Nvidia’s GTX 1080 redefines high-end gaming performance

Nvidia formally announced the GTX 1080 in Austin just over a week ago, but it held back on the GPU deep dive at the initial presentation. While the GTX 1080 isn’t scheduled to launch until May 27, Nvidia has lifted the curtain on the GPUs performance and technical advances. We’ve covered some of the card’s improvements in VR and overall positioning, plus new technologies like Ansel that give gamers more artistic freedom, so we’ll be focusing on areas where Nvidia hadn’t disclosed as much information. Unfortunately, Nvidia was unable to supply us with a GTX 1080 in time for launch, so we’ll have to defer benchmarks and performance comparisons for another day.

Pascal is an evolutionary step forward from Maxwell and many of the technologies that debuted in the GeForce 9xx family have been refined, improved, and enhanced for the GTX 1080. The base GPU packs 2560 cores with 20 SM blocks and 128 cores per block. There are 160 texture units and 64 ROPS, which is interesting — one common theory was that AMD and Nvidia would both ship substantially more ROPs this year to ensure that they didn’t become fill-rate limited in VR or at higher resolutions. Nvidia chose to deal with this another way, which we’ll explore shortly.

The GTX 1080 packs 25% more cores and 25% more texture units than the GTX 980 it replaces, along with a much higher base clock (1.61GHz vs. 1.1GHz for Maxwell) and significantly faster RAM (320GB/s of memory bandwidth, compared to 224GB/s for Maxwell). On paper, the 1080 looks much more like the 980 Ti. In practice, it often outperforms that card.

Memory bandwidth, memory compression
Pascal uses GDDR5X to increase its memory bandwidth, but Nvidia chose to stick with a 256-bit memory bus rather than the 384-bit bus that the GTX 980 Ti and GTX Titan X use. The secret sauce in Nvidia’s recipe? A more advanced version of the same delta color compression techniques that Maxwell used.

When Nvidia launched Maxwell, it claimed that Maxwell reduced its memory bandwidth needs by 19-30% over Kepler, depending on the game in question. Pascal delivers a further 11-28% improvement over Maxwell, again depending on the game in question. Considering that the GTX 1080 already offers 43% more bandwidth than the GTX 980, the added color compression is icing on the cake — or a smart way to ensure the card can scale to 4K or even beyond, depending on your point of view.

The outrageously pink image above shows the difference between Maxwell and Pascal’s color compression. You can see that Maxwell already does a pretty good job of compressing the frame, but there are still a significant number of areas where Maxwell couldn’t compress the data.

Here’s Pascal’s version of the same frame. If you’re thinking “Wow, that’s really pink,” you’re on the right track. What this means is that Pascal can extract memory compression savings from a much larger percentage of the frame than Maxwell could. In the long run there’s an inevitable diminishing marginal return to saving bandwidth via compression, but the feature worked extremely well in Maxwell and should give Pascal an additional edge in 4K gaming.

Display support, media encode/decode, and SLI bridges
Nvidia has announced that the GTX 1080 will support a maximum resolution of 7680×4320 at 60Hz if two DisplayPort 1.3 connectors are used to drive the display. The GPU is only certified for DP 1.2 but is listed as DP 1.3 and 1.4 “ready.” HDMI 2.0b and HDCP 2.2 are both supported as well. Media standard support has a few new bells and whistles that previous Maxwell cards lacked. Pascal now supports full encode and decode in both H.265 and 10-bit H.265. 12-bit (decode-only) is also supported, as is hardware decode for Google’s VP-9 codec.

Those of you who are familiar with multi-GPU configurations are also aware that Nvidia has previously used SLI bridges to connect one GPU to another. AMD abandoned this approach back in 2013 when it launched Hawaii; AMD GPUs now connect directly over the  

PCI Express 3.0 bus. Nvidia is still sticking with bridges for Pascal and GP104, but this time it’s introducing a new, higher-bandwidth bridge standard for modern GPUs. Existing bridges should function well up to 2560×1440 @ 60Hz, but if you want >60Hz refresh rates or to run SLI in 4K or 5K mode, you’ll see top performance if you use newer bridges (Nvidia did note that its LED bridges are still rated for anything up to 5K). It’s not entirely clear if older “stiff” bridges are limited the same way as the older “floppy” bridges (cross-GPU bandwidth was lower on the flexible bridges than on their “stiff” counterparts.)

This slide shows the difference in Shadows of Mordor between the old and new bridge. It’s important to note that this dramatic difference was captured in 4K Surround mode, which means three 4K displays running the same game for a total resolution of 15360×2160. While the new bridges are much smoother than the old ones, the game itself doesn’t maintain a playable frame rate at these resolutions and detail settings. Lower resolutions and detail settings might not show the same gains.

