Rumors about Nvidia’s next-gen graphics cards from the Geforce RTX 5000 series (“Blackwell”) have been in the headlines for months. PCGH tries to classify these seriously based on the current facts. Once again, Twitter user @kopite7kimi, who is known for such advance publications, would like to find out which graphics processors will be used on the next generation Geforce graphics cards and how they should be designed.
Nvidia should also rely on chiplets
As can be seen from the rumor mill around @kopite7kimi and other leakers, Nvidia is also said to be switching to a multi-chiplet module (“MCM”) and specifically the graphics units of the upcoming Geforce RTX 5000 (“Blackwell”), also known as “Ada Lovelace Next”, fundamentally remodel accordingly. This concept is known to already be in use with AMD’s Navi 31 and Navi 32 GPUs.
Recommended editorial contentHere you will find external content from [PLATTFORM]. To protect your personal data, external integrations will only be displayed if you confirm this by clicking on “Load all external content”:Load all external contentI agree that external content will be displayed to me. This means that personal data is transmitted to third-party platforms. Read more about our privacy policy.
Furthermore, the number of Graphics Processing Clusters (“GPCs”) and Texture Processing Clusters (“TPCs”) will not change much, but their internal structure will be fundamentally revised. This process could also be observed during AMD’s generation change from RDNA 2 to RDNA 3.
Recommended editorial contentHere you will find external content from [PLATTFORM]. To protect your personal data, external integrations will only be displayed if you confirm this by clicking on “Load all external content”:Load all external contentI agree that external content will be displayed to me. This means that personal data is transmitted to third-party platforms. Read more about our privacy policy.
However, the rumor mill considers it rather unlikely that Nvidia will switch from TSMC back to Samsung Foundry and its 3 nm node (“3GAA”), which is targeted for 2025, for the production of its graphics processors.
GB202, GB203, GB205, GB206 and GB207 for Blackwell
In the meantime, the user “Panzerlied” from the Chiphell forum, which is known for such information, has also confirmed that he has information about the graphics processors GB202, GB203, GB205, GB206 and GB207 for the upcoming graphics cards in the Geforce RTX 5000 series (“Blackwell “), which is currently listed as “Ada Lovelace Next” on an Nvidia roadmap. Nvidia itself has not yet commented on this.
GPU roadmap Source: Nvidia A leaked GPU roadmap from Nvidia suggests that “Ada Lovelace Next” and thus the Geforce RTX 5000 series will not appear until the year after next. Nvidia’s current lead over its competitors is too great, so the company will be in no hurry here.
Geforce RTX 5090 and 5080 are probably the start
The graphics processors AD102, AD103, AD104, AD106 and AD107 (“Ada Lovelace”) will most likely be followed by the new GPUs GB202 and GB203 (“David Blackwell”). Numerous further expansion stages with an even higher identifier are also expected here, which are planned for the Geforce RTX 5070 (Ti) and the smaller model series. Nvidia should stay true to its formula.
GDDR7 with 36 GiBit/s and PCIe 5.0 x16 are considered set
In addition to a connection via PCIe 5.0 x16, the new GDDR7 graphics memory will be used at least for the Geforce RTX 5090 and Geforce RTX 5080. This should achieve up to 36 gigabits per second and rely on three-stage pulse amplitude modulation (“PAM-3”) for signal transmission. It can be assumed that energy efficiency will be improved by around 25 percent in this sub-area. GDDR6X, on the other hand, already uses PAM-4.
The number sequence “23567” stands for GB202, GB203, GB205, GB206 and GB207 Source: Panzerlied via Chiphell The effective memory bandwidth would also benefit significantly from the fast GDDR7 memory with up to 36 GiBit/s:
128-Bit @ 36 Gbps: 576 GiByte/s192-Bit @ 36 Gbps: 846 GiByte/s256-Bit @ 36 Gbps: 1.152 GiByte/s320-Bit @ 36 Gbps: 1.440 GiByte/s384-Bit @ 36 Gbps: 1.728 GiByte/s512-Bit @ 36 Gbps: 2.304 GiByte/s
As can be seen from several entries in the Chiphell forum, Nvidia could also take a two-pronged approach:
GB202: GDDR6X (“PAM-4”) mit 512-BitGB203: GDDR6X (“PAM-4”) mit 384-BitGB205: GDDR7 (“PAM-3”) mit 256-BitGB206: GDDR7 (“PAM-3”) mit 192-BitGB207: GDDR7 (“PAM-3”) mit 128-Bit
But even this is usually just speculation, which is based on information that has been leaked so far, but cannot yet be viewed as binding. Nvidia itself has not yet said a word about the next generation of its Geforce graphics cards.
Performance predictions are almost impossible
The rumor mill predicts that the upcoming top model, the Geforce RTX 5090, will have up to 2.5 times the performance of an already very fast Geforce RTX 4090. However, great caution is advised here for the time being, because it is currently completely unclear whether DLSS 2 or even DLSS 3 with frame generation will be included in the equation. Performance predictions are currently almost impossible.
Geforce RTX 5090 with up to 18,432 CUDA cores
The specifications assigned to the Geforce RTX 5090 are also currently purely speculative. Accordingly, in the future it will go from 128 streaming multiprocessors and 16,384 shader units to 144 streaming multiprocessors and 18,432 CUDA cores. In addition, the L2 cache will be increased from 72 MiByte to 96 MiByte and the GPU clock speed will increase to more than 3 GHz. Videocardz has already compared the graphics processors of the three generations.
GPUs of the Ampere, Ada Lovelace and Blackwell generation in comparison Source: Videocardz The PCGH editorial team will report on this if valid information about Nvidia Geforce RTX 5000 is available.
Your opinion is asked!
What is your opinion on this topic? The PCGH editorial team would be pleased to hear your well-founded opinion in the comments to this report. To comment, you must be logged in to PCGH.de or the Extreme forum. If you don’t have an account yet, you can register here without obligation.
Those: @kopite7kimi via X (Twitter), Panzerlied via Chiphell, Videocardz
