We’ve been treated to a fresh update on higher-end Lovelace graphics cards from a regular source of Nvidia leaks on Twitter, who provided a few more specifics, including some eye-opening info on the RTX 4090.
This comes from Kopite7kimi, who gave us a whole bunch of details on these incoming GPUs only a week and a half ago.
Continue to update. RTX 4090, AD102-300-A1, 16384FP32, 384bit 21Gbps 24G GDDR6X, 450W, base 2235 boost 2520 actual max >2750; RTX 4080, AD103-300-A1, 10240FP32, 256bit 21Gbps 16G GDDR6X, 420W, RTX 4070, AD104-275-Kx(x is a number)-A1, 7168FP32, 160bit 18Gbps 10G GDDR6, 300W.July 4, 2022
As you might imagine, in such a short space of time, a lot of the purported specs spilled have stayed the same as in the previous leak from Kopite7kimi. But there are a few key differences here, notably some nippy-sounding clock speeds for the flagship Lovelace graphics card (building on what the leaker recently hinted at).
The speed that the RTX 4090 runs at is set to be a base clock of 2235MHz with a boost to 2520MHz, although the actual maximum that overclocking could reach is in excess of 2750MHz, the leaker believes (and it is just that – a belief in what they’ve been told by sources – so let’s not read too much into any of this just yet).
Kopite7kimi also pins a power usage of 450W on the Lovelace flagship, which matches what we’ve heard previously. The RTX 4090 will be a cut-down AD102 chip, of course, and power usage could be pushed a lot further (like to 600W) with the full-fat version (that could be a 4090 Ti, or maybe even a new RTX Titan).
The info on the RTX 4080 is again pretty much the same story, except now Kopite7kimi thinks that the 16GB of VRAM will be GDDR6X – the leaker wasn’t sure on this last time – and the prediction of that RAM being 18Gbps has been revised upwards to 21Gbps, so that’s some good (theoretical) news.
Power usage is pegged at 420W, which is again repeating what the leaker said a month ago, albeit they added a question mark to that wattage back then – so this seems to indicate that Kopite7kimi is becoming more confident that this is the level the RTX 4080 could be pitched at.
Finally, regarding the RTX 4070, there’s nothing new in terms of the spec since Kopite7kimi’s last major update. A 300W power usage is again mentioned, which has been the leaker’s best guess for some time now. Indeed, Kopite7kimi first floated that figure back in April, before even having much of a clue about the RTX 4080’s power demands.
Analysis: A Lovelace flagship that’s twice as powerful as the RTX 3090?
What’s quite telling here is that some of the power usage figures given appear to have been maintained as predictions for some time now, which suggests that these might just be set in stone, as it were. Or at least they could now represent a more likely correct estimate of where Nvidia will end up with TGPs for RTX 4000 graphics cards.
None of the info provided here is much of a surprise, then, and it’s more about consistency than fresh revelations – with the exception of those specific (purported) clock speeds for the RTX 4090. With 2750MHz plus being mentioned, Nvidia seems to be looking towards the 3GHz mark for the flagship, as opposed to the far more conservative clocks seen on current-gen Ampere (the 3090 has a boost speed of 1.7GHz, as officially rated by Nvidia – though it can, of course, be pushed faster – and even the Ti version only hits 1.86GHz).
And with the RTX 4090 loaded up with CUDA Cores as it is – 16,384 of them – coupled with that theoretical 50% increase in boost speed, this points towards a pretty powerful next-gen flagship. In fact, it could be twice as powerful as the RTX 3090, which is a performance leap we’ve seen predicted before, as you may recall.
While nothing is mentioned about other clock speeds here, looking at the relative CUDA Core counts, and the RTX 4080 coming pretty close to the RTX 4090 for TGP (420W versus 450W), it’s easy to imagine that the 4080 will push harder still with its clock speeds.
It makes sense that Nvidia would want to be driving as hard as it can to get the best gains possible from the advantageous drop to 5nm architecture with Lovelace, and indeed with the flagship, there’ll be plenty of room to build something even beefier than the initial 4090 (remember, we’re looking more at 600W for the 4090 Ti, as per the rumor mill).
Finally, it’s worth noting that even though it’s not new info, there’s still disappointment being aired online around the purported memory loadout of the RTX 4070, and that it might only be 10GB of VRAM. Previous rumors suggested a more palatable 12GB, but it seems that this configuration might be reserved for a 4070 Ti model – take all of this with the appropriate helping of salt, naturally.