Asynchronous compute
Asynchronous compute has been a hotly debated topic ever since Ashes of the Singularity debuted and showed AMD holding an advantage over Nvidia, ostensibly due to this particular capability. While that situation is rather more nuanced and game-specific, there are going to be a number of questions regarding how Pascal stacks up to the competition.

According to Nvidia, GP104 improves on Maxwell in some significant ways. Maxwell was only capable of performing draw-level preemption and could only switch to a different workload at a draw call boundary. What this meant practically was that there were significant penalties to running a mixed compute + graphics workload, and we saw that reflected in Maxwell’s performance when significant asynchronous workloads were running.

Unlike Maxwell, Pascal can perform much finer-grained preemption. In graphics workloads, it can preempt at the pixel level, flush the shader pipeline, and switch to compute. In compute workloads it can swap at the instruction level and return to doing graphics work. Nvidia claims that this takes 100 microseconds or less, and while the company didn’t offer competing figures for Maxwell, it should be significantly faster than what we saw last generation.

Asynchronous compute isn’t a feature most games rely on yet (Ashes of the Singularity is something of an exception), and we can’t deep dive into the question until we’ve got hardware. What I can say is that while Pascal significantly improves on Maxwell’s capabilities, it doesn’t offer the same set of compute capabilities that GCN does. The larger question is whether or not the difference between what the two companies support will have an impact on future DirectX 12 titles. The fact that Nvidia holds an estimated 75-80% of the gaming market is itself a powerful argument that developers should focus on building engines that cater to Nvidia’s architectures and GPU capabilities more so than AMD’s. At the same time, however, some developers have predicted that game engines may shift workloads towards compute engines no matter what — and that could potentially work in AMD’s favor in future DX12 titles.

I’ve always recommended evaluating GPUs based on the games you’re playing now, not the titles you might be playing in 12-24 months, and it’s difficult to predict how game engines might change in the next few years. At minimum, the changes Nvidia has made to Pascal should significantly reduce any asynchronous compute penalty relative to Maxwell. At best, we should see Pascal picking up some performance improvements from async compute, including in games where Maxwell took a performance hit.

The graph above shows how far the GTX 1080 has come relative to its predecessor, but there’s one caveat worth mentioning. According to Oxide, asynchronous compute is disabled on Nvidi cards by default, which means these test results may not tell us if Pascal actually benefits from async compute just yet. One final note: When Nvidia demoed its async compute capability at Austin, it did so using DirectX 11, not DX12. We weren’t able to discover more information on why it chose to demo using the older API, or what the performance ramifications were for that scenario.

Wrapping it all up
Pascal is a significant leap forward for Nvidia, thanks to a combination of higher clocks, increased core counts, and improved efficiency. The company is forecasting significant gains over and above GTX 980 in both traditional gaming and VR, with particularly impressive boosts arriving for VR titles. While we can’t speak to that specifically just yet, the on-paper gains are substantial.

One difference about this launch, however, is that AMD and Nvidia are taking very different approaches to the market. Nvidia has chosen to launch high-end parts first, with the GTX 1080 and 1070 taking over for the 980 Ti, 980, and GTX 970. AMD, in contrast, will launch an efficiency-focused GPU first, with Polaris 10 and 11 targeting the budget and mainstream segments in both mobile and desktop. This is the first time in a long while that the two companies have taken this approach, and it’ll be interesting to see how they compare in their respective brackets. It’s still not clear if Pascal’s VR performance gains will require substantial optimization or not, but VR enthusiasts who held off buying a new GPU when Oculus and Vive launched should be well-rewarded for their patience.

Current reviews show the GTX 1080 outperforming the GTX 980 by 25-35%, which is in-line with our expectations. This launch is going to put serious pressure on AMD to reduce the price of its Fury products — the non-Founders variant of the 1080 will sell for $500, which puts it head-to-head against Fury and Nano.

Read More

That mighty thud was CERN dropping 300TB of raw collider data to the Internet

 

 

Most of what CERN does sounds like the rarefied heights of sci-fi, accessible only to physicists with badges and pocket protectors, or academics who use esoteric software — not to mere mortals like you and me. What even happens in an atom smasher? CERN is hunting answers to big questions like what dark matter is and why there’s so much of it, why the fundamental forces seem to merge into one at extreme temperatures, why gravity behaves as it does, and other big-topic questions that seem to have one foot firmly in the realm of philosophy.

