|
 "I know you're going to zoom in on this"
|
#
?
Apr 5, 2016 19:05
|
|
|
|
| # ? Jun 3, 2024 17:32 |
|
|
EdEddnEddy posted:I flipped 2 780's (not Ti's, those came out like a month after I bought my freaking first 780 lol) for the 1 980Ti. And the Ti is faster most of the time. Pascal chips will ship. Probably not a Titan though. Although I guess there might be some feasible die-harvests but who knows how many.
|
|
#
?
Apr 5, 2016 19:06
|
|
Paul MaudDib posted:Yup, on the slide where they're listing IBM, HP, Dell, and Cray as their partners: "between now and [Q1 2017] we expect our hyperscale partners to consume everything we can make of Tesla P100". Well, it's not like they can't produce Tesla P100s and other graphic cards parallel, especially if we assume that the new consumer cards won't use HBM2 and have a lot smaller chips. I expect them to announce new consumer graphic cards at or shortly before Computex in early May. spasticColon posted:So no Pascal cards until next year then. What about Polaris? No way Nvidia is going to not release a new range of cards this year, they are just not announcing them on their GPU computing event because it's a GPU computing event.
|
|
|
#
?
Apr 5, 2016 19:09
|
|
|
If it really was 150B transistors including the memory, you can assume that DRAM will be roughly 1:1 bits:transistors so after the 16GB/128Gb of RAM you'll have 22B left for the processor. That's still a huge chip, like 3x as large as the 22-core Broadwell Xeon so I'm a bit skeptical. Maybe there's something on the board other than the core and RAM being counted too.
|
|
#
?
Apr 5, 2016 19:14
|
|
|
Ah well, computex I suppose. Honestly after watching all that it would seem almost weird to talk about mere gaming gpus lol.
|
|
#
?
Apr 5, 2016 19:17
|
|
|
I was hoping for some new consumer level cards, but I'm glad I watched anyway, there was a bunch of pretty drat cool stuff.
|
|
|
#
?
Apr 5, 2016 19:17
|
|
|
It's only Day 1? Is there no chance of more announcements for anything else?
|
|
#
?
Apr 5, 2016 19:19
|
|
|
Man who gives a crap about industrial uses other than uber-geeks, I just want a new x80 or whatever they're called.
|
|
#
?
Apr 5, 2016 19:20
|
|
|
GaussianCopula posted:Well, it's not like they can't produce Tesla P100s and other graphic cards parallel, especially if we assume that the new consumer cards won't use HBM2 and have a lot smaller chips. I expect them to announce new consumer graphic cards at or shortly before Computex in early May. Perhaps the rumors of limited HBM2 supply were right after all. 2 full quarters is an awfully long lead time for something that's in "volume production", and I think that was about the timeframe where HBM supply was expected to ramp significantly enough.
|
|
#
?
Apr 5, 2016 19:22
|
|
|
So they made gpus learn how to drive, model photo realistic lighting in real time, render a model of mount Everest in real time, process 21 teraflops of data out of a single card, conduct "unsupervised deep learning", and create art, but here I am still annoyed I dont know when the next card is coming out for me to mildly amuse myself with watching things blow up in video games. Man *reflects on life*
|
|
#
?
Apr 5, 2016 19:25
|
|
|
Eletriarnation posted:If it really was 150B transistors including the memory, you can assume that DRAM will be roughly 1:1 bits:transistors so after the 16GB/128Gb of RAM you'll have 22B left for the processor. ROM/flash? Other random support ASICs?
|
|
#
?
Apr 5, 2016 19:25
|
|
|
Eletriarnation posted:If it really was 150B transistors including the memory, you can assume that DRAM will be roughly 1:1 bits:transistors so after the 16GB/128Gb of RAM you'll have 22B left for the processor. But can it run Crysis 3?
|
|
#
?
Apr 5, 2016 19:30
|
|
|
feedmegin posted:ROM/flash? Other random support ASICs? Cache?
|
|
#
?
