xthetenth posted:That much memory bandwidth means memory speed isn't going to be a bottleneck period. However you don't need all of that bandwidth most/all of the time. It looks like Fury's being held back by something else, while the NV cards are better balanced and avoid the bottleneck that's holding Fury back. Either something's wrong with the drivers or something can't keep up with the rest of Fury. Overall though HBM's major speed boost was by letting them put more power to the rest of the card, not by taking memory bandwidth from sufficient to overkill. Yeah, something odd is happening with the Fury's memory bandwidth, look at the numbers here. Why is it only getting 387GB/s with a highly compressible solid black texture? And only 333GB/s on the incompressible one? The 980 Ti gets 364GB/s on the black texture and 234GB/s on the colored one, so the texture compression used on the Fury is pretty terrible but more importantly the Fury should be getting around 512GB/s and it's maxing out at only 387GB/s, something is going wrong here.
|
|
# ? Jun 26, 2015 14:27 |
|
|
# ? May 31, 2024 10:20 |
|
BurritoJustice posted:You could overclock two of them at max power target (275w) and overclock your CPU too with power to spare. Rephrasing. What kind of power draw should I shoot for on a custom PCB 980 Ti if I have a 660W power supply (a good one, Seasonic plat) and also overclock my CPU? Sidesaddle Cavalry fucked around with this message at 14:39 on Jun 26, 2015 |
# ? Jun 26, 2015 14:34 |
|
DaNzA posted:I guess the other question is how come nvidia can beat the HBM cards with that much more bandwidth using traditional GDDR5? and does that mean there's going to be a noticeable leap for them when they get HBM2 onto their Pascal next year. edit: and yes something is out of whack with Fiji's performance too, I'm still hoping its drivers because if its a hardware problem it probably will never be fixed. NV should get quite a nice bump in performance when they go to HBM2 and 14/16nm processes. A 40-50% performance improvement over the 980 in games would be my WAG. More would be quite doable if they're willing to get the GPU have a Hawaii-esque TDP. Again, its a WAG, but I don't think its unreasonable given how much transistors 14/16nm will let them cram on 1 die + the bandwidth improvements of HBM2. AMD's Arctic Islands is supposedly following in NV's footsteps (ie. focused on gaming and not GPGPU/compute) and will also have HBM2 and 14/16nm processes to work with. So long as they don't make any silly mistakes there is no reason not to expect them to have a GPU that is at least competitive with NV's Pascal. PC LOAD LETTER fucked around with this message at 14:44 on Jun 26, 2015 |
# ? Jun 26, 2015 14:40 |
|
Sidesaddle Cavalry posted:
If you are using a haswell cpu, expect ~400-500w load at the wall. So 350-450w actual psu load.
|
# ? Jun 26, 2015 14:51 |
|
Sidesaddle Cavalry posted:
It depends how high MSI sets the power target in the BIOS, but if past releases are anything to go off I'd imagine the 300-350w range minimum. Of course this is only if you also increase voltages and clock speeds to actually draw that much.