The LHC has been generating huge amounts of data for release to the general public since its first successful run in 2010. Continuing this trend, CERN just put out another big chunk of data for public analysis, some 300TB of partially organized results from the LHC’s operations since 2014 (when they did their first big data dump). CERN hopes to engage the curiosity of physicists around the world, whether amateur, academic or professional, and get them learning about particle physics and doing hands-on data analysis from their experiments. Many of the LHC’s current experiments and projects have a crowd-sourced component, relying on distributed computing like folding@home or SETI@home do. There are several experimental collider datasets on the CERN open data site that anyone can download. The LHC@home springboard page provides an overview of the distributed computing projects the LHC is currently involved in, including the LHCb, ATLAS, and ALICE.

The ATLAS project is probing for fundamental particles like the Higgs Boson, as well as looking for information about dark matter and extra dimensions. It records the path, energy and identity of particles traveling through the collider and then performs offline event reconstruction based on the data banked from the ATLAS detectors. This turns the raw stream of numbers into recognizable things like photons and leptons so that they can be analyzed. (If you’re a little rusty on your quantum chromodynamics, CERN put out a PDF primer about the LHC that should help to get you up to speed.) Since the ATLAS detectors create petabytes of data during each experiment, the project needs a substantial amount of computational muscle to run reconstructions and, in their words, extract physics from the data. They specifically call out to grad students — for physics majors trying to come up with a master’s project, it might not hurt to get involved with ATLAS.

ALICE is a different ball of wax. Where the ATLAS experiment deals with colliding protons and tries to identify particles, the ALICE detector studies conditions like those found just after the Big Bang. For part of each operating year, the LHC fires lead ions instead of protons, creating conditions so extreme that protons and neutrons in the lead nuclei can “melt,” freeing constituent quarks and gluons from their mutual bond and creating a quark-gluon plasma. It is the ALICE project’s mission to explore what happens to particles under these extreme conditions, leading to insights on the nature of matter and the birth of the universe.

Prototyping is in process for the High-Luminosity upgrade to the LHC, which will create a sort of broadband production of data so that particle physics, which relies on statistics, can prod the Standard Model faster. This is a massive, profoundly collaborative project. The upgrade itself relies on innovations in several fields including magnetics, optics and superconductors. It’ll have a set of shiny quadrupole superconducting niobium-tin magnets, used to better focus the particle beam, and new high-temperature superconducting electrical lines capable of supporting currents of record intensities.

Once they fire up the HL-LHC, the volume of data produced will make 300TB look as small as the particles they’re accelerating. It’s expected to start operation about 2025, but we’ll be watching their updates the whole way.

Read More

Volvo’s truly keyless entry: your smartphone

• y Bill Howard on March 15, 2016 at 11:27 am
• 2 Comments

Read More

How Google’s DeepMind beat the best in Go, and why that matters

 

 South Korea just finished hosting an historic tournament of the ancient game of Go. On one side, Google DeepMind’s AlphaGo program. On the other, South Korean Go champion Lee Sedol, widely considered one of the best players of all time. The contest was to be held over five days, with one match per day, but only three days were needed. AlphaGo won in three resounding victories, each more dominant than the last, and then went on to win 4-1 total. Google DeepMind’s achievement has been called bigger than IBM’s 1997 win against chess champion Garry Kasparov — just like Kasparov, Sedol went in predicting a clean sweep for the human side, but unlike Kasparov, ended up getting almost swept himself.
To be fair to the now-defunct human masters, Go is a game that seems to truly stump humans as well. It originated in China at least several thousand years ago, and while the exact story behind it isn’t known, its mechanics seem strikingly similar to medieval battlefield tactics. The original Chinese name literally means “encircling game,” and it involves laying your pieces so you trap your opponent’s pieces entirely — while, of course, trying to prevent that from happening to your own.

A fancy Go board, with a few pieces laid out.

Rather than shaving down the possibility space for moves, for instance by having a limited number of very specific moves associated with each piece, as in chess, a Go board represents a wide open field of possibility. It’s much larger than a chess board, 19 by 19 spaces according to tournament rules; the combination of a large number of spaces and non-restrictive rules governing moves in those spaces means that there are a uselessly large number of total possible moves. There are more possible positions for all stones on a Go board than there are atoms in the universe — though I don’t think this oft-cited figure includes WIMPs!

Computers have classically worked by quick-searching large numbers of possibilities, so if the number of possibilities becomes unsearchably large, as it does in Go, then a classical computer ought to be fundamentally incapable of playing well. But that’s just the first problem, because unlike chess, Go is all about attrition.

Chess Grandmaster Gerry Kasparov despairs in the face of IBM’s early attempt at machine learning — complete with Japanese surrogate to make the computer’s moves, just like today.