Apr 5, 2016 19:33
|
|
|
EdEddnEddy posted:Cache?
|
|
#
?
Apr 5, 2016 19:34
|
|
|
God damnit I came in totally expecting to find new cards!!!
|
|
#
?
Apr 5, 2016 19:40
|
|
THE DOG HOUSE posted:So they made gpus learn how to drive, model photo realistic lighting in real time, render a model of mount Everest in real time, process 21 teraflops of data out of a single card, conduct "unsupervised deep learning", and create art, but here I am still annoyed I dont know when the next card is coming out for me to mildly amuse myself with watching things blow up in video games. Man By enjoying things blowing up in video games you supported all those amazing things.
|
|
|
#
?
Apr 5, 2016 19:41
|
|
|
The fire sale is delayed!
|
|
#
?
Apr 5, 2016 19:41
|
|
|
SRAM is usually 6 transistors per bit. I don't know what kind of cache GPUs typically work with but going back to the 22-core Broadwell Xeon, it has 55MB of cache which would be 440Mb and ~2.6B transistors. I'm sure that some substantial portion of what's left after you take out the DRAM is cache, but it still has to fit on the same die (unless Nvidia is pulling a Pentium II) and I don't think they have 22B transistors in one die.feedmegin posted:ROM/flash? Other random support ASICs? Could be anything I guess, but the ROM is going to be insignificant unless that thing has a huge firmware image. I just downloaded a 980Ti BIOS out of curiosity and it's only 222KB. Eletriarnation fucked around with this message at 19:45 on Apr 5, 2016 |
|
#
?
Apr 5, 2016 19:42
|
|
|
I know its all speculation at best, if we see consumer cards announced at computex in May/June, how long is the gap between announcement and availability usually? I'm not super familiar with these things but I'd guess given we get an announcement then we'd see cards in August/September?
|
|
#
?
Apr 5, 2016 19:46
|
|
|
The actual GPU has 15.3 Billion transistors and they even managed to bump up the boost clocks by a considerable amount too. The P100 boost clock is as high as a Maxwell card at maximum overclock. https://devblogs.nvidia.com/parallelforall/inside-pascal/  
|
|
#
?
Apr 5, 2016 19:49
|
|
|
I would give them my money if I could...
|
|
#
?
Apr 5, 2016 19:53
|
|
|
Bleh Maestro posted:I would give them my money if I could...
|
|
#
?
Apr 5, 2016 19:54
|
|
|
MaxxBot posted:The actual GPU has 15.3 Billion transistors and they even managed to bump up the boost clocks by a considerable amount too. The P100 boost clock is as high as a Maxwell card at maximum overclock. Yeah, I just realized that 128Gb isn't 128B bits - it's 2^37 bits, which is ~137B. This plus a 15B transistor core puts us at the nice round 150B transistor number. Ed: 37th power, not 27th. Eletriarnation fucked around with this message at 20:08 on Apr 5, 2016 |
|
#
?
Apr 5, 2016 20:02
|
|
|
Paul MaudDib posted:Yeah, we'll get Pascal but it'll probably be GP104 and GP106. That and a lolhuge chip on the leading edge of an architecture on a node that's only been used for mobile. It's pretty much in line with Vega and NV working really hard to get it out early. Only Intel has a similar sized chip on a similar sized node.
|
|
#
?
Apr 5, 2016 20:38
|
|
|
RME posted:I know its all speculation at best, if we see consumer cards announced at computex in May/June, how long is the gap between announcement and availability usually? I'm not super familiar with these things but I'd guess given we get an announcement then we'd see cards in August/September? From the official launch date of the 970 / 980 there was practically no gap at all. It was a very good launch, and while availability was scarce - it was not a paper launch. If I recall there were rumors mere weeks leading up to it. And it was available like the next day after the announcement. However I wouldn't count on that just because it happened once. That was pretty rare, but I couldn't guess how rare. The most striking thing about the 970/980 release was how sudden it was though. It was a mere rumor one day then something you could buy the next day.
|
|
#
?
