Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Watch NVIDIA’s Special Event Livestream
NVIDIA is hosting a special event today to unveil their latest GeForce products. The latest products are said to be based on the brand new Pascal architecture and will deploy several key technologies aimed specifically for gamers and enthusiasts. These technologies include enhanced DirectX 12 and Vulkan API support, powerful VR capabilities and several new gaming initiatives built under NVIDIA’s GameWorks program.
Read more: http://wccftech.com/watch-nvidia-special-geforce-pascal-livestream-dx12-vr-gameworks/#ixzz47uYNOx3J
Comments
Because of this people, read no Nvidia Fanboys, will wait for some real benchmarks before buying into the PR machine.
GeForce GTX 1080
9 TFLOPS
8 GB GDDR5X
$599
May 27
GeForce GTX 1070
6.5 TFLOPS
8 GB GDDR5
$379
June 10
Nvidia is claiming 180 W for the GeForce GTX 1080.
I'm not really sure what to make of this. Widespread retail availability of a GDDR5X part on May 27 is flatly impossible, unless they mean May 27, 2017 or some memory vendor has secretly been producing GDDR5X for months and not told anyone. Maybe it will be a paper launch on May 27 with actual availability somewhere around November.
As for the GTX 1070, there isn't any trivially obvious reason why that can't possibly be a hard launch. Maybe it will be a hard launch or a paper launch or somewhere in between. But it's quite a drop-off from 9 TFLOPS to 6.5 TFLOPS. You could literally disable 1/4 of the chip with less of a drop-off than that. Either that's a different chip entirely or yields are awful or something strange is up.
My best guess is that the GTX 1070 is the "real" part for most of what's left of this year, but will be slower than current generation parts. They're announcing the GTX 1080 now, to try to make it look like the new generation is faster than the old.
It's also interesting to note that Nvidia is claiming a GP100 at a little over 10 TFLOPS and 300 W, and then a GeForce GTX 1080 at 9 TFLOPS and 180 W. Having double-precision support does take space and power, but traditionally the top end chips with double-precision support haven't been so much worse than the lower end chips without it. Maybe GP100 is all kinds of broken, and Nvidia fixed a bunch of things that were wrong with it in the next chip down for the GTX 1070 and GTX 1080. Or maybe the GTX 1080 specs are wishful thinking.
거북이는 목을 내밀 때 안 움직입니다
거북이는 목을 내밀 때 안 움직입니다
Also I would hope for a 20% improvement at lower wattage. They are die-shrinking after all.
Nvidia is promising 320 GB/s on the GTX 1080 with GDDR5X. For comparison, the GeForce GTX 980 Ti has 336 GB/s and the Radeon R9 390 has 384 GB/s, both with GDDR5. GDDR5X will make it so that you can get a given amount of bandwidth with a narrower bus, and that brings costs down, but that's an internal detail, not something that you should base purchasing decisions on.
http://www.geforce.com/hardware/10series/geforce-gtx-1080
2560 shaders
1607 MHz base clock
1733 MHz boost clock
256-bit GDDR5X
320 GB/s memory bandwidth
7680x4320 max resolution
180 W
If you consider only the cards available today, Nvidia is ahead of AMD in any efficiency metrics: performance per watt, performance per mm^2, performance per dollar to build cards, etc. Normally, when you're ahead, you want people to think it's a great time to buy now, and not to wait for the next generation. So it makes sense for AMD to hype Polaris a lot, but less so for Nvidia and Pascal.
Until Polaris launches, that is. If Polaris launches and all that Nvidia has shown is Maxwell, then unless Polaris is a huge flop, AMD will be clearly ahead of Nvidia. AMD showed off Polaris early this year, so a hard launch around the middle of this year is plausible--and even likely unless AMD needed multiple respins after the silicon they showed publicly. If Polaris is out and Pascal isn't, then Nvidia really wants people to wait for Pascal.
Nvidia seemed to be mostly hyping the GeForce GTX 1080, with only brief mention of the GTX 1070. But the GTX 1080 is a GDDR5X part, which means retail availability is perhaps half a year away or more. You can do a paper launch with only press availability much sooner than you can have hundreds of thousands of cards available at retail, and it sounds like Nvidia may go that route just to get people to wait.
But this casts the GP100 chip in a really bad light. Nvidia is claiming that as compared to the GTX 1080, GP100 needs 67% more power for about 15% more performance. That's brutal energy efficiency. Double precision does eat up some, but Tahiti wasn't that much worse than Pitcairn and Hawaii wasn't that much worse than Tonga. GK110 was more efficient than other Kepler chips, and while GF100 was markedly worse than GF104, that was only by about 20%, not 50%--and GF100 was easily the most broken GPU chip to launch in at least the last decade.
And that's in spite of GP100 having some huge energy efficiency advantages over the GTX 1080. HBM2 should be much more efficient than GDDR5X. More shaders clocked lower is usually much more efficient than fewer clocked higher, and Nvidia is promising to clock the GTX 1080 about 15% higher than GP100.
Both the GTX 1080 and GTX 1070 launch dates are on Fridays. That's rare in the tech world, but a Friday morning launch isn't the same sort of ominous news as a Friday evening launch. We'll see what happens, but I'd expect paper launches.
That leaves open the question of how it compares to AMD. There are rumored specs out on Polaris 10 and 11, but they're far from official. Even if AMD announced official specs today, I'm not inclined to treat dueling pre-launch paper specs as conclusive.