On a Go board, you have to constantly work toward both territory-right-now and a more abstract concept referred to as position. It’s all about boxing out your opponent, opening up territory for yourself while cutting them off from the best opportunities for future play. How do you tell whether this position has better long term potential than that one, when your ability to predict specific outcomes hits a mathematical wall a dozen or so moves into the future?

The moves from Game 1 of 5, March 8.

To answer these questions, the team at DeepMind created two basic components to the system. One neural network tries to get a feel for the overall — how is the board laid out, who’s in a better position, and what areas needs to be threatened and defended? This is the source of much of the difficulty, both for computers and humans, and with human players it’s where we most commonly run into annoying words like “hunch” and “intuition.”

The other major component uses knowledge about past games of Go to look a certain number of moves into the future, given different possible decisions for its own move — each possible path is considered for likely counter-moves, and good responses to each of these, and probable counter-moves to those… and on and on. On a 19-by-19 board, this sort of process needs to be directed by the high-level game- and board-awareness of the first algorithm, or else flail about with poor results.

The overall process is not unlike imagination: The system essentially plays out little mini games as far as it can, due to computational limitations, looks around in this imaginary space, and judges the desirability of this possible path versus all the other imaginary paths it has tested. That’s roughly similar to how some human beings describe their approach to the game. One DeepMind engineer described it in a live interview aired during Game 2 that the heatmap of interesting moves AlphaGo creates of any given snapshot is much like human intuition. “As a Go player, if I look at the board, there are immediately some points that jump out at me as moves that I might want to play. AlphaGo does the same thing.”

Perhaps that’s why even human masters often refer to a “feeling” when asked precisely why they choose one strategy over another — they no more “understand” the truly deterministic outcomes of their decisions than AlphaGo itself.

So, do human Go masters truly understand it? And if so, does that mean AlphaGo does, too? Sedol was surprised to have to deal with a number of moves in the first game (in particular, move 102 to the right, if you’re interested) and seemed to come to think of AlphaGo as a truly unpredictable and dangerous opponent. The program seemed to display genuine creativity — an illusion of course, but an important one.



In the second game, played on March 9, AlphaGo took an unusual opening approach, causing Lee pause. In the first game AlphaGo used far more of its allotted time than Lee, but in the second the human champion used up far more of his time than the computer. He ran out of time first, and continued hesitation brought him down to his final one-minute allotment of overtime before forfeiting due to timing out. It’s incredible — in a very real way, the computer psyched Lee out. What’s more incredible is that it’s actually possible that those early moves were made for specifically that reason. During that game Michael Redmond, the only Westerner to ever reach the top “9-dan” rank in Go, said that this latest version of AlphaGo plays an innovative game. He said its style was “already something I could learn from.”

At the end of the day, the fall-back explanation for human skill at Go is experience — we think it will work because it has worked in the past, even in a very abstract, “This move sort of looks like that move from six months ago!” kind of way. That’s precisely how a deep neural network learns. The difference is that while a human might be able to play at most a few thousand games a year, AlphaGo can play millions every day. That’s how it has acquired such astounding ability, fully a decade before even optimistic projections. By combining a computer’s brute force ability with a truly novel network-of-networks approach to breaking down the possibility space much the way a human mind does, DeepMind has been able to (at least) rival the best humanity has to offer.

This is an incredible achievement, regardless of whether AlphaGo had won the tournament overall. It shows that neural networks really can take us past previously impassable barriers. All Skynet nightmares aside, in the end Alphabet CEO Eric Schmidt said it best: “The winner here, no matter what happens, is humanity.”

Read More

AMD unveils 2016 GPU roadmap; Polaris offers 2.5x performance per watt, may utilize GDDR5

 

 

 At its Capsaicin event yesterday, AMD didn’t just unveil its $1,500 Fiji GPU — it also showed off its next-generation Polaris GPU and gave an update on its longer-term GPU roadmap. The new roadmap isn’t just significant for what it says, but what it doesn’t mention as well.

First, a bit of background. 14nm is expected to be a major new node for both AMD and Nvidia, but it’s particularly important to AMD’s situation. Not only is Team Red fighting to regain market share lost to Nvidia over the last few years, its GPUs have been less power-efficient and noisier than Team Green’s counterparts. (The Radeon Nano, it must be noted, was an exception to these trends, with markedly better power efficiency than any previous AMD GCN GPU.) This is particularly problematic in mobile, where every watt and decibel count.

Radeon Nano power efficiency at default and 150% board power level. The Nano was the first high-end GCN card to land within spitting distance of Maxwell.