Apr 5, 2016 21:17
|
|
|
Is Q2 2017 supposed to be consumer availability? This is the softest of paper launches otherwise. Little bit disappointed that FirePro Fiji beats this card in raw output. Also won't going big first delay smaller chips as well, since TSMC only has so much silicon? I get why Nvidia is doing what they are doing, but they're kind of leaving a door open for AMD here.
|
|
#
?
Apr 5, 2016 21:20
|
|
|
There's no way AMD launches major new products in a wide variety of price ranges (not more rebrands) this spring/summer and Nvidia doesn't... right?
|
|
#
?
Apr 5, 2016 21:25
|
|
|
FaustianQ posted:Is Q2 2017 supposed to be consumer availability? This is the softest of paper launches otherwise. Little bit disappointed that FirePro Fiji beats this card in raw output. AMD has not been making many noises that indicate an imminent launch of a next gen chip. Nvidia arguably knows more about it than the general public, because researching your competitors is generally good business, so if they have chosen the top end route for this launch, I wouldn't expect them believe AMD could undercut them and launch a new architecture at the consumer level. AMD may also be stymied by the weird deals worked out re: HBM2 access, that precludes them from getting it first vs Nvidia.
|
|
#
?
Apr 5, 2016 21:35
|
|
|
beejay posted:There's no way AMD launches major new products in a wide variety of price ranges (not more rebrands) this spring/summer and Nvidia doesn't... right? I think the last node shrink NV was first to was 65 nm with the 8800 GT. FaustianQ posted:Little bit disappointed that FirePro Fiji beats this card in raw output. The entire point of FirePro Fiji is that it's a ludicrous output monster in one regard for small data sets. It is exceedingly good at that but it's not a huge niche. However great inside a niche is great for AMD because that puts them far enough ahead to get a win. And maybe AMD does have a solid opening in the lower end. That'd be cool (and interesting for AMD to have the advantage in mobile). EoRaptor posted:AMD has not been making many noises that indicate an imminent launch of a next gen chip. Nvidia arguably knows more about it than the general public, because researching your competitors is generally good business, so if they have chosen the top end route for this launch, I wouldn't expect them believe AMD could undercut them and launch a new architecture at the consumer level. Isn't it pretty firm that AMD's launching Polaris this year and then Vega early next year and Vega's the parts with HBM2? I also wouldn't be surprised if NV thought that fighting off Knights Landing was more important than fighting off AMD. xthetenth fucked around with this message at 21:38 on Apr 5, 2016 |
|
#
?
Apr 5, 2016 21:36
|
|
|
EoRaptor posted:AMD has not been making many noises that indicate an imminent launch of a next gen chip. Nvidia arguably knows more about it than the general public, because researching your competitors is generally good business, so if they have chosen the top end route for this launch, I wouldn't expect them believe AMD could undercut them and launch a new architecture at the consumer level. Well, I was speaking more at consumer and workstation level with P11 and P10, not so much big chip HPC performance. However AMD has alluded to Vega being a Q1 2017 launch product and Nvidia is saying their product is only after Q1. Doubt that's enough time to make a difference, and most rumors of Vega make it an average 300-350mm² and not a staggering 610mm² compared to P100. Still, if producing P100 delays GP104 and 106 long enough, that will hurt Nvidia.
|
|
#
?
Apr 5, 2016 21:44
|
|
|
beejay posted:There's no way AMD launches major new products in a wide variety of price ranges (not more rebrands) this spring/summer and Nvidia doesn't... right? I'd hope so, they've had 3 years to figure it out
|
|
#
?
Apr 5, 2016 21:50
|
|
|
Sadly, I was on a plane during the show. Good to hear that Jen-Hsun hasn't lost his "touch". The 5.3TF / 1 TB/s numbers make KNL (~3TF / 500 GB/s) look puny. Paul MaudDib posted:Yup, on the slide where they're listing IBM, HP, Dell, and Cray as their partners: "between now and [Q1 2017] we expect our hyperscale partners to consume everything we can make of Tesla P100". Welp, I shouldn't have waited to upgrade. I guess I'll have to settle for running kernels on the ones in the lab. 
|
|
#
?
