|
http://cdn.overclock.net/6/67/677d7a15_P1100270-Copy_w_600.jpeg
|
# ? Jun 3, 2015 06:57 |
|
|
# ? May 9, 2024 19:35 |
|
Paul MaudDib posted:In particular the Titan X is a kick-rear end card for neural net stuff - you don't need double-precision for that stuff Honestly, if there were tiny-precision FP8 mode, we'd use it.
|
# ? Jun 3, 2015 07:05 |
|
Subjunctive posted:Honestly, if there were tiny-precision FP8 mode, we'd use it. it's weird how quickly this neural net stuff became a thing. I guess it makes sense, but man, I don't think anyone saw this coming back in the Fermi timeframe.
|
# ? Jun 3, 2015 07:08 |
|
Professor Science posted:hence why they're talking FP16 for Pascal. I'm biased, but from my perspective LeCun blew the doors off this stuff, and now everyone is smacking nails with the low-precision convolutional hammer. NVIDIA's loving it, based on their GTC keynote content.
|
# ? Jun 3, 2015 07:10 |
|
Their page looks really hype, but with all the rumours recently I just can't get hyped for anything they say. HBM is probably going to be the biggest leap 3d has seen in a long time, but if it's not faster yet for whatever reason, nobody will buy it. That said, will HBM help with VR at all? It's lower latency than GDDR 5, right? Surely that should help some.
|
# ? Jun 3, 2015 07:26 |
|
Etrips posted:How much of a difference is there between reference vs non-reference cards? In most cases the cards are the same or broadly similar. You see some like the lightning, classified, kingpin and iirc G4 gaming that have a ton of power phases, sometimes binned chips and so on that are there for the crowd that has to have huge overclocks. They generally aren't that great on the price/performance scale. The main thing reference and non reference means is what type of cooler the card has. Reference cards have to be compatible and play nice in all cases, including stuff like 4-way. That pretty much means those cards' heat has to go away from those cards. That means blowers with full shrouds. Those eject hot air out of the case, which is good. The problem is they use a single fan, smallish heatsinks and generally run hot and loud. The canonical example is the R9 290. That was hot and loud to the point that it had to throttle its clock to avoid overheating. Generally an aftermarket R9 290 performs like the reference 290X, it's higher priced cousin. Even with the best blower around the Titan X gets loud and relatively hot. Aftermarket cards come with different coolers. The card makers come up with them to differentiate their products. Generally they're open coolers with two or three fans blowing into their heat sink. For example the Gigabyte windforce cooler like the one on my GTX 970 has three 100mm fans and makes the card over thirteen inches long. These heatsinks generally do a much better job keeping the card cool and quiet. However the heat doesn't leave the case on its own. Unless you know you have good airflow in your case, more than one or two cards and that type of cooler is going to minimize any advantage they hold and maybe worse. Finally there's the cards set up for water cooling. Evga has its hydro copper cards designed to slot into a custom liquid cooling set, and the Zotac 980ti has an optional similar setup as well as beefy air on the card itself. These can be very cool and very quiet at once but you know you're in the demographic for custom loop cards already. The other is cards with an integrated liquid loop for the card itself. They're like similar CPU closed loop coolers in that they offer most of the upside of custom water and basically none of the work. The main candidates there are the R9 295, which generates something like twice the heat of a Titan and is cooler and quieter under heavy load and shows how well liquid can eat heat, and the evga 980 hybrid (comes in a kit that can fit a titan x too), which shows how much room it can open for a huge overclock. In one review it allowed a big enough overclock that compared to an overclocked reference model it just about paid for the $100 cooler. Why the hell did I write that on a phone. Also re: HBM, it should start getting real good in the long run, but we'll see if it paid of now like getting GDDR5 first did. Honestly where it should work miracles is APUs. Those things are painfully bandwidth starved. xthetenth fucked around with this message at 07:42 on Jun 3, 2015 |
# ? Jun 3, 2015 07:39 |
|
Professor Science posted:depends on arithmetic intensity of each computation of course, but I can't imagine it working well. latency of PCIe plus waking up the CPU and the various cache flushes on the GPU side (I THINK you'd have to flush L2 every time you want to send data to the CPU from within a kernel, I don't know if mapped pinned memory bypasses L2 on Maxwell or not) would put a serious damper on it. it'd probably be better to have a processing pass, keep track of exceptional locations where you actually need DP (if it's not needed at every location), then do the fixup on the CPU side, pipelining all of this with PCIe and probably triple-buffering. Where is zero-copy mapped - into the SMX/shared memory or into global memory? I think that's the core question to that first half. If it's the SMX then latency should be fine. But yeah, tracking error extrema might be the way to go. Latency of CPU wakeup shouldn't be an issue - if necessary you can spinlock it at an appropriate interval (I'd guess checking at 2-4x the rate of effective updates, ala Nyquist rate). e: It's been a while and I'm way out of this, does CUDA support multi-threaded dispatch (per device, per stream)? If not it might be advantageous to allocate a spinlocked thread per device-timeslice or SMX-timeslice anyway. Paul MaudDib fucked around with this message at 08:05 on Jun 3, 2015 |
# ? Jun 3, 2015 07:54 |
|
Holy poo poo, thanks for the long phone post write up. It was very informative. I think I will wait for a non-reference card by ASUS or MSI.
