That's about $40 cheaper than I was expecting the GTX 1070 to sell for. Nvidia is really putting some heat on AMD with this card the same way they did with the 970 in 2014.
Might have to pick one up, tbh
I have heard people saying Maxwell has problems switching between render and compute workloads, that you have to flush everything and wait for the GPU to idle to switch from render to compute workloads and vice-versa. While supposedly AMD doesn't have that kind of penalty. From your experience programming on Maxwell and GCN is this true? It might explain why the 970 is getting murdered by the 390 in games like Quantum Break and Hitman.
Yes, AMD has a nice async compute system, BUT, it's really extremely dependent on how your setup is. For example, now I'm writing for the PS4, and while the async compute stuff is really nifty, there's other drawbacks that are platform specific. For example, the PS4 has a unified memory system. It's fast and has multiple buses, but because you can designate memory areas that are accessible by both the CPU and GPU, you can run into some bus and cache stalls. The good thing with consoles though, is that you get a lot more flexibility about processor and memory usage, so it's about finding your weak spots, figuring out what's going on, and optimizing.
I'm actually just getting started on a fairly big project (and making crazy bank, tbh), that's why I haven't been posting much
Oh good, I was scared you got deported. Supposedly Pascal doesn't have this big penalty switching between compute and render workloads, so I'm really interested to see how the GCN friendly games like Hitman and Quantum Break stack up on Pascal.Yes, AMD has a nice async compute system, BUT, it's really extremely dependent on how your setup is. For example, now I'm writing for the PS4, and while the async compute stuff is really nifty, there's other drawbacks that are platform specific. For example, the PS4 has a unified memory system. It's fast and has multiple buses, but because you can designate memory areas that are accessible by both the CPU and GPU, you can run into some bus and cache stalls. The good thing with consoles though, is that you get a lot more flexibility about processor and memory usage, so it's about finding your weak spots, figuring out what's going on, and optimizing.
I'm actually just getting started on a fairly big project (and making crazy bank, tbh), that's why I haven't been posting much
Also, what do you think about the memory bandwidth on the 1080 vs 1070? They're using 256 bit buses so even at 10 GHz effective the 1080 has slightly less memory bandwidth than does that 384 bit 980 Ti with its 7 GHz GDDR5. Do you think 1070 is likely to be killed by that 256 bit bus with 7GHz or 8 GHz effective GDDR5 at 1440p? I imagine it'll really hurt it in 4k. But then again the Fury X murders everything out there for memory bandwidth with a 4096 bit bus and chips clocked at 1GHz, and it is still losing to the 980 Ti in even 4k gaming performance most of the tine.Yes, AMD has a nice async compute system, BUT, it's really extremely dependent on how your setup is. For example, now I'm writing for the PS4, and while the async compute stuff is really nifty, there's other drawbacks that are platform specific. For example, the PS4 has a unified memory system. It's fast and has multiple buses, but because you can designate memory areas that are accessible by both the CPU and GPU, you can run into some bus and cache stalls. The good thing with consoles though, is that you get a lot more flexibility about processor and memory usage, so it's about finding your weak spots, figuring out what's going on, and optimizing.
I'm actually just getting started on a fairly big project (and making crazy bank, tbh), that's why I haven't been posting much
Actually, for 4k, the rendering pipeline is a little different. It's more of a tiled-type system. I'm still reading up on it since this thing I'm working on has to support 4k on the Neo. But there's some trickery going on so you don't have to render a full 4k frame every vsync and the bandwidth that consumes. The thing with these 'advanced' DX12 style units is that there's basically mul hreading support on the card itself. As long as you can load up workloads that can be done independently and in parallel, you can really max these things out throughput-wise. IIRC, you can also chain workloads to automatically trigger when you join one or more threads, so that's a lot of stuff you had to sync from the CPU, but now the cards are smart enough so you can send up work queues and they'll mostly be entirely GPU bound.
RandomGuy, here's some info about the upcoming AMD cards.
Based on official word from AMD, this card isn't going to be competing with GTX 1070 and GTX 1080, but instead trying to expand the userbase of VR capable PCs. So most likely what we'd be looking at here is R9 390 level performance for a lower price than the 390 sells for, and using a lot less power, as they have already demoed Polaris 11 (the slower version of the chip) at the GTC show in January and its power consumption was really impressive. Unless they give a really steep price discount to the upcoming R9 480x vs the currently available R9 390, it looks like the GTX 1070 at $380 is going to be a much better buy. We still have to wait for benchmarks to see how GTX 1070 handles compute effects, but shoring up this weakness of the current gen Maxwell with the next gen Pascal seems to have been a focus for Nvidia, and I expect this will have been taken care of.