Unlike Nvidia (at least for the moment), AMD has every reason to rush cards to market and get on with the next generation. Of course, also unlike Nvidia, AMD showed off working cards in January.
I remain skeptical.
Announcing a date is not at all the same thing as shipping a product.
+ same DX12 support level as maxwell, and no confirmation if async is still software, or actually hardware now.
Display port is only 1.2 version certified, 1.3/1.4 upgradeable :pleased:
% wise overclock range is actual lower than 980 Ti because the 1080 already boosts to 1733Mhz.
970/980>1070/1080 price is $50 increased at base level.
Actual performance increase over 980ti/Titan X per 1080p/4k/VR almost impossible to determine due to all the PR speak, and perf/watt mumbo jumbo, But i've seen similar calculations in comments on all major tech websites come to the ballpark of ~26% overall performance increase over 980 Ti stock
Oh and 15w TDP increase over 980.
So basically, my recommendation is, don't buy an early card with a GP100 chip, as there's probably something much better coming soon after it. My other recommendation is, don't pre-order a GeForce GTX 1070 or especially a GTX 1080, as you might be waiting six months--and might look like a pretty bad deal by the time you get it. But if the GTX 1070 or 1080 is available at retail and the price and performance are agreeable to you, then have at it. When that eventually happens.
So what about Micron? They started sampling GDDR5X in late March. And then there's this:
https://www.micron.com/about/blogs/2016/february/gddr5x-has-arrived
"We plan to hit mass production this summer."
https://www.micron.com/support/faqs/products/dram/gddr5x
"Micron is the first memory supplier in the industry ramping GDDR5X product. This puts Micron in the leading position on high-speed signaling with traditional memory components."
Starting to mass produce GDDR5X this summer is not at all similar to video cards that use GDDR5X being commercially available this summer. It takes time to send the wafers through the fabs, ship them to companies that assemble video cards, actually assemble the cards, and then ship completed cards around the world to be available to buy.
Even if everything else goes flawlessly for Nvidia and they're just waiting on memory, widespread availability of the GTX 1080 isn't coming until late this year at best. Think November, not July. And everything else going flawlessly for Nvidia isn't likely to happen, either. This is why Polaris went with GDDR5: AMD will likely be the only GPU vendor with next generation products available for several months. AMD isn't going exclusively GDDR5 on 14 nm, though; Vega will show up early next year sporting HBM2.
Notably missing yesterday was any mention of lower end cards. You'd have to believe that a GeForce GTX 1060 or GTX 1050 or something is coming eventually, right? If they're not coming until after the cards that have been announced, that would mean that AMD has basically the entire sub-$600 market all to themselves for several months.
And then there are laptops. Polaris 11 is likely to be a great laptop card, with Polaris 10 perhaps appropriate for high end gaming laptops. Maybe Nvidia could throttle back clock speeds to make the chip they talked about yesterday into a 100 W laptop card. But what about 30 W and 50 W and 75 W laptop cards--the ranges that constitute essentially all discrete laptop cards?
Even so, I'm actually more bullish on Nvidia than I was a day ago. GP100 made it look like Nvidia was going to get absolutely slaughtered in all of the efficiency metrics this generation. Yesterday's announcements look a lot more competitive, and look like Nvidia is getting the gains from the die shrink that they should have.
Rather, what's likely to happen is that AMD will have most of the market to themselves for several months, with Nvidia just trying to convince people to wait for Nvidia cards that are just as good but not yet available. If one ignores those several months, it's not at all clear who wins this generation. It wouldn't be surprising if this is basically a repeat of GCN versus Kepler: AMD is available first, Nvidia gets there some months later, then a while after that, Nvidia takes the lead, and then AMD eventually catches up. The timeframe should be much more compressed, though, with Vega following less than a year after Polaris, as compared to 20 months from Tahiti to Hawaii.
거북이는 목을 내밀 때 안 움직입니다
Look at the price difference between them, too. A 58% premium for the top bin over the second bin of the same chip? I can't think of any other example ever where it was that large of a premium for a mid-range or higher GPU chip. Nvidia might be paying a fortune for early GDDR5X chips, or they might just be asking a huge price premium because there are hardly any chips and because they can.
And also look how severely cut down the GTX 1070 is. The GTX 1080 has 38% more TFLOPS than the GTX 1070. When does that ever happen outside of low end parts? For comparison:
Titan X/GTX 980 Ti: 9%
GTX 980/GTX 970: 32%
Titan/GTX 780: 13%
GTX 680/GTX 670: 26%
GTX 580/GTX 570: 13%
GTX 560 Ti/GTX 560: 16%
GTX 480/GTX 470: 24%
GTX 285/GTX 275: 5%
GTX 280/GTX 260: 30%
Fury X/Fury: 20%
R9 390X/R9 390: 15%
R9 380X/R9 380: 14%
HD 7970/HD 7950: 32%
HD 7870/HD 7850: 45%
HD 6970/HD 6950: 20%
HD 6870/HD 6850: 35%
HD 5870/HD 5850: 30%
I guess Pitcairn did have a further cut down second chip, but that's a small die at 212 mm^2, not really in the same class as GP104.
There's also the issue that more memory bandwidth only benefits you if you can use it. Nvidia is claiming that the GTX 1080 will have 38% more compute, but only 25% more bandwidth. If the GTX 1080 isn't seriously throttled by bandwidth, then the GTX 1070 might not benefit from the extra bandwidth at all.