- This is a 70% boost over the current top consumer graphics card.
- The card has 568 tensor cores for machine learning and ray tracing.
- Nvidia Lovelace will be released later this year.
As the planned introduction of Nvidia Lovelace later this year approaches, more spec and benchmark leaks are beginning to surface online, and boy howdy, the flagship Lovelace card might be an absolute monster of a GPU.
The newest rumours comes from Twitter hardware leaker kopite7kimi, who claims Nvidia is testing a variation of their AD102 GPU with 18,176 CUDA cores, roughly twice the amount of cores of the Nvidia RTX 3090 Ti, the top consumer graphics card in terms of raw performance.
“the beast”
PG137-SKU0
AD102-450-A1
18176FP32
48G 24Gbps GDDR6X
total board power ~800W— kopite7kimi (@kopite7kimi) July 25, 2022
If the rumored specifications are accurate, Kopite7kimi’s nickname is apt. The AD-102-450-A1 GPU will support 18,176 CUDA cores, 48GB of GDDR6X VRAM, and 800W of TGP, which is in accordance with Lovelace’s power-hungry architecture.
These specifications reveal additional stuff. Each stream multiprocessor (SM) should have 128 CUDA cores, therefore the card would have 142 SMs. 142 ray tracing cores if each SM has one, like Nvidia Ampere.
If Lovelace follows Ampere’s example and incorporates four tensor cores per SM, the card will have 568 tensor cores to accelerate machine learning and ray tracing. This is a 70% boost over the RTX 3090 Ti, although real performance increases might be higher.
We don’t know whether this is an RTX 4090 Ti engineering sample or another high-end BFGPU. A few months ago, we heard about a Titan-class AD102 GPU with more cores and TGP, but this would likely be a workstation graphics card.
This new design is expected to be a consumer card for high-end gaming PCs and workstations.
[embedpost slug=”nvidia-will-pay-a-5-5-million-sec-fine-for-inadequate-disclosures-over-crypto-mining/”]