The specs on the new Nvidia cards look sick as .. 8 GB of bull free VRAM on the 1070 is nice, too.. Regarding the GDDR5X memory on the 1080, what exactly is that gonna lead to benchmark/graphics wise?? Is it gonna provide any noticeable gains?? Less pop up?? What?
Are you talking 4k on the Neo or 4k in general? Do you see the memory bandwidth of the 1070 hurting it in say 1440p vs the 980 Ti there, given the 980 Ti has a 336GB/s memory bandwidth while the 1070 will only have either 224GB/s or 256GB/s, depending on if they use 7GHz or 8GHz chips? What about 1080 at 1440p with 320GB/s?
The 1070 is going to have significantly less memory bandwidth than the 1080 and 980 Ti / an X. The 980 Ti / an X have a 384 bit memory bus while the 1070 and 1080 have a 256 bit one. So with 10 GHz GDDR5X the memory bandwidth of the 1080 is
10 GHz * 256 bit = 2560 Gb/s = 320 GB/s.
With the 980 Ti / an X using 7 GHz GDDR5 its memory bandwidth is
7 GHz * 384 bit = 2688 Gb/s = 336 GB/s.
With 7 GHz GDDR5 the 1070 would have memory bandwidth
7 GHz * 256 bit = 1792 Gb/s = 224 GB/s,
same as the 980. Hopefully it'll have 8 GHz GDDR5 to increase the memory bandwidth to
8 GHz * 256 bit = 2048 Gb/s = 256 GB/s.
Rendering for 4k in general requires some trickery, it's just massive amount of data and processing. The bandwidth is entirely dependent on the game/what you're rendering. You use bandwidth to read/write textures, vectors, shadow maps, depth buffers, stencil buffers, etc. If you do little compute, then bandwidth becomes more prominent, because you're likely maxing that out instead of GPU processing (Bandwidth would be your bottleneck). If you're doing a lot of compute, then GPU clock speed will likely be your bottleneck. Most modern games do multi-pass rendering though, that's why you want bandwidth to be high. On the other hand, having a better compute engine allows for better parallelism. For example if your compute work doesn't necessarily affect the current render pass, you could run them both in parallel, with bandwidth used exclusively for the render. This wasn't necessarily possible with DX11, that's why APIs like Vulcan came in (and now DX12). So, regardless of bandwidth, there's gains to be had that way. But you probably need to rework your render pipeline to take advantage of that. If you use an middleware engines like Unreal, Unity or Phyre, you get support for that stuff basically for free, which is nice.
That's not correct. The X in GDDR5X means it can prefetch 16 bytes at a time as opposed to 8 bytes that GDDR5 did.
So the calculation goes from "DDR clock * 2 * (Bus Width / 8)" -> "DDR clock * 4 * (Bus Width / 8)" for GDDR5X (these are theoretical maximums in both cases).
That makes the 1080: 20 GHz * 256 bit = ~640 GB/s
See here: http://wccftech.com/micron-gddr5x-me...-graphic-card/
That's funny, they're advertising 320GB/s.
http://www.geforce.com/hardware/10se...force-gtx-1080
Sorry, I wasn't clear I was giving effective clock when I said 10 GHz, since that's how it's advertised.
, maybe it's only the Founder's Edition cards launching May 27th and June 10th. I wouldn't want to buy those anyways since the Nvidia stock cooler sucks.
http://nvidianews.nvidia.com/news/a-...force-gtx-1080
Availability and Pricing The NVIDIA GeForce GTX 1080 "Founders Edition" will be available on May 27 for $699. It will be available from ASUS, Colorful, EVGA, Gainward, Galaxy, Gigabyte, Innovision 3D, MSI, NVIDIA. Palit, PNY and Zotac. Custom boards from partners will vary by region and pricing is expected to start at $599. The GeForce GTX 1080 will also be sold in fully configured systems from leading U.S.-based system builders, including AVADirect, Cyberpower, Digital Storm, Falcon Northwest, Geekbox, IBUYPOWER, Maingear, Origin PC, Puget Systems, V3 Gaming and Velocity Micro, as well as system integrators outside North America. The NVIDIA GeForce GTX 1070 "Founders Edition" will be available on June 10 for $449. Custom boards from partners are expected to start at $379. - See more at: http://nvidianews.nvidia.com/news/a-....hy5yFjTe.dpuf
Normally, you don't hit peak anyways, because you're bound to have prefetch cache misses, so a theoretical 640GB/s is really about 480GB/s... that's both for GDDR5 and GDDR5X...