Apr 5, 2016 21:59
|
|
|
I guess the good news for me is that I can expect a lot of Tesla cards to show up on Ebay and I'll be able to try them out at home for a lot cheaper than the hundreds of GPU / CPU hours I'd have to burn on AWS to do some meaningful experiments with my deep learning hobby. It should also mean that I should sell my 970 GTX since I won't have the time to game really until 2017 while I dick around with that. If I want to give these new cards a try, I'm sure I'll be able to get a crack at them via AWS for maybe $.70 / hr compute time. Hopefully by then I'll have figured out how to train my models efficiently enough I won't have to spend $4k to do it in the first place. My GPGPU boner is pretty raging hard.
|
|
#
?
Apr 5, 2016 22:11
|
|
|
xthetenth posted:I also wouldn't be surprised if NV thought that fighting off Knights Landing was more important than fighting off AMD. necrobobsledder posted:I guess the good news for me is that I can expect a lot of Tesla cards to show up on Ebay and I'll be able to try them out at home for a lot cheaper than the hundreds of GPU / CPU hours I'd have to burn on AWS to do some meaningful experiments with my deep learning hobby. It should also mean that I should sell my 970 GTX since I won't have the time to game really until 2017 while I dick around with that. If I want to give these new cards a try, I'm sure I'll be able to get a crack at them via AWS for maybe $.70 / hr compute time. Hopefully by then I'll have figured out how to train my models efficiently enough I won't have to spend $4k to do it in the first place. 
|
|
#
?
Apr 5, 2016 22:12
|
|
|
mobby_6kl posted:^^^ drat, it'll be nice if HBM can keep considerably shrinking cards down while maintaining (or exceeding) previous gen performance. Getting kinda tired of video cards that are a foot long with 2-3 fans, especially when tooling around with smaller form builds.
|
|
#
?
Apr 5, 2016 22:12
|
|
|
I would be shocked if Nvidia didn't release a new card in all of 2016, it might be a long rear end time until we get consumer big Pascal with HBM but for them to not release anything at all this year would be way out of step with their usual release schedule.
|
|
#
?
Apr 5, 2016 22:29
|
|
|
Anime Schoolgirl posted:I dunno if that's the case, given the history of the Knights series... Yeah, and with P100, it's going to be fighting uphill this generation as well. However, if it didn't have a big GPU competitor... MaxxBot posted:I would be shocked if Nvidia didn't release a new card in all of 2016, it might be a long rear end time until we get consumer big Pascal with HBM but for them to not release anything at all this year would be way out of step with their usual release schedule. I expect a similar cadence to AMD. Maybe three cards soon and the biggest next year, maybe two and two.
|
|
#
?
Apr 5, 2016 22:37
|
|
|
At this point I'm gonna spend my money on an X34 or XB270HU instead, and maybe trade in my 780 Ti's for a single 980 Ti before depreciation takes over.
|
|
#
?
Apr 5, 2016 22:37
|
|
|
Paul MaudDib posted:At this point I'm gonna spend my money on an X34 or XB270HU instead, and maybe trade in my 780 Ti's for a single 980 Ti before depreciation takes over. Might want to look into the possibility of flashing a MAC compatible Bios onto them as well. The guy that bought my 780s on Ebay runs an MacGPU site and looks like he just flashes MAC compatible Bios's onto them and sells them at a huge markup.
|
|
#
?
Apr 5, 2016 22:43
|
|
|
|
| # ? Jun 3, 2024 17:32 |
|
|
EdEddnEddy posted:Might want to look into the possibility of flashing a MAC compatible Bios onto them as well. The guy that bought my 780s on Ebay runs an MacGPU site and looks like he just flashes MAC compatible Bios's onto them and sells them at a huge markup. Selling PC things to Mac users at huge markups is after all a tried and true Apple approved method, too 
|
|
#
?
Apr 5, 2016 22:48
|
|