|
# ? Jun 26, 2015 16:29 |
|
Oddly enough I managed to get bites on my waterblocks before the cards. Trying to decide if I want to sell my Titan X and get one of those HoF 3 pin water 980 ti cards ><
|
# ? Jun 26, 2015 17:14 |
|
DaNzA posted:I guess the other question is how come nvidia can beat the HBM cards with that much more bandwidth using traditional GDDR5? and does that mean there's going to be a noticeable leap for them when they get HBM2 onto their Pascal next year. xthetenth posted:That much memory bandwidth means memory speed isn't going to be a bottleneck period. However you don't need all of that bandwidth most/all of the time. It looks like Fury's being held back by something else, while the NV cards are better balanced and avoid the bottleneck that's holding Fury back. Either something's wrong with the drivers or something can't keep up with the rest of Fury. Overall though HBM's major speed boost was by letting them put more power to the rest of the card, not by taking memory bandwidth from sufficient to overkill. The 20nm fiasco probably hit AMD very hard, being area-limited. We know that Fiji has an area of 596 mm^2. We also know that only an additional 4mm^2 of silicon could have been placed on the interposer. (which brings the total up to 600mm^2, a nice round number. I think it's significant.) Which probably means that AMD's original design spec for 20nm Fiji was not planned to take up every last scrap of spare area on the interposer. My take on this is that had the 20nm transition gone off without any hitches, Fiji would be a smaller chip in terms of physical area, but the smaller process would mean that they would have more room to have even more transistors than it presently has, filling out all the features that we find lacking in it. (ROP count, HDMI 2.0, etc.) I think that what we see as "top of the line" Fiji XT right now is a cut-down version of a much larger and beefier chip, limited by physical die space. Sort of like Tonga before it, there's a full-fat version lurking somewhere that can't be built on 28nm and have it fit on an interposer. Hell, maybe that's where Tonga came from, AMD having to chop bits off of Fiji XT to get it down to size. (I'm joking on that last part. That's not true, I made that up.) I think they're still grappling with the ramifications of TSMC making GBS threads the bed, is what they're doing, and that the Fury X is AMD impressively making the best of a bad situation. SwissArmyDruid fucked around with this message at 17:52 on Jun 26, 2015 |
# ? Jun 26, 2015 17:47 |
|
fiji likely has severe yield issues, a combination of being huge and using HBM + interposers GPUs have been stuck at 28nm because mobile drives process tech, so the higher perf processes gpus like are not really developed as quickly. when you sell a rounding errors worth of cards compared to mobile SoCs, you aren't usually catered to 28nm is also a local minimum on cost per transistor; iirc 20nm and "16nm" are actually more expensive right now once that changes and there's a 14/16 High Perf process out there then we'll see die shrunk GPUs
|
# ? Jun 26, 2015 17:56 |
|
TDR hotfix will be published today: http://forums.guru3d.com/showpost.php?p=5108246&postcount=251
|
# ? Jun 26, 2015 18:18 |
|
Malcolm XML posted:GPUs have been stuck at 28nm because mobile drives process tech, so the higher perf processes gpus like are not really developed as quickly. when you sell a rounding errors worth of cards compared to mobile SoCs, you aren't usually catered to
|
# ? Jun 26, 2015 18:38 |
|
I am loving the new 980ti for 1440p sweetness. Going from 30 fps on medium to 60 fps on ultra in Witcher 3 and GTAV is a great change. I had considered waiting for pascal, but there are plenty of games I have to enjoy before then. GTAV, Witcher 3, Fallout 4, and so on. Pascal could conceivably be 12-18 months away. I already shipped off my old 760 to one of my friends to use in his girlfriend's computer.
Filthy Monkey fucked around with this message at 20:54 on Jun 26, 2015 |
# ? Jun 26, 2015 20:49 |
|
An interesting Fury X review from JayzTwoCents. Overclocked Fury X vs Overclocked reference 980 ti vs Overclocked 3rd party 980 ti's vs EVGA 970 SSC SLI. https://www.youtube.com/watch?v=iEwLtqbBw90 Ak Gara fucked around with this message at 20:57 on Jun 26, 2015 |
# ? Jun 26, 2015 20:55 |
|
His conclusion is a fury is good if you are going to overclock it but would not over clock a 980ti.
|
# ? Jun 26, 2015 21:08 |
|
Ak Gara posted:An interesting Fury X review from JayzTwoCents. Wtf it's like under the 980 (non-Ti) in several cases, even OC'd. It competes in Metro but who even plays metro?
