These days it seems like everything in the GPU world is beginning to fall back to earth. We all know prices have begun to reflect the impact of post-pandemic gravity. Also, several weeks ago AMD’s upcoming RDNA3’s specs were taken down a peg by the rumor mill. Now, it’s Nvidia’s turn in the barrel. Noted GPU rumor monger @kopite7kimi has released a “summary” of the AD102 die’s rumored specs, and they’re decidedly more sane than what they previously reported.
The biggest news, in our opinion, is the Total Board Power (TBP) has been significantly reduced. It was first reported to be an eye-popping 850W back in February. This naturally caused a lot of consternation as it’s totally unheard of for a consumer GPU. That number was dialed back a bit to 600W in April. This aligned with leaked PCI Express 5.0 specs that showed you could push that much power through a single GPU cable. Now the official rumor is the RTX 4090’s TBP is a much more palatable 450W. The reduction in power consumption is due to a downgrading of the total number of CUDA cores.
Previously it was rumored the RTX 4090/AD102 would offer around 140 Streaming Multiprocessors (SM). This would have granted it 18,432 CUDA cores. It’s now being reported it will only offer 126 SMs, bringing its CUDA core count down to 16,128. That’s a 13 percent drop, so it’s kind of a big deal. However, the full AD102 die will reportedly still feature those “missing” 2,304 CUDA cores according to Videocardz. Therefore, it sounds like the RTX 4090 Ti could be the 600W beast with everything enabled. Despite the downgrade, the RTX 4090 is still rumored to offer twice the performance of the RTX 3090, which is hard to believe. It’s not clear if that metric is for rasterization performance, ray tracing, or both.
What’s left unchanged is the memory specs, which are still 24GB of 21Gb/s GDDR6X. This is the same configuration as the RTX 3090 Ti, which also has a TBP of 450W. It’s difficult to predict how much memory bandwidth will be available though. Nvidia is rumored to be upping the L2 cache significantly for Ada Lovelace, which will have a big impact. It’ll still feature a 384-bit memory bus though, like the current flagship GPU.
The latest Nvidia rumors follow AMD’s RDNA3 specs also getting downgraded recently. Interestingly, this same leaker notes in his tweet that they’re “disappointed” in AMD’s next-gen. It’s unclear why, as it was previously noted that despite the downgrade the company was still hitting its performance goals.
All will be revealed soon, as Nvidia’s Ada Lovelace GPUs are supposedly launching in “early Q3.” That vague timelines has been translated to mid-July, which is two months before Nvidia typically launches its GPUs. There’s been no explanation for why Nvidia is moving up its timeline, but it will likely beat AMD to market. The company’s RDNA3 GPUs aren’t expected to arrive until the end of the year. This launch timeline makes sense since AMD just launched its RDNA2 refresh one week ago. It certainly wouldn’t want to cannibalize its own just-launched GPUs, whereas Nvidia might be less concerned about that. After all, it’s previously stated it might sell its Ampere GPUs alongside Ada Lovelace.