Idk, based on the rumors and what we know about HBM and the interposer they are using, I think it will be a hard launch, but I also think that the supply will be low at the start. If the pricing is as high as some of the rumors claim, I guess it will deter some people from buying though. Had this card launched a month or two ago, I probably would have bought one. But now I figure I might as well wait until 16nm. I don't think HBM itself is going to have a major impact on performance. There doesn't seem to be many games hitting the bandwidth cap on current cards, even at 4K.

Do this for us and yourself. Downclock your card GPU and memory to 1/2 of currently stable clock you use. Benchmark/bench in game. Then return to normal just memory clock (keep GPU at 1/2) to see if doubling memory bandwidth has effect in comparison to having balanced vram+GPU clocks. I am sure that double bandwidth will show good performance increase even without increasing GPU clocks/beefing up GPU.

Do this for us and yourself. Downclock your card GPU and memory to 1/2 of currently stable clock you use. Benchmark/bench in game. Then return to normal just memory clock (keep GPU at 1/2) to see if doubling memory bandwidth has effect in comparison to having balanced vram+GPU clocks. I am sure that double bandwidth will show good performance increase even without increasing GPU clocks/beefing up GPU.

Huh? why would I downclock to half? And how would this prove/disprove diminishing returns on memory bandwidth? If a game requires 200gb/s of bandwidth, increasing my card from 224 to 448 isn't going to show an improvement in performance. In your scenario I'm going from 112 to 224. Of course it's going to have an effect, regardless to the core clock speed. Obviously they should be balanced, but that's my point -- they already are balanced and above the threshold for most games. Regardless, if Nvidia/AMD felt like memory bandwidth was a concern, they could easily be shipping GPU's with a larger memory bus than they do now. Both companies have shipped cards with a 512 bit interface. On a Titan X, die size might be a concern, but Nvidia could easily fit a 512 bit bus on a 980. They didn't because it's simply unnecessary. Now HBM is a bit different, you get some power savings by switching to it, along with a doubling in available bandwidth. I did the math in a previous post (based on http://www.cs.utah.edu/thememoryforum/mike.pdf), but essentially it works out to about a ~18-20w savings over using a GDDR5 interface. Which is pretty good, because that 20w savings is going to go directly into improving core performance (40% increase in shaders according to rumors). But unless games start using more VRAM, the bandwidth increase will not bring significant performance improvements directly, it will come indirectly from the power savings. I'm not saying HBM is unnecessary, but given the obvious delays caused by it, I don't think it was worth the trouble for this particular card. Only the power savings is worth having, the doubling in memory bandwidth is unnecessary, in my opinion.

Since my GTX 680 runs GTA 5 just fine, I don't see any point in upgrading anytime soon :| Will the 390 series finally deliver more than 100% more performance at less than I paid for the 680? If so, they may have a sale. But so far, no sale to anyone.

Huh? why would I downclock to half? And how would this prove/disprove diminishing returns on memory bandwidth? If a game requires 200gb/s of bandwidth, increasing my card from 224 to 448 isn't going to show an improvement in performance. In your scenario I'm going from 112 to 224. Of course it's going to have an effect, regardless to the core clock speed. Obviously they should be balanced, but that's my point -- they already are balanced and above the threshold for most games. Regardless, if Nvidia/AMD felt like memory bandwidth was a concern, they could easily be shipping GPU's with a larger memory bus than they do now. Both companies have shipped cards with a 512 bit interface. On a Titan X, die size might be a concern, but Nvidia could easily fit a 512 bit bus on a 980. They didn't because it's simply unnecessary. Now HBM is a bit different, you get some power savings by switching to it, along with a doubling in available bandwidth. I did the math in a previous post (based on http://www.cs.utah.edu/thememoryforum/mike.pdf), but essentially it works out to about a ~18-20w savings over using a GDDR5 interface. Which is pretty good, because that 20w savings is going to go directly into improving core performance (40% increase in shaders according to rumors). But unless games start using more VRAM, the bandwidth increase will not bring significant performance improvements directly, it will come indirectly from the power savings. I'm not saying HBM is unnecessary, but given the obvious delays caused by it, I don't think it was worth the trouble for this particular card. Only the power savings is worth having, the doubling in memory bandwidth is unnecessary, in my opinion.