Previously, AMD had claimed that it would deliver a 2x performance-per-watt improvement with Polaris as compared with its 28nm-class hardware. Now, the company has bumped that prediction to 2.5x improved performance-per-watt. Readers should keep in mind that like all metrics, performance-per-watt is not an absolute. Because silicon power consumption is not a simple linear curve, these figures are likely best-case estimates based on midrange hardware, not the worst-case scenario when comparing top-end parts clocked at maximum frequency. Even so, a 2.5x efficiency bump is a noteworthy generational leap.

AMD’s GPU roadmap

AMD’s GPU roadmap for the next 18 months is shown below:

AMD’s GPU roadmap. Navi is a star, not a planet of blue cat people.

There are several interesting bits of information we’d like to discuss. First, AMD’s next-generation GPU family, codenamed Vega, is now forecast to arrive hard on the heels of Polaris. After measuring the red boundary boxes and Polaris’ relative position compared to the summer 2016 launch time frame, the position of the Vega box suggests a launch at the tail end of 2016 or very early 2017 timeframe.

The fact that Vega is specifically labeled as using HBM2 also suggests that this is the major high-end architecture revamp that Team Red fans have been waiting for, with much larger frame buffers and, if rumors are true, a true next-generation GPU architecture from AMD (Polaris is still based on GCN, albeit a heavily improved version).
 
The fact that Vega is labeled as HBM2 whereas Polaris isn’t labeled with any type of memory at all suggests that Polaris will either use GDDR5 or GDDR5X. I’m inclined to think GDDR5 is more likely — GDDR5X still requires significantly more power than GDDR5, and isn’t expected to be ready for a volume launch in a matter of months. PC Perspective believes Polaris will still use HBM1, which is also possible.

This slide shows the problem of stacking 2D RAM configurations indefinitely. Polaris could still offer 8GB of RAM in 8x integrated circuits, however, as opposed to the 16 seen here.

If AMD is using conventional GDDR5 for Polaris, it’s probably for practical reasons. When the company debuted Fiji last year, it took pains to note that GDDR5 made progressively less sense as frame buffer sizes and clock speeds ramped up. AMD chose to adopt HBM in 2015, partly because it wanted to deliver an enormous amount of memory bandwidth and was willing to accept a smaller maximum VRAM frame buffer to do it. The company has always maintained that HBM made more sense than GDDR5 for the Fury X and Fury Nano, but never claimed that it would adopt HBM across the entirety of its product stack.



One of the rumors we’ve heard about Polaris is that AMD will target the mainstream and upper-midrange of the GPU market. This is where the money is in enthusiast gaming; Steam’s Hardware Survey is currently broken and not displaying any GPUs from NV at all, but archival data from September showed the GTX 970 as the single-most popular GPU from AMD and Nvidia, with 3.97% of the total install base. The Steam Hardware survey is inexact at best — it often takes months to be updated with new cards and at present all of the Nvidia data seems to have been rolled into Intel’s data — but it still offers objective proof that this is the sweet spot for AMD to target if it wants maximum revenue for its efforts. If AMD intends to deliver a GPU with vastly improved performance per dollar, performance-per-watt, and hold open the option for 8GB cards, GDDR5 may still be the better choice. HBM2 may not be ready for launch and HBM was never more than a niche launch, meant to be cost-effective above a high price point but not to take the technology mainstream.

Whither Greenland?

If you’ve paid attention to the various AMD “roadmaps” passed around in enthusiast circles, you’ve probably seen references to Greenland sprinkled across them in the past year. It’s never been clear exactly what the specs of Greenland were, and the various claims made about it transformed over time. Either way, it’s not listed on this roadmap.

If we assume that Greenland existed at all, this data suggests AMD canceled the part in favor of pulling Vega in and launching an all-new architecture more quickly. Alternately, the code name could still have some bearing on future Zen APUs. We know AMD is working on these parts, but the company has said nothing about a launch date or what features and capabilities they’ll offer.
I’ve speculated in several stories that AMD might use HBM2 memory in APUs to finally break the bandwidth limitations that have held integrated graphics back for nearly twenty years. I still think this is likely in the long run, but whether or not AMD rolls the technology out will depend on both HBM/HBM2 costs and Zen’s ability to command favorable margins and higher absolute prices compared to Carrizo and Kaveri/Godavari.

As of this writing, the cheapest DDR4-2133 on Newegg is $59, while DDR4-3200 is $85. Given that AMD’s integrated GPUs have often been entirely memory bandwidth-bound, it’s entirely possible that the first wave of Zen APUs will lean on the 50% improved memory bandwidth offered by high-speed DDR4-3200 — and 51.2GB/s of memory bandwidth is nothing to sneeze at in a dual-channel system.

Read More