|
# ? Jun 26, 2015 21:11 |
|
Ak Gara posted:An interesting Fury X review from JayzTwoCents. I didn't really hear the noise he was bitching about until I turned the volume up. Apparently I'm not the only one and he's kinda mad about that
|
# ? Jun 26, 2015 21:18 |
|
This guy did a better job recording it https://youtu.be/XfyQzroYnrI
|
# ? Jun 26, 2015 21:24 |
|
Should I hold out for the non-reference cooling 980 tis to come back in stock, or are the reference cooler cards OK? I don't care about overclocking.
|
# ? Jun 26, 2015 21:25 |
|
d3rt posted:Should I hold out for the non-reference cooling 980 tis to come back in stock, or are the reference cooler cards OK? I don't care about overclocking. I have two ref titan x cards and honestly the noise doesn't bother me. The nvidia stock cooler noise is very easy on the ears. That being said it will be louder than non ref coolers. You could also always install a evga acx cooler later.
|
# ? Jun 26, 2015 21:28 |
|
Alereon posted:I don't think this is really true, I think TSMC is working as fast as they can to get 20nm ready, it's just MUCH more difficult than any of the previous process transitions they've done. Even Intel has run into serious difficulties with manufacturing below 20nm, and they've been a process node ahead of the rest of the industry for some time. i strongly suspect that tsmc skips 20nm and goes directly to 16/14nm since 20nm missed the mark samsung is shipping millions of phones w/ a 16/14nm SoC apple a9 will likely be 16/14nm
|
# ? Jun 26, 2015 21:33 |
|
d3rt posted:Should I hold out for the non-reference cooling 980 tis to come back in stock, or are the reference cooler cards OK? I don't care about overclocking.
|
# ? Jun 26, 2015 21:34 |
|
d3rt posted:Should I hold out for the non-reference cooling 980 tis to come back in stock, or are the reference cooler cards OK? I don't care about overclocking. The Gigabyte and MSI cards with custom pcbs are harder to find, and get snapped up pretty fast still. The Asus version still has yet to be spotted in the wild.
|
# ? Jun 26, 2015 21:34 |
|
Filthy Monkey posted:The EVGA ACX cards aren't that hard to get. Put an auto-notify in on the evga site, or watch newegg. They come up pretty frequently. They aren't as quite as good as the Gigabyte or MSI customs, but I would totally take one over a reference blower. Woot, looks like I got my waterblocks sold, should make unloading the 980s easier.
|
# ? Jun 26, 2015 21:39 |
|
Bleh Maestro posted:Wtf it's like under the 980 (non-Ti) in several cases, even OC'd. It competes in Metro but who even plays metro? Whoa that 980 Classified must be totally "overclocked", big boy Nvidia bully tactics making AMD cards look slow d3rt posted:Should I hold out for the non-reference cooling 980 tis to come back in stock, or are the reference cooler cards OK? I don't care about overclocking. I would not buy a reference 650$ card unless you really need the blower to push out hot air (SLI) or for watercooling mods. Wait for the MSI 6G or whatever EVGA S(S)C+ ACX+2.0 there is. sauer kraut fucked around with this message at 22:18 on Jun 26, 2015 |
# ? Jun 26, 2015 22:01 |
|
I think this slipped under the radar here, since both English-speaking review sites that use thermal imaging inexplicably left the cover on: The Fury Xes VRMs are going thermonuclear even at stock settings. Overclockers dream indeed
|
# ? Jun 26, 2015 22:28 |
|
repiv posted:I think this slipped under the radar here, since both English-speaking review sites that use thermal imaging inexplicably left the cover on: Dammmn. You'd think they could have / should have just put an actual block on it then to cover the GPU and VRM instead of just that little copper pipe.
