- Joined
- Dec 16, 2014
New Nvidia cards are coming out soon and ray tracing is the big new thing. They're also really fucking expensive holy shit.Nvidia will launch its next-generation GeForce RTX 20-series graphics cards on September 20, starting with the RTX 2080 and RTX 2080 Ti, followed by the RTX 2070 in October. These will use the new Turing architecture, which boasts more cores than the previous generation Pascal architecture, along with significant updates that should deliver more performance per core. Also included are new technologies to enable real-time ray tracing in games, and deep learning Tensor cores.
There's a lot to cover, and Nvidia is rightly calling this the most significant generational upgrade to its GPU since the first CUDA cores in 2006. Turing promises better performance than existing GPUs, and has the potential to fundamentally change what we expect from graphics. Here's everything you need to know about the RTX 2080 Ti, RTX 2080, and RTX 2070, the Turing architecture, pricing, specs, and more.
Pricing and release dates for the GeForce RTX series
Nvidia has only announced three GeForce RTX models so far. We don't know when or even if lower tier cards will exist. Most likely, but they may not arrive until 2019. Here are the launch dates and prices so far:
GeForce RTX 2080 Ti Founders Edition: $1,199, September 20
GeForce RTX 2080 Ti Reference: $999, September 20?
GeForce RTX 2080 Founders Edition: $799, September 20
GeForce RTX 2080 Reference: $699, September 20?
GeForce RTX 2070 Founders Edition: $599, 'October'
GeForce RTX 2070 Reference: $499, 'October'?
It's not all good news for the RTX 20-series, as pricing for all three classes of GPU has increased substantially. Call it a lack of competition (AMD's GPUs already struggle to compete against the 10-series parts), or the cryptocurrency bubble bursting (there are reportedly a lot of 10-series graphics cards left to sell), or just plain greed. The bottom line is that launch prices on the Founders Edition cards are up to 50 percent higher than the outgoing 10-series parts.
Pre-orders are available, and while we don't generally recommend buying expensive hardware before independent reviews have been published, many places offering pre-orders are currently sold out. What's worse is we don't even know if the lower 'reference' prices will be seen at launch, or if they're merely recommendations. Based on past experience, we expect Founders Edition and factory overclocked cards priced similarly to the FE to be the main option for the first month or two.
The RTX 2070 launch date hasn't been firmly set by Nvidia yet, with only a statement of October 2018. Given the likely demand for the higher end 2080 parts, we anticipate late October. Again, prices will probably be higher for the first month or two. Then again, with Black Friday and the holiday shopping season going on, we might get a few surprises.
GeForce RTX specifications
Nvidia unveiled many core details of the Turing architecture at SIGGRAPH, and followed up by announcing the below specs for the GeForce RTX graphics cards. After much speculation, we now know what to expect. Mostly.
The number of CUDA cores in each model has increased by 15-20 percent across the line, though clockspeeds have dropped slightly as well. In theoretical TFLOPS (that's trillions of floating-point operations per second), the GeForce RTX cards are 14-19 percent faster than the GTX 10-series.
Nvidia equips all the new models with 14 GT/s GDDR6, improving bandwidth by anywhere from 27 percent (RTX 2080 Ti) to as much as 75 percent (RTX 2070). That's assuming there aren't any other tweaks to the memory subsystem, like the improved compression technologies and tiled rendering in Pascal.
Along with faster cores and memory, the Turing architecture adds Tensor cores for deep learning and RT cores for real-time ray tracing. Both have the potential to dramatically change what we can expect from future games in terms of graphics.
Turing architecture and performance expectations
While we have the numbers for the CUDA cores, GDDR6, Tensor cores, and RT cores, there's a lot more going on with the GeForce RTX and Turing architecture. We've provided a deep dive into the Turing architecture elsewhere, which we'll update with additional details closer to launch, but here's the short summary.
Nvidia has reworked the SMs (streaming multiprocessors) and trimmed things down from 128 CUDA cores per SM to 64 CUDA cores. The Pascal GP100 and Volta GV100 also use 64 CUDA cores per SM, so Nvidia has now standardized on a new ratio of CUDA cores per SM. Each SM now includes eight Tensor cores and an unspecified number of RT cores, plus texturing units (which we assume to be half as many as in Pascal). The SM is the fundamental building block for Turing, and can be replicated as needed.
For traditional games, the CUDA cores are the heart of the Turing architecture. Nvidia has made at least one big change relative to Pascal, with each SM able to simultaneously issue both floating-point (FP) and integer (INT) operations—and likely Tensor and RT operations as well. Nvidia said this makes the new CUDA cores "1.5 times faster" than the previous generation.
That might be marketing, but Nvidia's preview benchmarks suggest an average performance increase of around 50 percent for the RTX 2080 over the GTX 1080. Combined with the increase in CUDA core counts and the higher bandwidth of GDDR6, in GPU-limited benchmarks it's not unreasonable to expect 50-75 percent more performance from the GeForce RTX models compared to the previous generation parts.