|
# ? Jun 3, 2015 08:13 |
|
Would buying a gtx 980 (not ti) be a terrible idea now? I just want a blower design card and the blowers on gtx 970 seem to either suck (evga) or have no information online about them (Asus turbo gtx 970). Kind of want to keep it under 500 bucks.
|
# ? Jun 3, 2015 08:30 |
|
Decided to do another check on R9 290 stock: this one was from early May: Looks like there has been significant draw down in stock but still plenty of supply out there that they can't get rid of before the 390/X relaunch. Releasing a rebranded 290/X as a 390/X will probably result in a nice price drop on their 'older' hardware. If you've got the cash to buy a 980 or 980Ti you won't care but some decent bargains could pop up.
|
# ? Jun 3, 2015 08:30 |
Thelonious Monk posted:Would buying a gtx 980 (not ti) be a terrible idea now? I just want a blower design card and the blowers on gtx 970 seem to either suck (evga) or have no information online about them (Asus turbo gtx 970). Kind of want to keep it under 500 bucks. I'd say a 980 at this point is a bad idea, when you are already spending $500 the $150 bump to a 980Ti is not much to pay for the extra performance it can give you. A 970 is also a great choice because it's very close to the 980 for a good chunk less money. Almost all of the blower designs are going to be the same as the reference one which is a perfectly good one for a blower, most people just preference an open shroud design over a bower because their setup works well for that type of cooler. Really you should be able to buy any of the bower cooled 970s and get perfectly good performance out of it, just not much head room for OCing and it will be a bit louder.
|
|
# ? Jun 3, 2015 08:40 |
|
AVeryLargeRadish posted:I'd say a 980 at this point is a bad idea, when you are already spending $500 the $150 bump to a 980Ti is not much to pay for the extra performance it can give you. If you're TDP-constrained, that might not be the case. The TDP jump from 980 to 980 Ti is substantial, and could definitely tip a given configuration from feasible to infeasible. E: actually, in that case an overclocked 970 might still be better, hmm.
|
# ? Jun 3, 2015 08:51 |
|
Subjunctive posted:If you're TDP-constrained, that might not be the case. The TDP jump from 980 to 980 Ti is substantial, and could definitely tip a given configuration from feasible to infeasible. Thelonious Monk posted:Would buying a gtx 980 (not ti) be a terrible idea now?
|
# ? Jun 3, 2015 08:52 |
|
If only all the retailers in my country weren't ripoffs. I paid 600€ for a non G1 Windforce 980 (all the other manufacturers were out of stock) about a month ago, now they offer Ti for 800€ and thats for the reference one, definitely outside my budged even if I sold the 980 or didn't get it in the first place at all.
|
# ? Jun 3, 2015 09:21 |
|
So from what I can gather, AMD is going for shear power while Nvidia is going more for max efficiency? Which then allows Nvidia to power up even more if the situation calls for it? Right? So AMD is Vegeta and Nvidia is Goku right after the hyperbolic time chamber?
|
# ? Jun 3, 2015 09:58 |
|
West007 posted:If only all the retailers in my country weren't ripoffs. Same situation in most of Europe. It is not so much the retailers but more a combination of VAT and the € dropping off a cliff.
|
# ? Jun 3, 2015 10:15 |
|
Star War Sex Parrot posted:Please stop trying to crush my mITX dreams. Silverstone had introduced a 700w sfx psu, you're gonna be fine.
|
# ? Jun 3, 2015 10:35 |
|
Hamburger Test posted:Same situation in most of Europe. It is not so much the retailers but more a combination of VAT and the € dropping off a cliff. Mostly VAT, 1 EUR is still 1.11 USD.
|
# ? Jun 3, 2015 10:37 |
|
Riso posted:Mostly VAT, 1 EUR is still 1.11 USD. It was close to 1.4 one year ago.
|
# ? Jun 3, 2015 10:49 |
|
Don Lapre posted:
If that's the GPU + the RAM, it's smaller than the area taken by GPU + RAM on any other card. Truga posted:Their page looks really hype, but with all the rumours recently I just can't get hyped for anything they say. HBM is probably going to be the biggest leap 3d has seen in a long time, but if it's not faster yet for whatever reason, nobody will buy it. Here is the link to their page: http://www.amd.com/en-us/products/graphics/graphics-technology June 16, 9am PST. However I think it's definitely gonna be a paper launch.
