NVIDIA GeForce RTX 50 “Blackwell” Flagship Reportedly Options GB202 GPU With Up To 192 SM Models & 512-bit Bus

NVIDIA’s next-gen GeForce RTX 50 “Blackwell” GPU rumors have began to roll out from dependable leakers corresponding to Kopite7kimi.

NVIDIA GeForce RTX 50 “Blackwell” Flagship Reportedly Expands Upon RTX 40 Collection With Elevated SM Depend, Wider Bus Interface, Elevated Cache & Extra

NVIDIA’s GeForce RTX 50 “Blackwell” GPU rumors already started a number of months in the past when the final of the GeForce RTX 40 GPUs had been accomplished making their option to the market. Labeled as “Ada-Subsequent”, these next-gen chips would be the foundation of NVIDIA’s model new gaming lineup which targets a 2025 launch date in line with the official roadmap however rumors are additionally suggesting that the launch could occur earlier.

As I discussed earlier than, GA100 is 8*8, and GH100 is 8*9. GB100 could have a fundamental construction like 8*10. GB202 seems to be like 12*8.

— kopite7kimi (@kopite7kimi) September 28, 2023

So beginning with the main points, Kopite7kimi posted on X about two configurations of Blackwell GPUs. The primary one is the HPC/AI-oriented chip often called GB100 which has lately been said to make the most of the TSMC 3nm course of node and focusing on a late 2024 launch (announcement throughout GTC 2024).

The GB100 GPU is anticipated to be the primary HPC chip from NVIDIA to make the most of an MCM design and will probably be based mostly on an 8 GPC cluster which incorporates 10 TPCs per cluster and every cluster will carry 2 SMs for a complete of 160 SM items on the totally enabled die. The highest die will even function an 8192-bit extensive bus interface which can help the most recent HBM requirements corresponding to HBM3e.

Each Ampere & Hopper function totally different FP32/FP64 core rely arrangments but when NVIDIA had been to comply with the 128 FP32 core rely per SM for Blackwell, it might find yourself with a doable 20,480 FP32 cores on a totally enabled die. The next is how the NVIDIA HPC elements examine towards Blackwell GB100:

• A100 (Ampere) – 8 GPCs / 64 TPCs / 128 SMs / 64 Cores Per SM / 8192 Cores / 5120-bit
• H100 (Hopper) – 8 GPCs / 72 TPCs / 144 SMs / 128 Cores Per SM / 18,432 Cores / 5120-bit
• B100 (Blackwell) – 8 GPCs / 80 TPCs / 160 SMs / 128 Cores Per SM / 20,480 Cores  / 8192-bit

Transferring again to the gaming half, the GB202 GPU is rumored to function a vastly totally different GPU config as we’ve got seen within the earlier gaming/HPC launches. The chip is anticipated to accommodate 12 GPCs with a complete of 8 TPCs which might complete as much as 96 TPCs on the total die or 192 SMs. As soon as once more, if NVIDIA is to make use of the identical 128 FP32 cores per SM, you stand up to 24,576 cores which might mark a 33% uplift in core configuration over the total AD102 GPU. After all, we’ve got but to see a gaming GPU with the total AD102 GPU so NVIDIA is more likely to launch a cut-down GB202 die with its next-gen GeForce RTX 50 gaming lineup too with a higher-end variant making its option to the market when GPU yields turn into higher or if there is a must deal with the competitors.

NVIDIA has moved away from including simply conventional cores to its GPUs and now consists of numerous various kinds of cores and accelerators for AI, Tensor, Neural Processing & ray tracing operations inside its GPUs so by the point NVIDIA introduces Blackwell, the prevailing Ada Lovelace configuration could very properly be an outdated design.

Kopite7kimi additionally reiterates that the NVIDIA GB202 “Blackwell” GPU for GeForce RTX 50 GPUs goes to get a a lot wider 512-bit bus interface, a 33% improve over the 384-bit extensive bus interface that is being featured on present flagship chips.

GB100 8192-bit, GB202 512-bit.

— kopite7kimi (@kopite7kimi) September 28, 2023

There are additionally some rumors coming in from Chiphell Boards which recommend NVIDIA’s GeForce RTX 50 “Blackwell” flagship incorporates a 50% improve in core rely, a 52% uplift in reminiscence bandwidth, a 78% improve in cache dimension, and a 15% improve in core frequency, all leading to a 70% uplift within the total GPU efficiency capabilities. It’s nonetheless a bit too early to inform what the ultimate specs NVIDIA will resolve for its GeForce RTX 50 flagship graphics card and the corporate is understood to work on a number of SKUs earlier than deciding which one really makes it to the market and since we’re a yr away from launch, it will likely be unwise to name something remaining this early. However based mostly on these stories, a GeForce RTX 50 GPU would function:

• 24,576 CUDA Cores (GB202 GPU)
• 32 Gbps Reminiscence Speeds (GDDR7)
• ~3000 MHz Peak GPU Clock Speeds
• 128 MB L2 Cache (For GPU)

Samsung and SK Hynix are already reported to have began sampling the next-gen GDDR7 DRAM modules to NVIDIA for its next-gen GPU lineup. The brand new modules are anticipated to function as much as 32 Gbps pin speeds, delivering as much as 2 TB/s of bandwidth throughout a 512-bit bus interface. That can mark an enormous improve in GDDR bandwidth capabilities and a 2x improve over the present quickest RTX GPU such because the 4090.

All of that is fascinating stuff for positive however we’ve got to do not forget that these are rumors and we’ve got to attend and see how a lot finally ends up being true by the point the next-gen GeForce RTX sequence launch.

NVIDIA GeForce GPU SKUs:

Information Supply: VideoCardz