All Turing GPUs announced so far will be manufactured using TSMC's 12nm FinFET process. The TU102 used in the RTX 2080 Ti has 18.6 billion transistors and measures 754mm2. That's a huge chip, far larger than the GP102 used in the GTX 1080 Ti (471mm2 and 11.8 billion transistors) and only slightly smaller than the Volta GV100. While the full TU102 has up to 72 SMs and a 384-bit GDDR6 interface, the RTX 2080 Ti disables four SMs and one of the 32-bit GDDR6 channels. That leaves room for a future RTX Titan, naturally.
The TU104 trims the SM counts and memory interface by a third, giving a maximum of 48 SMs and a 256-bit interface. The RTX 2080 disables two SMs while the RTX 2070 disables 12 SMs, but both keep the full 256-bit GDDR6 configuration. Nvidia has not revealed die size or transistor count for the TU104, but it should fall in the 500-550 mm2 range, with around 12-13 billion transistors. Again, that's a substantially larger chip than the GP104 used in the GTX 1080/1070.
TSMC's 12nm process is a refinement of the existing 16nm process, perhaps more marketing than a true die shrink. Optimizations to the process technology help improve clockspeeds, chip density, and power use—the holy trinity of faster, smaller, and cooler running chips. TSMC's 12nm FinFET process is also mature at this point, with good yields, allowing Nvidia to create such large GPU designs.
Looking forward, TSMC is readying its 7nm process for full production, and we should see it in a limited fashion by the end of the year (eg, for AMD's Vega 7nm professional GPUs). Don't be surprised if late 2019 sees the introduction of a die shrink of Turing, bringing sizes down to more manageable levels.
What the RT cores and ray-tracing mean for games
Why is ray-tracing such a big deal, and what does it mean for games? We wrote this primer on ray-tracing when Microsoft unveiled its DirectX Ray Tracing (DXR) API. DXR hasn't reached the final public revision yet, but that's expected to happen around the time GeForce RTX cards begin shipping. Nvidia clearly had a lot of input on DXR, and while initial demonstrations like the above Star Wars clip used a DGX-Station with four GV100 GPUs to achieve 'cinematic' 24fps results, Turing is clearly what Nvidia was aiming for.
Not only can a single Turing GPU run the same demonstration as the DGX-Station—which only costs $60,000 if you're wondering—but it can do so at 60fps. That's because the RT cores in Turing are roughly ten times faster for ray tracing than using compute shaders to accomplish the same work. However, doing full ray tracing for real-time games is still a bit impractical.
Nvidia instead suggests using the RT cores for hybrid rendering. Traditional rasterization used for geometry and textures, while ray tracing provides lighting and shadows, reflections, ambient occlusion, and other effects. At least 11 games have announced support for Nvidia's RTX ray tracing. Here's the current list:
Assetto Corsa Competizione from Kunos Simulazioni/505 Games
Atomic Heart from Mundfish
Battlefield V from EA/DICE
Control from Remedy Entertainment/505 Games
Enlisted from Gaijin Entertainment/Darkflow Software
Justice from NetEase
JX3 from Kingsoft
MechWarrior 5: Mercenaries from Piranha Games
Metro Exodus from 4A Games
ProjectDH from Nexon’s devCAT Studio
Shadow of the Tomb Raider from Square Enix
Several of these games should release in 2018, while others are coming in 2019. Shadow of the Tomb Raider will apparently launch without RTX effects enabled, with a post-launch patch adding the feature. Given the September 14 release date, one week before the official launch of the GeForce RTX cards, plus waiting for Windows 10 Redstone 5 and the full DXR API (currently prepping for arrival in September or October), that shouldn't be much of a concern. Getting games that support brand new hardware features within weeks of the hardware launch is still much faster than the usual rate of adoption.
How machine learning and the Tensor cores affect graphics
If you're thinking the Tensor cores are pointless when it comes to 'real' graphics work, you're wrong. Deep learning and AI are revolutionizing many industries, and games are another potential market. But how can the Tensor cores help with graphics?
Nvidia has specifically talked about DLSS, Deep Learning Super Sampling, a new AI-based anti-aliasing algorithm that can offer improved image quality compared to other AA algorithms like TAA (Temporal Anti-Aliasing). The idea is to train a neural network with high quality AA images as the 'ground truth' model—the desired result. Once trained, DLSS can provide real-time enhancements like the removal of jaggies, plus it can also combine resolution upscaling with anti-aliasing. Nvidia hasn't fully disclosed how DLSS is being implemented, but upscaling 1080p to 4k seems likely.
GeForce RTX 2080 Ti Founders Edition: $1,199
GeForce RTX 2080 Ti Reference: $999,
GeForce RTX 2080 Founders Edition: $799
GeForce RTX 2080 Reference: $699
GeForce RTX 2070 Founders Edition: $599,
GeForce RTX 2070 Reference: $499,
GeForce RTX 2080 Ti Reference: $999,
GeForce RTX 2080 Founders Edition: $799
GeForce RTX 2080 Reference: $699
GeForce RTX 2070 Founders Edition: $599,
GeForce RTX 2070 Reference: $499,