|
# ? Jun 3, 2015 12:29 |
|
Paul MaudDib posted:Same - a gifted HP Z400 workstation had a "450W" PSU that could be over-currented by an R9 280. It had an 80+ rating and everything. (Ask) me about the pile of poo poo tier HP 8300s sitting across from my desk in our it hallway with power supplies that burst in to flame! (Or at least decided to let the magic blue smoke go free)
|
# ? Jun 3, 2015 13:23 |
|
Paul MaudDib posted:In particular the Titan X is a kick-rear end card for neural net stuff - you don't need double-precision for that stuff, and latency doesn't matter because you can queue up a billion threads and wait it out. My Master's thesis was a simulation done in single-precision, as was the source program. It's not universally applicable but I think a lot of people overestimate the need for double precision nowadays, just because it's there. If you can get away with floats (and for most applications you can, especially with some effort) then 12GB for $1k is a goddamn steal - a K40 with 12Gb memory will set you back $3500. That's why I'm thinking about shelling out the extra $350 for the Titan X. Doubling my working set for $350 is a no brainer, even if I have to spend another $350 down the line if I SLI. Here ya go I just checked the heavy load test mode option. Should i be running a load on it?
|
# ? Jun 3, 2015 13:50 |
|
I'd be real surprised if the 300 series were a paper launch. It doesn't match well with straight rebrands. For that matter, if it's straight rebrands why the hell did they wait this long selling 200 series cards rather than printing off stickers and new boxes.Thelonious Monk posted:Would buying a gtx 980 (not ti) be a terrible idea now? I just want a blower design card and the blowers on gtx 970 seem to either suck (evga) or have no information online about them (Asus turbo gtx 970). Kind of want to keep it under 500 bucks. I personally wouldn't. The 980 is a lot more expensive than the 970 and doesn't perform much better, while being not much less than. 980ti with much less performance. Best buy used to sell nvidia branded reference cooled 970s, might be worth a check. Otherwise might I ask why you want a blower, and will you be swayed by benchmarks showing aftermarket open coolers being much cooler and quieter? Is there a specific build reason for a blower like trying to fit things into a small case?
|
# ? Jun 3, 2015 14:03 |
|
Assuming prices on the 980 drop even more, from AMD releases or whatever, which brands have the best custom coolers?
|
# ? Jun 3, 2015 14:24 |
|
Will a 760 4gb be able to run games at 1080 p with med-high graphics for awhile yet? Was considering getting a 970 but I may wait for awhile longer if the 760 is able to stay relevant
|
# ? Jun 3, 2015 14:24 |
|
Ak Gara posted:So from what I can gather, AMD is going for shear power while Nvidia is going more for max efficiency? Which then allows Nvidia to power up even more if the situation calls for it? Right? Does this mean that Nvida and Amd are going to fuse via the potara earrings to become Nvimd at some point? As a 970 haver this thread is really making me want a 980ti, even though its total overkill for 1080p and it probably wouldnt work with a 550w psu? I'm still getting framedrops in ets2 at 200% scaling....
|
# ? Jun 3, 2015 14:41 |
|
If you're cool spending that much more within the same generation for a card that likely won't let you hold out for another generation, go for it, but I doubt that 980ti will last that much longer. I'm worried probably more than is rational about the 970's longevity but I'm holding out for the HBM2 cards, at which point I'm going to flip my 970 when it should still have pretty good value. A 970 or a 290(X) are definitely the top price/performance buys of the generation, especially for people who got in pretty soon after the 970 hit. Edit: Looks like a 390X has been spotted with a hybrid cooler. Very nice. If it behaves anything like Hawaii, better cooling should be handy to drop the wattage a bit and really lower the noise and temps, as well as get the heat out of the case. xthetenth fucked around with this message at 15:27 on Jun 3, 2015 |
# ? Jun 3, 2015 14:55 |
|
track day bro! posted:Does this mean that Nvida and Amd are going to fuse via the potara earrings to become Nvimd at some point? AMD is probably going to join up with an evil corporation in order to beat up Nvidia at some point. Intel's inbuilt HD Graphics 4000 thing is pretty drat powerful. Maybe AMD's gpu's could come with a big Intel logo on them?
|
# ? Jun 3, 2015 15:41 |
|
Ak Gara posted:AMD is probably going to join up with an evil corporation in order to beat up Nvidia at some point. It takes an Iris pro part to really compete with an AMD APU. On the other hand that's a lot of catching up Intel has done.