The thing is, with GDDR5X, if you don't have a lot of misses, it should be 2x as fast as GDDR5.
The one they were demoing that was clocked at 2.1GHz was a PNY, IIRC. It was on one of the screen caps from the presentations I saw.
That 2.1 GHz 1080 running the upcoming Doom in Vulkan:
Once they uncap the framerate it only drops below 100 fps twice, both times for a split second. They didn't say the resolution, I hope they didn't do something cheap and run it 1680x1050 or something.
RandomGuy I do have to say the GTX 1070 looks like the most cut down 70-series chip Nvidia has ever released. The drop from 9 TFLOPS to 6.5 TFLOPS would correlate well with the GTX 1070 having the 1980 cuda cores we heard from one of the leaks about GTX 1080 a few days before release (GTX 1080 has 2560 cuda cores per official specs from Nvidia, they haven't released the specs for 1070). So still take these reports of better than an X performance with a grain of salt. CEOs often give boldface lies in these presentations. Like Lisa Su from AMD saying the Fury X would be an overclocker's dream last Computex when you couldn't overclock a ing thing on it for many months after release (I'm not even sure if you can now a year later). So definitely wait for benchmarks before buying a 1070. I'll post them here and tag your name when they become available. I probably wouldn't expect 1070 benchmarks until a couple of days before release though, maybe June 7th or 8th. It still should be a really nice card for $380 though because of those clockspeeds, I imagine when benchmarks come out it'll be a little faster than an X at 1080p, a little slower at 1440p, and a decent bit slower at 4k.
Last edited by baseline bum; 05-08-2016 at 01:02 PM.
Isn't the CPU doing most the FLOPS? (floating point operations)
not for games, no
Maybe he just means for Kerbal Space Program lol
Nvidia unveils GTX 1080 and GTX 1070: a new level in GPU power
"Irresponsible levels of performance."
By Richard Leadbetter Published 07/05/2016
Nvidia has officially revealed the GeForce GTX 1080 and GTX 1070, based on its new Pascal architecture, offering an astonishing leap in performance and power efficiency over its existing 900-series Maxwell cards. According to Nvidia, the GTX 1080 is faster than two GTX 980s in SLI, shipping to users on May 27th at $599. Meanwhile, the GTX 1070 offers an X level performance for just over one third of the price - $379 - and is set to launch on June 10th.
Both products are based on Nvidia's GP104 processor, featuring a 7.2 billion transistor count. The GTX 1080 features 2560 CUDA cores, offering 9 TFLOPs of performance, and is paired with 8GB of Micron's state-of-the-art GDDR5X (or G5X, as Nvidia called it). Meanwhile, no core count has been announced for the GTX 1070, but we do know that it ships with 8GB of standard GDDR5 and has a 6.5 TFLOP performance level.
Clock-speeds and overclocking headroom look remarkable. Boost clock on GTX 1080 is rated at higher than 1700MHz, but a real-time Unreal Engine 4 demo shown on stage saw the card overclocked on air, rock-solid at 2114MHz. On-screen stats using the EVGA Precision X monitoring software saw temperature at just 67 degrees Celsius, with the G5X memory operating at 5508MHz - 10gbps effective.
In terms of direct comparisons with the firm's previous performance leader, Nvidia revealed a power efficiency/performance graph which seems to suggest something along the lines of a 20 per cent performance increase for GTX 1080 running at 180W, compared to an X operating at 250W.
GTX 1080 offers a sizeable performance boost over an X, at a much lower power level. It was also seen overclocked by 400MHz, running at 2.1GHz vs the max 1400-1500MHz you can extract from previous king of the hill.
Of course, Add-in-board manufacturers will be supplying their own products around the suggested retail prices, but Nvidia will be selling its own 'Founder's Edition' variants through select partners, bumping up prices to $699 for the GTX 1080 and $449 for the GTX 1070. Nvidia boss Jen-Hsung Huang promises "crazy overclockability" for these parts perhaps suggesting some kind of custom power delivery mechanism. We'll look into this and report back, but all Pascal cards will be overclockable - this isn't just limited to the more expensive variants.
In our recent 'In Theory' piece, we pieced together a potential scenario for the performance level of the new wave of Pascal GPUs, based on available information derived from the official spec for the much larger GP100 chip, combined with the more plausible leaks from the Far Eastern press. We suggested that Pascal could offer an X-level performance for GTX 970 money - based on an established precedent that saw the 970 beat its big-chip predecessor (GTX 780 Ti) once an overclock was in place. Some might say that Nvidia has exceeded our expectations here, assuming that an X level performance is indeed delivered at stock clocks. However, GTX 1070 gets a $40 price-bump compared to its $330 predecessor. Regardless, the price-to-performance ratio is simply remarkable.
