Nvidia seemingly confirms some RTX 4000 specs with the H100 Hopper announcement
Judging by Nvidia’s past gaming GPU release schedules, 2022 should be the year when Team Green is launching a new generation, which is rumored to bear the Lovelace codename. Unlike previous years, though, the gaming GPU codename is not the same as the compute GPU one, but the two usually share the same architecture, so the recently announced H100 Hopper compute GPUs could actually provide some interesting clues about the Lovelace models that are expected to launch in the second half of 2022. Moore’s Law Is Dead is pointing out that the Hopper architecture confirms two important aspects related to Lovelace’s performance and TGP specs that might have been exaggerated in previous rumors.
First off, the rumored chiplet designed for Hopper does not seem to be true, or it could launch at a later time. MLID believes that Nvidia presenting a monolithic H100 is enough to confirm that Lovelace is also monolithic. AMD’s chiplet (MCM) design for RDNA3 expected to launch later this year as well could have the upper hand when it comes to energy efficiency and die yields. Nvidia’s Lovelace, on the other hand, may deliver better performance due to the sheer size of the die.
The second aspect pointed out by MLID refers to the expected TGP specs of the upcoming Lovelace GPUs. There were some rumors suggesting that Lovelace could hit 850 W TGPs, but, from the official Hopper specs, we now know that Lovelace cannot require more than 700 W (limited to 350 W for the PCIe version). MLID is suggesting a TGP between 450 and 600 W for the upcoming RTX 4090 / 4080 high-end gaming GPUs, which is still considerably higher than Ampere's initial 350 W TGP.
Nvidia surprised everyone with the 4 nm fabrication process for Hopper, and while this is not yet confirmed anywhere, it is not that far-fetched to assume Lovelace may also be produced on TSMC’s N4 nodes. This could ensure a slight advantage over AMD’s RDNA3 GPUs, which are rumored to be produced on custom TSMC 5 nm nodes, yet MLID believes that the two competing GPU models will still be neck and neck when raw performance is considered.