GPU leaker @Kopite7kimi (opens in new tab) is at it again today with another update on the upcoming RTX 4080 specifications. His latest rumors include a core count reduction for the RTX 4080 from the estimated 10,240 to 10,752 range down to 9728. This translates to a 10% reduction in core count.
Thankfully, this change doesn't appear to change the 4080's estimated Time Spy Extreme score of 15,000, which Kopite lists again in his latest Tweet. Performance appears to remain the same, for now at least. Kopite's other indicated specs for the RTX 4080 also remain intact, including the 420W power limit, 16GB of 21Gbps GDDR6X memory, and AD103 die.
Regardless of the source, apply a healthy dose of skepticism to these rumors and "leaks." Mistakes and incorrect guesses are inevitable, and the hardware specs might not be fully nailed down yet. One critical missing piece of information are the RTX 40-series core clocks. There's speculation, based on the higher power levels, that Nvidia will increase core clocks substantially relative to the existing Ampere line, but there's no hard data as yet.
I'm not a chatterbox, but I have to make some updates. I hope you don't mind. a possible RTX 4080, PG136/139-SKU360AD103-300-A19728FP32256bit 16G 21Gbps GDDR6Xtotal power ~420WTSE ~15000Now I have completed the latest update for 4090, 4080 and 4070.August 5, 2022
RTX 40-Series Titan Rumor and L2 Cache Updates
Another Tweet from @kopite7kimi suggests the Titan branding may be coming back to the GeForce lineup in some shape or form. This RTX 40-Series Titan "has a chance" at using the full-fat AD102 chip, but Kopite hasn't confirmed whether or not that will actually happen. More likely is it will be a nearly complete AD102, since 100% functional chip yields on the presumably large GPU will likely be low.
There's an interesting question if we do in fact get a new Titan card. Will Nvidia once again have a Titan "hybrid" SKU that blurs the lines between Nvidia's A-series workstation GPUs and the company's GeForce gaming GPU lineup, with lots of memory and enhanced drivers? Or will we instead get an RTX 4090 Ti that skips the professional driver enhancements? Either way, we can safely guess a card with 48GB of GDDR6X memory won't come cheap. Kopite7kimi appears to have provided an educated guess for now is all.
In the same Tweet, Kopite also shared some L2 cache updates for the AD103 and AD104 40 series dies. The fully unlocked AD103 die will apparently have 64MB of L2 cache, while the AD104 die will have 48MB — that's the same as what the earlier leaks suggested back at the start of the year.
However, Kopite also suggests the RTX 4090 might have a reduced L2 cache size, or perhaps that AD102 won't have proportionately larger caches. Or it might be a move to differentiate the potential RTX 4090 Ti (or Titan) from the "lesser" RTX 4090.
1. The TITAN have a chance to use a full-fat chip. I will update that when I confirm it. 2. A full-fat AD103 will have 64M L2, and AD104 will have 48M. Only RTX 4090 should have a L2 cut. (99%😁)August 5, 2022