Tuesday, June 11th 2024
Possible Specs of NVIDIA GeForce "Blackwell" GPU Lineup Leaked
Possible specifications of the various NVIDIA GeForce "Blackwell" gaming GPUs were leaked to the web by Kopite7kimi, a reliable source with NVIDIA leaks. These are specs of the maxed out silicon, NVIDIA will carve out several GeForce RTX 50-series SKUs based on these chips, which could end up with lower shader counts than those shown here. We've known from older reports that there will be five chips in all, the GB202 being the largest, followed by the GB203, the GB205, the GB206, and the GB207. There is a notable absence of a successor to the AD104, GA104, and TU104, because NVIDIA is trying a slightly different way to approach the performance segment with this generation.
The GB202 is the halo segment chip that will drive the possible RTX 5090 (RTX 4090 successor). This chip is endowed with 192 streaming multiprocessors (SM), or 96 texture processing clusters (TPCs). These 96 TPCs are spread across 12 graphics processing clusters (GPCs), which each have 8 of them. Assuming that "Blackwell" has the same 256 CUDA cores per TPC that the past several generations of NVIDIA gaming GPUs have had, we end up with a total CUDA core count of 24,576. Another interesting aspect about this mega-chip is memory. The GPU implements the next-generation GDDR7 memory, and uses a mammoth 512-bit memory bus. Assuming the 28 Gbps memory speed that was being rumored for NVIDIA's "Blackwell" generation, this chip has 1,792 GB/s of memory bandwidth on tap!
The GB203 is the next chip in the series, and poised to be a successor in name to the current AD103. It generationally reduces the shader counts, counting on the architecture and clock speeds to more than come through for performance; while retaining the 256-bit bus width of the AD103. The net result could be a significantly smaller GPU than the AD103, for better performance. The GB203 is endowed with 10,752 CUDA cores, spread across 84 SM (42 TPCs). The chip has 7 GPCs, each with 6 TPCs. The memory bus, as we mentioned, is 256-bit, and at a memory speed of 28 Gbps, would yield 896 GB/s of bandwidth.
The GB205 will power the lower half of the performance segment in the GeForce "Blackwell" generation. This chip has a rather surprising CUDA core count of just 6,400, spread across 50 SM, which are arranged in 5 GPCs of 5 TPCs, each. The memory bus width is 192-bit. For 28 Gbps, this would result in 672 GB/s of memory bandwidth.
The GB206 drives the mid-range of the series. This chip gets very close to matching the CUDA core count of the GB205, with 6,144 of them. These are spread across 36 SM (18 TPCs). The 18 TPCs span 3 GPCs of 6 TPCs, each. The key differentiator between the GB205 and GB206 is memory bus width, which is narrowed to 128-bit for the GB206. With the same 28 Gbps memory speed being used here, such a chip would end up with 448 GB/s of memory bandwidth.
At the entry level, there is the GB207, a significantly smaller chip with just 2,560 CUDA cores, across 10 SM, spanning two GPCs of 5 TPCs, each. The memory bus width is unchanged at 128-bit, but the memory type used is the older generation GDDR6. Assuming NVIDIA uses 18 Gbps memory speeds, it ends up with 288 GB/s on tap.
NVIDIA is expected to double down on large on-die caches on all its chips, to cushion the memory sub-systems. We expect there to be several other innovations in the areas of ray tracing performance, AI acceleration, and certain other features exclusive to the architecture. The company is expected to debut the series some time in Q4-2024.
Source:
kopite7kimi (Twitter)
The GB202 is the halo segment chip that will drive the possible RTX 5090 (RTX 4090 successor). This chip is endowed with 192 streaming multiprocessors (SM), or 96 texture processing clusters (TPCs). These 96 TPCs are spread across 12 graphics processing clusters (GPCs), which each have 8 of them. Assuming that "Blackwell" has the same 256 CUDA cores per TPC that the past several generations of NVIDIA gaming GPUs have had, we end up with a total CUDA core count of 24,576. Another interesting aspect about this mega-chip is memory. The GPU implements the next-generation GDDR7 memory, and uses a mammoth 512-bit memory bus. Assuming the 28 Gbps memory speed that was being rumored for NVIDIA's "Blackwell" generation, this chip has 1,792 GB/s of memory bandwidth on tap!
The GB205 will power the lower half of the performance segment in the GeForce "Blackwell" generation. This chip has a rather surprising CUDA core count of just 6,400, spread across 50 SM, which are arranged in 5 GPCs of 5 TPCs, each. The memory bus width is 192-bit. For 28 Gbps, this would result in 672 GB/s of memory bandwidth.
The GB206 drives the mid-range of the series. This chip gets very close to matching the CUDA core count of the GB205, with 6,144 of them. These are spread across 36 SM (18 TPCs). The 18 TPCs span 3 GPCs of 6 TPCs, each. The key differentiator between the GB205 and GB206 is memory bus width, which is narrowed to 128-bit for the GB206. With the same 28 Gbps memory speed being used here, such a chip would end up with 448 GB/s of memory bandwidth.
