NVIDIA GeForce RTX 4090, 4080 And 4070 Alleged Specs Revealed Ahead Of Launch
The GeForce RTX 30 series is already super fast. The higher-end cards in that family chew up and spit out games at high resolutions the way cards from five years ago were just starting to master 1080p. It gets a little bit silly at times when we're benchmarking these graphics cards because it is sometimes difficult to find benchmarks that can really put the hurt on something like a GeForce RTX 3080 Ti—to say nothing of the 3090 or 3090 Ti.
That's why talking about the GeForce RTX 4000 series can get our heads spinning, because these cards are sounding like a monumental leapp ahead of the Ampere generation. Most of what we think we know about the upcoming GeForce generation comes from leakers. We've heard stories of massive GPUs pulling incredible power numbers, and tall tales of up to 48GB of memory, but the reality is probably a little more modest.
The latest leaks from kopite7kimi sound a lot more realistic than some of the stuff we've heard before. In a tweet today, he provided some "updates" on the expected specifications of the GeForce RTX 4090, 4080, and 4070. The tweet includes significant specifications for the upcoming GPUs, and in a follow-up tweet, he provided his expectations for the cards' total board power.
The memory configuration on the RTX 4090 will supposedly be some 24GB of GDDR6X memory running at a blistering 21Gbps on a 384-bit memory bus. This is, notably, the exact same configuration as the GeForce RTX 3090 Ti, which lends some credence to the idea that the two cards (4090 and 3090 Ti) use closely-related PCBs, as has been rumored in the past.
Matching up with earlier rumors, kopite7kimi also puts the RTX 4090 down for a 450W total board power. This coincides with the maximum amount of power that the 12VHPWR connector is allowed to carry without having the extra four sense pins. Almost nobody has a power supply with that capability yet—there's only one on the market—so it seems unlikely we'll see cards exceed 450W nominal power in the short-term.
Stepping down to the GeForce RTX 4080's rumored specifications, we see a much larger leap than in the 3000 series between x80 and x90 cards. The GeForce RTX 4080 is supposedly going to ship with a different processor altogether compared to the top-end card, known as AD103-300. We have seen a "103" GPU before; there's a GA103 that is used in the top-end Ampere mobile GPUs. The Ada Lovelace version will allegedly have 10240 CUDA cores, giving it 80 SMs of a possible 84.
This latest leak pegs the GeForce RTX 4080 with a narrower 256-bit memory interface using 16GB of GDDR6 memory, possibly at 18Gbps. This seems conservative to us; we'd expect the x80 card to ship with GDDR6X, at least, and likely at the higher 21Gbps clock rate. The rumored configuration would end up with less memory bandwidth than even the RTX 3070 Ti, although even 256-bit GDDR6X at 21 GT/s would be behind the RTX 3080 and its much wider bus. It's an interesting rumor for sure; perhaps the much larger L2 cache on Ada Lovelace can make up the difference, or NVIDIA has devised a way to effectively increase memory bandwidth through a new compression technology or Infinity Cache-like feature.
The leaker seems less confident in his power estimate for the GeForce RTX 4080, listing it with a question mark. The estimate is 420W, which isn't all that far from the GeForce RTX 4090. Given the huge difference in allotment of functional units, it's possible that the GeForce RTX 4080 is clocked much higher than its larger sibling to keep them from being too far apart in final performance. That would push it out of the optimal area on the performance/watt curve, potentially making the 4090 a more efficient card.
Finally, for the RTX 4070, the leak predicts that it will use AD104-275 with 7168 CUDA cores in 56 SMs. Based on the leaked specifications for the Ada Lovelace GPUs, that would be the same 4-SM cut from the full-fat AD104 GPU as we saw in AD103 for the RTX 4080. He predicts 10GB of 18Gbps GDDR6 memory on a 160-bit bus for the RTX 4070, another short memory bandwidth allocation putting it on par with the RTX 3060. Seems dubious to us, but again, Ada Lovelace is drastically increasing the on-die GPU cache compared to Ampere. Maybe that makes up the difference.
Unsurprisingly, given that it's a much smaller GPU, the RTX 4070 gets put down for a downright-reasonable 300W TBP. Remember when a 300W graphics card seemed absurd? It wasn't all that long ago that people were making memes comparing the R9 290X's blower to a jet engine.
One sore point of the GeForce RTX family for many PC enthusiasts has been the sharp rise in price at every tier. The GeForce GTX 1080 shipped with a nominal MSRP of $599, which dropped to $499 after nine months. The GeForce RTX 2080 debuted at $699, and the RTX 3080 piled another $100 on top of that.
Unfortunately, it looks like that pricing trend will continue. As the last part of the original tweet states, "do NOT expect a lower MSRP." Hopefully that means that we at least won't see even higher pricing for discrete GPUs, but it does present an opportunity for a competitor (like AMD or perhaps even Intel) to slip in with lower pricing and steal some market share. We'll see what happens later this year when the higher-end Arc cards and AMD's Radeon RX 7000 series hit the market.