The GTX 1080's epochal launch all but overshadowed its cut-down counterpart – that is, until the price was unveiled. NVidia's GTX 1070 is promised at an initial $450 price-point for the Founders Edition (explained here), or an MSRP of $380 for board partner models. The GTX 1070 replaces nVidia's GTX 970 in the vertical, but promises superior performance to previous high-end models like the 980 and 980 Ti; we'll validate those claims in our testing below, following an initial architecture overview.
The GeForce GTX 1070 ($450) uses a Pascal GP104-200 chip. The architecture is identical to the GTX 1080 and its GP104-400 GPU, but cuts-down on SM presence (and core count) to create a mid-range version of the new 16nm FinFET architecture. This new node from TSMC is nearly half the size of Maxwell's 28nm Planar process, and switches the company over to FinFET transistor architecture for reduced power leakage and overall improved performance-per-watt efficiency. The trend is symptomatic of an industry trending toward ever-smaller devices with a greater concern on the power envelope, and has been reflected in nVidia's architectures since Fermi (GTX 400 series running notoriously hot) and AMD's since Fiji (sort of – Polaris claims to make a bigger push in this direction). On the CPU side, Intel has been driving this trend for several generations now, its 10nm process making promises to further extend mobile device endurance and transistor density.
Had investigators walked into our Thermal-Lab-And-Video-Set Conglomerate, they'd have been greeted with a horror show worthy of a police report: Two video cards fully dissected – one methodically, the other brutally – with parts blazoned in escalating dismemberment across the anti-static mat.
Judging by some of the comments, you'd think we'd committed a crime by taking apart a new GTX 1080 – but that's the job. Frankly, it didn't really matter if the thing died in the process. We're here to make content and test products for points of failure and success, not to preserve them.
The test results are in from our post-review DIY project, which started here. Our goal was a simple one: As a bit of a decompression project after our 9000-word analysis of nVidia's GeForce GTX 1080 Founders Edition, we decided to tear-down the GTX 1080, look underneath, and throw a liquid block onto the exposed die. The “Founders Edition” of the GTX 1080 is effectively a reference model, and as such, it'll quickly be outranked by AIB partner cards with regard to cooling and OC potential. The GTX 1080 overclocks reasonably well – we were hitting ~2025-2050MHz with the FE model – but it still feels limited. That limitation is a mix of power limit and thermal throttling.
Our testing discovered that thermal throttles occur at precisely 82C. Each time the card hits 82C absolute, the clock-rate dips and produces a marginal impact to frametimes and framerate. We also encountered clock-rate stability issues over long burn-in periods, and would have had to further step-down the OC to accommodate the 82C threshold. Even when configuring the VRM blower fan to 100% speed, limitations were encountered – but it did perform better, just with the noise levels of a server fan (~60dB, in our tests). That's not really acceptable for a real-world use case. Liquid will bring down noise levels, help sustain higher clock-rates at those noise levels, and keep thermals well under control.
The video (Part 3) is below. This article will cover the results of our DIY liquid-cooled GTX 1080 'Hybrid' vs. the Founders Edition card, including temperatures, VRM fan RPM, overclocking, stability, and FPS. Our clocks vs. time charts are the most interesting.
In the process of tearing apart the new nVidia GTX 1080 video card, we discovered solder points for an additional 8-pin power header positioned at a 90-degree corner to the original 6-pin header. This is shown in our tear-down video (embedded at the bottom of this post), but we've got a photo above, too.
We're building our own GTX 1080 Hybrid. We're impatient, and the potential for further improved clock-rate stability – not that the 1080 isn't already impressively stable – has drawn us toward a DIY solution. For this GTX 1080 liquid cooling mod, we're tearing apart $1300 worth of video cards: (1) the EVGA GTX 980 Ti Hybrid, which long held our Best of Bench award, is being sacrificed to the Pascal gods, and (2) the GTX 1080 Founders Edition shall be torn asunder, subjected to the whims of screwdrivers and liquid cooling.
Here's the deal: We ran a thermal throttle analysis in our 9000-word review of the GTX 1080 (read it!). We discovered that, like Maxwell before it, consumer Pascal seems to throttle its frequency as temperatures reach and exceed ~82C. Each hit at 82C triggered a frequency fluctuation of ~30~70MHz, enough to create a marginal hit to frametimes. This only happened a few times through our first endurance test, but we've conducted more – this time with overclocks applied – to see if there's ever a point at which the throttling goes from “welcomed safety check” to something less desirable.
Turns out, the thermal throttling impacts our overclocks, and it's limited the potential of a GPU that's otherwise a strong overclocker. And so begins Part 1 of our DIY GTX 1080 build log – disassembly; we're taking apart the GTX 1080, tearing it down to the bones for a closer look inside.
All the pyrotechnics in the world couldn't match the gasconade with which GPU & CPU vendors announce their new architectures. You'd halfway expect this promulgation of multipliers and gains and reductions (but only where smaller is better) to mark the end-times for humankind; surely, if some device were crafted to the standards by which it were announced, The Aliens would descend upon us.
But, every now and then, those bombastic announcements have something behind them – there's substance there, and potential for an adequately exciting piece of technology. NVidia's debut of consumer-grade Pascal architecture initializes with GP104, the first of its non-Accelerator cards to host the new 16nm FinFET process node from TSMC. That GPU lands on the GTX 1080 Founders Edition video card first, later to be disseminated through AIB partners with custom cooling or PCB solutions. If the Founders Edition nomenclature confuses you, don't let it – it's a replacement for nVidia's old “Reference” card naming, as we described here.
