Graphics Cards Break the 512-Injured Storage Barrier!
Blackwell GB202 GPU
There will not be a 512-injured storage capacity connector for the graphics cards manufactured by flagship product, the Blackwell GB202 GPU, according to rumors. Instead, it will have 384-bit GDDR7.
The well-known hardware leaker kopite7kimi made a mistake in their own forecast, which said that Nvidia’s future-oriented flagship graphics processor unit would have a 512-bit memory interface if it was based on the Blackwell architecture. It may not happen. It seems like Nvidia’s GPU processing unit, and this is being referred to asGB202 GPU, is going to keep to make use of a 384-bit memory bus, but it will switch to using GDDR7 type of memory instead.
It is anticipated that Nvidia’s GB202 GPU would come equipped with as many as 24,576 CUDA cores, which represents a 33% increase in comparison to the amount of CUDA cores that were included in the in AD102 Graphics (18,432 CUDA cores).
It is unknown if Nvidia and TSMC will tweak Taiwan’s 3nm-class nodes for GPUs or whether they will keep to the default settings for those nodes. However, rumors claim the business will utilize one of TSMC’s 3nm-class manufacturing methods to produce its Graphics Processing Units (GPUs) depending on the Blackwell architecture.
“he thinks he might have committed a qualitative mistake,” added kopite7kimi in a subsequent message. “he thinks he obviously made an analytical mistake.” “he wrongly matched the dimensions of Sandra Lovelace’s L2 [cache] as well as [memory controller] into Blackwell, among others [things that led into an unfounded supposition with respect to GB202’s 512-got reminiscence connectivity].”
They responded affirmatively when requested whether this mean a 384-bit storage capacity interface for the GB202 GPU, and they clarified that the component would also utilize GDDR7 in their response.
Utilizing GDDR7 memory over a 384-bit interface would also bring advantages in addition to those provided by utilizing a 512-bit bus, which would allow Nvidia to dramatically boost the amount of performance given to its next-generation powerhouse graphics cards (likely to be called the the company GeForce RTX 5090).
If Nvidia decides to employ Micron’s 32 GT/s 16 Gb ICs, then the RTX 5090 will have 1.536 TB/s of memory bandwidth. This is an increase over the 1.008 TB/s that the RTX 4090, which is now the most powerful graphics card on the market, receives. Nvidia’s premium consumer board will only be able to accommodate 24 GB of memory, despite the company’s use of 16 GB integrated circuits.