Bearing in mind we were already quite optimistic about Pascal's chances based on the GP100 specs, the bottom line here is that on the day, Nvidia exceeded expectations last night by some distance. Jen-Hsun Huang stated several times that Pascal took over two years to complete with the input of "thousands" of engineers and a multi-billion dollar investment. The inference here is that the firm intends to maintain its massive market dominance by outspending AMD in terms of research and development.
GTX 1080 offers up 7.2bn transistors, enabling 9TFLOPs of rendering power, backed by 10gbps bandwidth offered by Micron's new GDDR5X technology. Price-wise, it replaces the GTX 980 Ti at $600.
The firm also revealed new technology at the heart of the Pascal architecture, called 'simultaneous multi-projection', with the GPU able to create 16 independent viewports simultaneously in one pass with stereo support - and no performance penalty. Two scenarios for this tech were showcased, the first being full perspective correction on satellite displays on a tri-screen set-up, positioned in a 'curved' surround arrangement.
The second was for virtual reality. Right now, VR is rendered by creating a full resolution output, warping the output to match the lenses on the HMD - and in the process, losing a lot of visible resolution. With simultaneous multi-projection, each eye view is a composite of four projections, meaning that 'wasted' resolution is massively reduced, and the end result is anything from a 40 to 80 per cent increase in performance (the on-screen demo seemed to be in the 45-50 per cent area, rising from 65fps to 96fps with the single-pass technology enabled). [UPDATE 7/5/16 3:48pm: Having discussed this in more depth with Nvidia this morning, the performance increase comes from two components - the reduction in resolution as stated, and also the hardware acceleration provided for the stereo viewpoint, which essentially halves the processing overhead for geometry.]
Other new technologies revealed at the event were software-based in nature, with the reveal of a VRWorks audio set-up based on path-traced sound using a physically modelled audio simulation. This is paired with physically-based haptic feedback for VR controllers, in combination with PhysX. Nvidia seemed to suggest that all of this new VR technology will be open sourced, and revealed that a series of tech demos - dubbed Nvidia VR Funhouse - will be made available in a single Steam download.
It's a little more expensive than its predecessor - by around $40 - but assuming that Nvidia's claims stack out, GTX 1070 should see something in the region of a 40 to 55 per cent increase in overall performance, with stock clocks offering an X levels of performance. In theory, once overclocked, it should beat a an X pushed to its limits.
Nvidia also revealed Ancel - an in-game 3D camera system, allowing you to pause gameplay, adjust special effects, brightness, field of view and then create your own 2D, 3D and surround screenshots using anything up to 1000x resolution. The Division, The Witness, Lawbreakers, The Witcher 3: Wild Hunt, Paragon, No Man's Sky and Unreal Tournament are all set to receive Ancel support. Although it seems to operate at the driver level, clearly developer support is essential here in getting Ancel properly integrated.
There was no mention of DirectX 12 at the event - something of a sticking point for Nvidia cards thus far - but accompanying press materials describe "new asynchronous compute advances" that "improve efficiency and gaming performance", so it'll be interesting to see how Pascal performs there in an area where the existing Maxwell cards have had some issues.
Overall, it's difficult not to be impressed with the show Nvidia put on here. We knew that GTX 1080 would be the new flagship, but to post a circa 20 per cent improvement to an X performance is immense, factoring in the 28 per cent reduction in TDP. For its part, assuming that the an X comparison holds true, the GTX 1070 offers a potential 40-55 per cent improvement in performance over the GTX 970, costing just $40 more. It could be as radical a proposition as GTX 970 was back at launch - a product we described at the time as "the GPU that nukes almost the entire high-end graphics market from orbit."
Going into the event, Nvidia staff gave us the impression that Pascal would offer the same level of ambition and performance that defined the GPU classics of the past - graphics cards like the GTX 970 and the 8800GT. Of course, the jury's out until both products are in the Digital Foundry lair with the benchmarks complete, but in the meantime, hopes are sky-high based on this remarkable showing.
Nvidia's Pascal gaming cards were unveiled at an event hosted in Austin, Texas that we attended. Nvidia paid for travel and accommodation.
Last edited by baseline bum; 05-08-2016 at 05:33 PM.
There are currently 1 users browsing this thread. (0 members and 1 guests)
Posting Permissions
Reply With Quote
