Cuda cuties — Getting the most of Max-Q Turing: “We want every half-watt savings to drive the GPU.” Sam Machkovech – Apr 2, 2020 7:01 am UTC Here’s a gallery of April 2020 laptops with the latest Nvidia Max-Q GPUs. First up, the Razer Blade Advanced, with an RTX 2080 Super inside. The Acer Nitro…
What kind of GPU year can we expect from Nvidia, one of the two largest consumer-grade GPU producers in the world? The answer is somewhat up in the air, because Nvidia is in a solid-yet-fluid position. Market worries and announcement-filled event cancellations hover on one end, while the company’s surprisingly bullish financial guidance stands out on the other.
Either way, we’ve reached April without the company’s usual announcement of some new desktop hardware by March’s end, and we still don’t know when wholly new desktop GPUs might come (more on that later). Instead, we start this month with a different wave of products: a new slate of laptop-grade GPUs, albeit not that new.
Nvidia has announced a wave of “Max-Q” GPUs coming to laptops from 25 OEMs by the end of April, and most, but not all, come from the company’s RTX line of GPUs. This month’s wave of GPUs consists of three new laptop SKUs (RTX 2080 Super, RTX 2070 Super, GTX 1650 Ti) and slight updates to four existing SKUs (RTX 2070, RTX 2060, GTX 1660 Ti, GTX 1650). Each of these GPUs is built upon the manufacturer’s Turing 12nm architecture.
But as we learned from last year’s “Super” line of desktop GPUs, those updates mostly consist of mild-yet-welcome jumps on nearly identical silicon, which is what we’re seeing in two new Max-Q Super options. Both new Super Max-Q cards benefit the most from an apparent jump in maximum boost clocks, though it remains to be seen how those numbers bear out in the wild. Last year’s cards launched with different maximum wattage counts, and that may be the same this time around as well. And they have very slight bumps in tensor core capacity, which is directed primarily at DirectX-based ray tracing and proprietary Nvidia RTX effects (like DLSS, which we get into later). Otherwise, they both continue to have 8GB of GDDR6 memory with a maximum of 448GB/s memory bandwidth, and they have each received a slight bump in “CUDA core” processing units (up 4.6 percent on the 2080 Super and 11.1 percent on the 2070 Super).
A press-only briefing ahead of the cards’ reveal showed Nvidia coming up with ways to get further gains from these cards, which is why they tooted the horn of a few new features exclusive to this year’s models. The biggest, dubbed “Dynamic Boost,” sees Nvidia partnering with OEMs to manage the shared thermal workload of both the CPU and GPU and to fudge up to 15W of power from one motherboard element to the other based on any application’s live frame-time data. Meaning, if these sensors, as activated on the driver level, detect that either the GPU or CPU has been pushed to its maximum and that the other half could spare some power, the system will then redirect its total wattage accordingly.
“All of this is free”
Nvidia representatives estimate that this system-level tweak will result in a performance boost of roughly 4 to 8 percent in any game. “But all of this is free,” Nvidia’s marketing director Mark Aevermann said in a phone interview with Ars Technica. “[The performance] was sitting there locked away because a controller wasn’t smart enough to do this in the past.” Worth noting, this feature requires total buy-in from an OEM before shipping a laptop with a new Nvidia GPU, owing to the system-level optimizations that Nvidia needs to confirm. So don’t expect it to get patched into existing laptops with older GPUs. Additionally, as of press time, AMD CPUs do not work with Nvidia Dynamic Boost and that support will arrive “shortly thereafter,” Aevermann said.
Aevermann mentioned additional tweaks designed to drive even smaller performance gains, including voltage optimizations for both the GPUs’ GDDR6 memory and for the GPUs’ general voltage regulators. These result in 1-2W of power recovery, Aevermann said, which they admitted was mild but insisted was crucial: “We want every half-watt, quarter-watt, one-watt savings to drive the GPU core.”
Last on the optimization list is a feature that sounds great: “Advanced Optimus.” This is Nvidia’s term for a new connector protocol to guarantee that a laptop’s integrated-graphics solution doesn’t get in the way of an Nvidia Max-Q GPU’s direct connection to a laptop’s display. Aevermann said that this is simply a more streamlined way for OEMs to guarantee screen technologies like variable refresh rates (G-Sync, FreeSync) and frame rates up to 144fps and beyond, even though we’ve seen both in Nvidia-powered laptops from the past two years. But this new toolset may also come with a licensing cost or other issues, because as of press time, only one upcoming laptop, the Lenovo Legion 7, has been confirmed to implement Advanced Optimus, with “more coming later.” (“Advanced Optimus, like Dynamic Boost—we offer it to OEMs, but it’s up to them to implement it or not,” Aevermann said to Ars.)
Nvidia appears to hope that consumers are as excited by these nitty-gritty tweaks as they are by the RTX-exclusive Deep Learning Super Sampling (DLSS) image-processing system receiving a huge update late last month. DLSS 2.0 is indeed an impressive proposition, since its machine-trained anti-aliasing system can upscale 1080p footage to a whopping 4K pixel count with very little in the way of noticeable noise or distortion. So long as developers implement a temporal anti-aliasing system (TAA) in their game, then forward the results to Nvidia, an automatic model can be generated. If more games did so, that’d certainly be good news for anyone seeking value in their pricier RTX GPUs, laptop or otherwise.
Our own anecdotal testing of last week’s DLSS 2.0 patch to the beautiful, ray-tracing-filled game Control revealed that the system mostly works as advertised and the images above aren’t whistling dixie: they’re borne out by our own testing of the same game. But the effect is much better when a 1080p signal is processed to 4K resolution. The DLSS 2.0 up-conversion process continues to struggle with lower resolutions, particularly its attempts to scale 960p resolution to 1440p, so laptop buyers aiming for an RTX system with a 1440p monitor should be mindful. Aevermann said that its DLSS system will continue to be incrementally updated over time, so we hope lower-source resolutions receive greater emphasis going forward, especially for laptop systems that don’t always ship with 4K monitors.
While we were preparing this report, rumors about Nvidia’s next series of Ampere-generation GPUs began picking up steam, with one report suggesting we’ll have to wait until September 2020 to see any “RTX 3000” series cards. Anyone hoping to sink their teeth into interesting new Nvidia hardware will have to settle on this month’s Max-Q line in the meantime. We hope to go hands-on with newly outfitted laptops by this month’s end and will report back with our findings.
Listing image by Razer