At the entry level, there is the GB207, a significantly smaller chip with just 2,560 CUDA cores, across 10 SM, spanning two GPCs of 5 TPCs, each. The memory bus width is unchanged at 128-bit, but the memory type used is the older generation GDDR6. Assuming NVIDIA uses 18 Gbps memory speeds, it ends up with 288 GB/s on tap.
NVIDIA is expected to double down on large on-die caches on all its chips, to cushion the memory sub-systems. We expect there to be several other innovations in the areas of ray tracing performance, AI acceleration, and certain other features exclusive to the architecture. The company is expected to debut the series some time in Q4-2024.
141 Comments on Possible Specs of NVIDIA GeForce "Blackwell" GPU Lineup Leaked
We only just 16GB on xx80(S) and then on xx70TiS, after staying with 8GB on xx80 for two generation and then upgrading to 10/12GB and for three generations on xx70(Ti) and then upgradig to 12GB. Don't expect NV to give you another upgrade after just on generation!
GTX 980 2015 4GB
GTX 980 Ti 2015 6GB
GTX 1080 2017 8GB
GTX 1080 Ti 2017 11GB
RTX 2080 2018 8GB
RTX 2080 Ti 2018 11GB
RTX 3080 2020 10/12GB
RTX 3080 Ti 2021 12GB
RTX 4080 2022 16GB
RTX 5080 2024 or 2025 16GB?
Good luck selling that.
Primarily, I don't think higher density GDDR7 dies are available yet, and when they finally are, they will come with a profit-eating higher cost. Nvidia will likely justify it on the $3000 5090, and maybe eventually on the 5080 Ti/Super/Ultra/Whatever but for the cost-effective models at xx50/60/70 we're going to get screwed because those cards are all about maximising Nvidia's profit, not pleasing end-users.
"If you want more VRAM, spend more money, you filthy peasants."
- Jacket Man, probably.It was with the 30-series, mostly due to spikes tripping the safeties on many high-end 850W units rather than average power draw of a 3080-equipped system exceeding 850W sustained power draw.
only circumstances I'd consider the 7900XTX is if you can purchase it for $300 less
Nvidia is factoring that in when deciding how much VRAM to equip their xx90 cards with. We've had two generations now with 24GB and prosumer workloads are demanding more VRAM than that.
Games themselves could probably use more VRAM as well if Nvidia wasn't intentionally holding back the market in that regard. You can't expect devs to use VRAM most people don't have and by extension this means Nvidia has great influence over how much VRAM games will use. If they continue to put 8GB VRAM cards out then 8GB of VRAM will continue to be enough so long as they are the dominent player in the market. Of course I expect it to increase but only when Nvidia absolutely has to, as goes with the trend of giving people as little as possible for their money whether that be up front or in regards to longevity.
It's ironic, in the enterprise market Nvidia caters to what customers what while in the gaming market gamers cater to what Nvidia wants. If gamers are going to argue for Nvidia that 8GB is still fine, perhaps they deserve 8GB cards ad infinium. It's a self perpetuating prophecy.
I am done with GPU's, this machine will all likely, be my last, it's been fun, but I just can't bother anymore. At certain tiers nVidia, you have priced yourself out of the market, everyone has their limits and you pissed on their concerns. No one like being punked or ripped off.
Would you believe me if I said I had so many Nvidia driver problems that I switched to AMD and never had any again?
Edit: forgot to mention
Canvas
Omniverse
NVENC
www.nvidia.com/en-us/studio/canvas/
www.nvidia.com/en-us/omniverse/
Things may change if we ever see a GPU with two GEN5 power connectors. In that case, we need at least a 1200W PSU (ideally 1500W+) to power something like the upcoming 5090 which could end up sipping 800W, if not that then maybe a 6090.
- DLSS4.x with ultra super framerate acceleration, definitely better than 3, so you can't miss it, and it can only run on Blackwell because Huang said so
- Even betterrer RT performance, which is what Nvidia is going to tout as the 'real performance gap' versus Ada, don't mind raster, its no longer relevant, there is always an RT game to distort reality
- Some subscription to some Nvidia service model (3 months of free GF Now?)
- A lot of marketing to drive the above home
-FSR SR FG SS XXX 4.1, definitely better than FSR 1.0, and just as good as DLSS because it's open source and Lisa Su Bae said so
-still lacking in RT performance, but more sponsored games with half assed effects to show they're barely behind in games where you struggle to spot the difference.
- +++++ VRAM so they can claim at least one spec advantage and cater to people who want more than a decade of high textures, well past the GPU's ability to push good fps in modern AAA titles
-some ridiculous marketing or PR blunder because it wouldn't be Radeon Technologies Group without them managing an own goal too
I just turn down settings rather than spending more money lol