Anticipation is high for GP104's improvements over Maxwell, particularly in the area of asynchronous compute and command queuing. As the industry pushes ever into DirectX 12 and Vulkan, compute preemption and dynamic task management become the gatekeepers to performance advancements in these new APIs. It also means that LDA & AFR start getting pushed out as frames become more interdependent with post-FX, and so suddenly there are implications for multi-card configurations that point toward increasingly less optimization support going forward.
Our nVidia GeForce GTX 1080 Founders Edition review benchmarks the card's FPS performance, thermals, noise levels, and overclocking vs. the 980 Ti, 980, Fury X, and 390X. This nearing-10,000-word review lays-out the architecture from an SM level, talks asynchronous compute changes in Pascal / GTX 1080, provides a quick “how to” primer for overclocking the GTX 1080, and talks simultaneous multi-projection. We've got thermal throttle analysis that's new, too, and we're excited to show it.
The Founders Edition version of the GTX 1080 costs $700, though MSRP for AIBs starts at $600. We expect to see that market fill-in over the next few months. Public availability begins on May 27.
First, the embedded video review and specs table:
In this seventeenth episode of Ask GN, we discuss component selection and coil whine avoidance, GPU utilization and its seeming lock to 99% load, fan speeds, wireless mice, and more. Timestamps for all posted questions are below. As always, leave comments on the video page for potential inclusion in the next episode of Ask GN.
A quick thanks to Joe Vivoli of NVIDIA for helping us determine the answer to the first question. One viewer asked, paraphrased, “why does it seem like my GPU only hits 99% load, meanwhile the CPU will hit 100% load?” This was an excellent question for which we did not have an answer, but it seems that, after consultation with engineers, it effectively boils-down to a rounding issue in the software.
A frame's arrival on the display is predicated on an unseen pipeline of command processing within the GPU. The game's engine calls the shots and dictates what's happening instant-to-instant, and the GPU is tasked with drawing the triangles and geometry, textures, rendering lighting, post-processing effects, and dispatching the packaged frame to the display.
The process repeats dozens of times per second – ideally 60 or higher, as in 60 FPS – and is only feasible by joint efforts by GPU vendors (IHVs) and engine, tools, and game developers (ISVs). The canonical view of game graphics rendering can be thought of as starting with the geometry pipeline, where the 3-dimensional model is created. Eventually, lighting gets applied to the scene, textures and post-processing is applied, and the scene is compiled and “shipped” for the gamer's viewing. We'll walk through the GPU rendering and game graphics pipeline in this “how it works” article, with detailed information provided by nVidia Director of Technical Marketing Tom Petersen.
NVidia's bombastic GTX 1080 & GTX 1070 release signified the shift from “Reference” card nomenclature to one more fitting of a Kickstarter campaign, branding its $700 version of the 1080 the “Founder's Edition.” In our ensuing video coverage of the card, viewer comments indicated a clear disconnect with nVidia's intentions regarding the “Founder's Edition” GPU, its differences between the “normal” GTX 1080, and the GTX 1080s from add-in board partners. We're here to demystify that.
NVidia CEO Jen-Hsun Huang hosted the official GTX 1080 ($700) unveil. On stage, Huang indicated – whether intentional or not – that a few different versions of the “reference” GTX 1080 would ship. The price structure was a markedly affordable $600 MSRP for the GTX 1080, or $700 for the GTX 1080 Founder's Edition (or “Legendary Edition,” as we've taken to calling it). Based upon the stage presentation, the Founder's Edition also carried with it a mark of higher overclocking support.
We've learned that's not quite how it works. Here's the deal, plain-and-clear:
Austin, Texas this weekend hosted nVidia's Pascal GTX 1080 unveil event, headed-up by nVidia CEO Jen-Hsun Huang and dotted with high-end demo rigs from well-known case modders (including BS Mods). The week's #OrderOf10 event prefaced tonight's press conference – a puzzle we solved just ahead of the show – and ultimately unlocked a count-down timer that unintentionally coincided with Dreamhack US. Tonight, nVidia announced its new GeForce Pascal architecture GPUs, including the GTX 1080 video cards.
The Pascal architecture was first unveiled at GTC a few years back, when nVidia's roadmap through 2018 was posted (concluding with Volta). Earlier this morning, we published a Pascal architecture deep-dive that fully detailed the new Pascal SM (Streaming Multiprocessor) and memory subsystem as it relates to GP100. Although GP100 belongs exclusively to the Tesla P100 Accelerator Card, Pascal as an architecture applies globally to the platform – and that includes the imminent GeForce cards. Pascal as it relates to GeForce is almost certainly running a “lite” version of the architecture, but will carry-over potential game-changers from the GP100. A 16nm FinFET process node from TSMC heralds nVidia's move away from the 28nm process that both major GPU manufacturers have resided on for the past few years. Other changes, like a reduction in cores-per-SM while increasing or maintaining warp schedulers and dispatch units, will work with this datapath organization change to improve overall efficiency and performance-per-watt of Pascal.
If this architecture discussion interests you, we'd urge you to read our deep-dive on Pascal.
(Article title & content updated).
We moderate comments on a ~24~48 hour cycle. There will be some delay after submitting a comment.