|
# ? Jun 26, 2015 23:38 |
|
The pipe should be plenty though. I'd first assume there was something wrong with the way its fit to the VRM or the thermal paste wasn't applied properly. The latter issue in particular is something that has plagued many cards over the years. edit: \/\/\/\/\/\/\/Yuuup. Whatever process they're going to use for GPU's will have to be very different from what they use for cellphones if they're going to try and keep improving performance over the previous generation. PC LOAD LETTER fucked around with this message at 23:54 on Jun 26, 2015 |
# ? Jun 26, 2015 23:41 |
|
Malcolm XML posted:i strongly suspect that tsmc skips 20nm and goes directly to 16/14nm since 20nm missed the mark While I don't know whether or not you're right on the 16/14nm guess, I will say that there is a difference between high-power and low-power processes on a given node. the fact that they can make a tiny cellphone SoC with 14nm doesn't mean they could also build a 250W GPU chip that takes up 600mm^2.
|
# ? Jun 26, 2015 23:41 |
|
Paul MaudDib posted:While I don't know whether or not you're right on the 16/14nm guess, I will say that there is a difference between high-power and low-power processes on a given node. the fact that they can make a tiny cellphone SoC with 14nm doesn't mean they could also build a 250W GPU chip that takes up 600mm^2. that's my point, we're mobile process driven and everyone seems to have skipped 20nm since only mobile cpus have the volume needed to justify such a quick and expensive transition so 14nm is gonna be the next point at which high end asic makers are ok w/ the increasingly expensive bs that it takes to make tiny transistors, so the foundries will invest in more than just mobile processes
|
# ? Jun 27, 2015 00:06 |
|
repiv posted:I think this slipped under the radar here, since both English-speaking review sites that use thermal imaging inexplicably left the cover on: Lol this sealed it for me. I was staying as positive as I could about the Fury X but this is the straw that breaks the back (again for me). I hate cards running that freaking hot Not exactly apples to apples of course, but for a little reference the effect even the "least efficient" water cooling did to the 290x penus penus penus fucked around with this message at 00:26 on Jun 27, 2015 |
# ? Jun 27, 2015 00:12 |
|
Even with a cut down chip or whatever that doesn't bode well for the Fury non-X, right? How is air going to cool it sufficiently?
|
# ? Jun 27, 2015 00:51 |
|
literally boiling the water, love it
|
# ? Jun 27, 2015 00:54 |
|
Does the water block cover the vrms? If not, then the Fury non x will be better off. If those are the temps with a water block though...
|
# ? Jun 27, 2015 00:57 |
|
THE DOG HOUSE posted:Lol this sealed it for me. I was staying as positive as I could about the Fury X but this is the straw that breaks the back (again for me). I hate cards running that freaking hot I'm gonna be laughing my rear end off if this goes the way of the 290x/390x, where unfucking the cooling makes a significant difference in the performance of the card.
|
# ? Jun 27, 2015 00:59 |
|
Fauxtool posted:literally boiling the water, love it Nah, it's under too much pressure if they did it right. Damnit AMD, this isn't what anyone meant by explosive gaming.
|
# ? Jun 27, 2015 00:59 |
|
THE DOG HOUSE posted:Does the water block cover the vrms? There is a copper tube with water flowing through it that is supposed to be going over the VRM's to cool them. Should be plenty good to do the job so something is wrong there.
|
# ? Jun 27, 2015 00:59 |
|
PC LOAD LETTER posted:There is a copper tube with water flowing through it that is supposed to be going over the VRM's to cool them. Should be plenty good to do the job so something is wrong there. :|
|
# ? Jun 27, 2015 01:05 |
|
I thought it was a really cool bit of engineering, but I guess it doesn't count if it doesn't work
|
# ? Jun 27, 2015 01:06 |
|
It does mean that we may get non-reference fury Xes
|
# ? Jun 27, 2015 01:13 |
|
What if the VRM's are specced to run at high temperature in the first place?
|
# ? Jun 27, 2015 01:17 |
|
|
# ? May 31, 2024 10:20 |
|
They certainly could, and should, be. They did that with the R9 290/X too. 100C is pretty drat hot though for any IC component. Especially one that is supposed to be water cooled.
|
# ? Jun 27, 2015 01:23 |