|
# ? Jun 3, 2015 15:48 |
|
The early Broadwell-desktop benchmarks show it competing quite well against AMD's APUs, I think.
|
# ? Jun 3, 2015 15:52 |
|
Of course, HBM could allow for near-SoC sizes out of APUs with incredible performance. Provided their TDP doesn't go too high, AMD could produce the killer platform for light computers with good performance and moderate graphics capabilities with four gigs of HBM on die for shared system/video RAM. Sounds like the perfect office box or media center PC to me.
|
# ? Jun 3, 2015 16:01 |
|
Subjunctive posted:The early Broadwell-desktop benchmarks show it competing quite well against AMD's APUs, I think. Yeah, all but one are Iris Pro iirc and that's a big part that performs well. I think that integrated graphics are starting to get to where they could really use higher bandwidth memory.
|
# ? Jun 3, 2015 16:02 |
|
xthetenth posted:On the other hand that's a lot of catching up Intel has done. Yep. I got myself a google pixel (has i7-5500U) on sometime early May. I decided to install steam and check out if I can run any games on it, expecting to maybe be playable in ancient. Turns out, it runs KSP and Guns of Icarus Online, as long as I halve the resolution and play at 1280x850, at a steady 20-30 FPS. It's not stellar, but it's more than playable for the kinds of games I'd bother to play on a laptop. Used to be one could just *barely* run minecraft on intel gpus not long ago, now it runs unity games just fine. I was honestly surprised. On the other hand: Angry Fish posted:I really think Ars Technica really did a good job explaining AMDs new mobile platform: The future for AMD may not be as bleak as it looks right now. Who knows
|
# ? Jun 3, 2015 16:54 |
|
Yeah, Carrizo is a good looking part focusing on a market segment that should hopefully pay off pretty well. Really good and power efficient move playback is a neat feature for a low power laptop to have.
|
# ? Jun 3, 2015 17:04 |
|
xthetenth posted:Yeah, Carrizo is a good looking part focusing on a market segment that should hopefully pay off pretty well. Really good and power efficient move playback is a neat feature for a low power laptop to have. I'm skeptical, because the problems with $200-600 laptops have almost nothing to do with CPU and everything to do with everything else. Terrible screens, garbage hinges, keyboards, and trackpads. If AMD can really shave $200 off the cost of a CPU and let me have a Thinkpad with a 1080p IPS screen, 5ghz wifi, and decent keyboard for $600, they will be my heroes.
|
# ? Jun 3, 2015 17:06 |
|
Twerk from Home posted:I'm skeptical, because the problems with $200-600 laptops have almost nothing to do with CPU and everything to do with everything else. Terrible screens, garbage hinges, keyboards, and trackpads. If AMD can really shave $200 off the cost of a CPU and let me have a Thinkpad with a 1080p IPS screen, 5ghz wifi, and decent keyboard for $600, they will be my heroes. That's true enough. If they can manage to get the price down to where an fhd ips can happen, that would be a hell of a product for anyone who watches video. Honestly I just want viable products so AMD stays in it, and that doesn't require making everything on those machines better.
|
# ? Jun 3, 2015 17:11 |
|
Yeah, I really do want AMD to remain competitive. If they weren't refreshing their GPU lineup this month, I am sure nvidia would have been happy enough to milk the TitanX a bit longer, or to up the price on the 980ti. Competition is a good thing. Given the specs, it is possible that the Fiji cards could end up slightly faster than a 980ti, provided that the drivers are as well optimized as nvidia's. In reality though, we know that probably won't be the case.
|
# ? Jun 3, 2015 17:26 |
|
Filthy Monkey posted:Yeah, I really do want AMD to remain competitive. If they weren't refreshing their GPU lineup this month, I am sure nvidia would have been happy enough to milk the TitanX a bit longer, or to up the price on the 980ti. Competition is a good thing. 4k shaders is a lot. It could very well be big. And optimized drivers is an interesting subject. Compare 780ti vs 290X at launch and now. GCN aged a hell of a lot better than Kepler, and I think a good bit was the gains it got with the omega driver.
|
# ? Jun 3, 2015 17:40 |
|
Here are some pictures of what is believed to be a R9 390X (NOT Fiji) with a hybrid cooler: http://videocardz.com/56096/powercolor-radeon-r9-390x-devil-pictured
|
# ? Jun 3, 2015 18:07 |
|
|
# ? May 9, 2024 19:35 |
|
xthetenth posted:Compare 780ti vs 290X at launch and now. GCN aged a hell of a lot better than Kepler, and I think a good bit was the gains it got with the omega driver. By now, do you mean right now? The driver released on Monday fixed a big regression with Kepler, some games are running significantly faster now.
|
# ? Jun 3, 2015 18:15 |