Entertainment New Nvidia Lovelace spec leak is downright scandalous
New Nvidia Lovelace spec leak is downright scandalous
As we get closer to the anticipated launch of Nvidia Lovelace later this year, more spec and benchmark leaks are starting to pop up online, and boy howdy, the flagship Lovelace card could be an absolute monster of a GPU.
The latest news comes courtesy of reliable Twitter hardware leaker kopite7kimi (opens in new tab), who says that Nvidia is currently testing a variant of its AD102 GPU with an astonishing 18,176 CUDA cores, or nearly double the number of cores of the Nvidia RTX 3090 Ti, currently the best graphics card on the consumer market in terms of raw performance.
“the beast” PG137-SKU0AD102-450-A118176FP3248G 24Gbps GDDR6Xtotal board power ~800WJuly 25, 2022
Kopite7kimi calls the card “the beast”, and that is entirely appropriate if the leaked specs are anywhere close to what the actual card has at launch. In addition to the 18,176 CUDA cores, the AD-102-450-A1 GPU will also support 48GB of GDDR6X VRAM and could be rated for about 800W of TGP, which is in line with a lot of what we’ve heard about Lovelace’s power-hungry design.
There are some other things that we could deduce from these specs. Assuming that there isn’t some further revolutionary architectural change in the stream multiprocessors (SM), each SM should have 128 CUDA cores, which would mean that the card would have 142 SMs. If each SM only has a single ray tracing core, as is the case with Nvidia Ampere, that would also mean 142 ray tracing cores.
If Lovelace follows Ampere’s lead and includes four tensor cores per SM, that would mean 568 tensor cores capable of accelerating the advanced FP32 operations required for machine learning and ray tracing features of the card. All in all, this is about a 70% increase over the RTX 3090 Ti, but the actual performance gains could be even greater than that.
We can’t say for certain whether this is an engineering sample of an RTX 4090 Ti or some other higher end variant of the next-gen BFGPU. A couple of months back, we got word of a Titan-class AD102 GPU with higher core counts and TGP, but this would likely be a workstation graphics card, assuming it was ever actually released.
The expectation is that this new design is a consumer card straddling the border between a card found in the very best gaming PCs and professional workstations.
Analysis: Holy CUDA, this is going to be an absolutely massive graphics card
The RTX 3090 Ti is huge. It weighs several pounds, is over a foot long, and putting one in an upright tower case requires additional physical support so that the torque on the PCIe connector doesn’t damage the board. It simply won’t fit in a lot of cases on the market.
Let’s say for argument’s sake that this is an RTX 4090 Ti. Just figuring out how you’re going to fit this kind of card in even a full tower case is a serious design challenge, because if this card is sucking up as much as 800W of energy, it’s going to require an enormous cooling solution to keep it running safely.
What’s more, the RTX 3090 Ti is already a $2,000+ graphics card. How much more expensive is this card going to be if this is the hardware it’s packing? We need to take any leak with a grain of salt, if not a heaping spoonful, but everything is pointing towards a massive increase in the number of cores, VRAM, and power draw for the flagship Lovelace cards.
We can’t know for sure until there’s some kind of official announcement, but if it’s anything close to what these specs indicate, our concerns about the rising costs of graphics cards look downright quaint.