Nvidia Announces RTX 3070, 3080, 3090 GPUs: Ampere Crushes Turing
When data leaked on Nvidia’s rumored upcoming GPUs last week, I wasn’t sure what it might imply about Nvidia’s upcoming products and product stack. Now that we have more information on what the upcoming hardware will look like, I can say those fears were unfounded. Nvidia isn’t going to repeat the price-increasing moves it took with Turing — and Ampere sounds like it’ll deliver real gains over Turing (or Pascal).
Let’s go over what we know:
Ampere and the RTX 3000 family pack up to 28B transistors and the cards are built at Samsung, on the company’s 8N process. This is not an EUV node. Samsung’s 8nm is an extension of the company’s 10nm technology. Compared with 10nm, the 8nm node offered either 10 percent improved die area or a 10 percent performance improvement. We can’t do a direct comparison between TSMC 12nm and Samsung 8nm (well, not until we have cards, of course), but Samsung is claiming a full-node improvement.
These cards use GDDR6X, with an improved signaling system that allows for higher clocks and they offer a surprising feature — improved storage I/O (dubbed RTX I/O). Jen-Hsun didn’t spend a lot of time on RTX I/O, but it sounds conceptually very similar to the fast storage performance options we’ve seen Microsoft and Sony talking about for their own upcoming consoles. Don’t be surprised if there’s a lot of common DNA around these ideas.
The RTX 3080 can perform 2x the ray/triangle intersection calculations per clock compared with Turing, with 58 RT TFLOPS (these numbers are vague enough to not be worth comparing much, but more does generally equal better) and the ability to leverage concurrent RT + graphics or concurrent RT + compute.
Overall performance improvements are expected to put the RTX 3070 on the level of the RTX 2080 Ti. Companies sometimes play fast and loose with their performance claims, so always keep that in mind, but that’s an effective 1.51x improvement in performance for $500. This is exactly the kind of realignment gamers have wanted to see.
The RTX 3080 is a $600 card with up to 2x the performance of the RTX 2080. “Up to” is suspicious phrasing, here, but we can probably assume Nvidia will keep a 1.2x – 1.3x performance differentiation between the two GPUs. If the RTX 3070 can match an RTX 2080 Ti, the RTX 3080 will decisively outperform it, for just $600.
When you factor in DLSS improvements, the gains are even larger. DLSS has improved substantially since Turing debuted, and while it isn’t available for every game, the performance improvements of using it are very real. DLSS 2.0 is higher quality than DLSS 1.0, and as the feature continues to evolve the benefits of using it are increasing.
Meanwhile, there’s also the RTX 3090 — a GPU Nvidia is claiming can run at 60fps, even in 8K.
At $1,500, the RTX 3090 isn’t cheap — but Nvidia’s premium GPU pricing has never really been the issue we took with its product positioning. During the Turing launch, it was the price increases lower down the stack that rankled, and Nvidia has solved that problem by picking more aggressive launch targets.
There’s a lot more to talk about with the RTX 3000 series, including some of the new software capabilities and the cooling system. Look for that content in a bit. Overall, my initial impression of the positioning on these GPUs is very positive — Nvidia clearly is bracing for a fight when RDNA2 launches, and the claimed 1.9x performance improvement over Turing is going to be sharply compared and contrasted against AMD’s claimed 1.5x further improvement over RDNA. On paper, Nvidia now leads if you apply these claims against existing GPU power draw, but broad targets can’t compete with actual measurements where efficiency is concerned.
We don’t have to wait long to find out: The RTX 3080 debuts on September 17, while the RTX 3090 will launch on September 24. The RTX 3070 is expected to be available in October.