node in the fourth quarter of 2022, and according to rumors, it turned out better than expected, however, as impressive as it sounds, it will also be very expensive. According to insiders, TSMC’s 3nm wafers are on 25% more expensive than TSMC’s 5nm wafers, which could lead to higher prices for next-gen GPUs like the NVIDIA GeForce RTX 50 “Blackwell” series. The CEO of NVIDIA also reportedly visited Taiwan to talk with the CEO of TSMC about the safety of 3nm wafers for their next-generation GPU lineup.
RedGamingTech channel shared details from its own sources that Blackwell GPUs will have a massive re-architectural overhaul CUDA. SM Blackwell units are expected to receive “new structure”as well as further optimizations and additions to the anti-aliasing, ray tracing, and path hardware blocks.
Given NVIDIA’s recent push to use path tracing in modern AAA games, it looks like we’ll get even more powerful hardware to handle them. It is possible that RT units will be replaced by stronger PT units.
Blackwell GPUs will support memory GDDR7and while GDDR6X still has some headroom, if the new GDDR7 delivers higher efficiency, then this would definitely be a better choice.
Samsung is already working on its all-new GDDR6W DRAM which should deliver even better performance and higher capacity, although it hasn’t been that long since NVIDIA returned to Samsung memory in the consumer chip space with technology from Micron. NVIDIA uses Samsung HBM solutions for its HPC and AI chips.
The decision to use GDDR7 does not seem far-fetched at all, given that Cadence has already submitted the first samples for verification testing of the new memory standard. Even Samsung has announced that it will offer transfer rates up to 36Gbps with a PAM3 signal on GDDR7 chips, and Micron is expected to introduce its solution sooner or later. There’s still more than a year to go in between the GDDR6W and GDDR6X upgrade unless an RTX 4090 Ti is announced.
Kopite7kimi also says that Blackwell may have 512-bit memory bus interface, although NVIDIA has long gone down this path. The last 512-bit card from the manufacturer was the GTX 285 back in 2009. NVIDIA’s upcoming monstrous Blackwell GPU can achieve over 2TB/s of bandwidth using 36Gb/s chips.
- 128-bit at 36 Gb/s – 576 GB/s.
- 192-bit at 36 Gb/s – 846 GB/s.
- 256-bit at 36 Gb/s – 1152 GB/s.
- 320-bit at 36 Gb/s – 1440 GB/s.
- 384-bit at 36 Gb/s – 1728 GB/s.
- 512-bit at 36 Gb/s – 2304 GB/s.
GeForce RTX 50 may receive an interface PCIe Gen 5 and expected clock frequencies over 3 GHz. The GB102 GPU should be the flagship of the gaming line and is said to support 144 SM and 96 MB of L2 cache. This is the same amount of SM and cache as the full featured Ada AD102 GPU, but it looks like the underlying architecture changes will provide a significant performance boost.
As a result, performance is likely to increase in double in terms of Ada GPUs that offered a significant leap from the previous generation, such as the RTX 4090 with a stripped-down AD102 core, but there were significant improvements across all segments (rasterization/ray tracing/DLSS).
Further improvements could take Blackwell GPUs to the next level, it remains to be seen what NVIDIA has in store.