At clocks your card is, it is balanced, isn't it? If you downclock both to 1/2 will it still be balanced? I think it will be. So as you bench balanced 1/2 clock and then 1/2 GPU + 1x vram clock, you will measure how much performance you can gain even with diminishing returns. It can give hint of situation like: AMD takes 290x and slaps on it HBM controller+memory and doubles memory bandwidth... Will it benefit performance wise? And how much? I'll do it on my card in 2 hours. Probably on firestrike/metro/AvP benchmarks. Yeah, and why I wrote about 1/2 clocks scenario? Because you have no way of doubling memory bandwidth ratio if you keep GPU at standard clock.

The rumor has been June, and still is June. Patience.

At clocks your card is, it is balanced, isn't it? If you downclock both to 1/2 will it still be balanced? I think it will be. So as you bench balanced 1/2 clock and then 1/2 GPU + 1x vram clock, you will measure how much performance you can gain even with diminishing returns. It can give hint of situation like: AMD takes 290x and slaps on it HBM controller+memory and doubles memory bandwidth... Will it benefit performance wise? And how much? I'll do it on my card in 2 hours. Probably on firestrike/metro/AvP benchmarks. Yeah, and why I wrote about 1/2 clocks scenario? Because you have no way of doubling memory bandwidth ratio if you keep GPU at standard clock.

Yeah I guess I agree with the logic in your test, I feel like it's overly simplified, but I'd definitely be interested to see the results. I'm at work, but I can try testing later. In my mind though, if memory bandwidth is balanced on current cards (which I think it is) then doubling the bandwidth of memory shouldn't have a significant impact on performance at all, unless it's a game that specifically gets impacted by a memory bottleneck. In my experience, overclocking GDDR5 on more recent cards does little to nothing even despite 10-15% increases. I also think that if Nvidia felt like doubling bandwidth was worth it, they could have easily put a 384/512bit bus on a 980. Again, don't get me wrong, I think HBM is the future and I think it's worth switching too eventually -- but clearly they having problems manufacturing chips with it. I'm not sure it was worth the delay or potential rumored shortages after launch.

Can maxwell go that low in the clocks? On my cards lowest I can go is -105 on the core and down to 6ghz on the vram in MSI AB.

I can patiently wait a few weeks. Nvidia did a good job of ignoring Kepler, so I'm interested to see what this 390x will be like.

Just like AMD ignored TeraScale3 when they released GCN.

Yeah I guess I agree with the logic in your test, I feel like it's overly simplified, but I'd definitely be interested to see the results. I'm at work, but I can try testing later. In my mind though, if memory bandwidth is balanced on current cards (which I think it is) then doubling the bandwidth of memory shouldn't have a significant impact on performance at all, unless it's a game that specifically gets impacted by a memory bottleneck. In my experience, overclocking GDDR5 on more recent cards does little to nothing even despite 10-15% increases. I also think that if Nvidia felt like doubling bandwidth was worth it, they could have easily put a 384/512bit bus on a 980. Again, don't get me wrong, I think HBM is the future and I think it's worth switching too eventually -- but clearly they having problems manufacturing chips with it. I'm not sure it was worth the delay or potential rumored shortages after launch.

So, I clocked it 450/685 vs 450/1370. Firestrike could not be measured because driver ignored max clocks and run on default 925/1375. (probably something in driver) But Metro+AvP held clocks way I wanted. Gain was actually smaller than I thought, It was 8.4% in metro2033 and 5.7% in AvP. If you can, try same thing. Maybe nVidia's architecture can benefit more from bandwidth, so we may see bigger gains in pascal from HBM2. Edit: I did run both on maximum details 1080p, so maybe there was too big load on shaders than on texturing. +metro was choked by physX. But in the end, that is how people usually run games 🙂 Edit2: Metro without Advanced PhysX shown 9.3% gain. Edit3: But DoW2:Retribution only 1.9% gain. Edit4: Conclusion. There is performance to be extracted, let's say up to 10% on HD7970. R9-290x has: 37.5% more shaders, 37.5% more TMUs, 100% more ROPs, but only 21% higher memory bandwidth, so I would say there may be more performance to extract. (benchmarks show that it is 20~45% better so there are cases where it needs higher bandwidth since chip itself is in total more than 40% better) Considering speculations about r9-390x stats. (if true that they are at least double of HD7970) It will not be memory bandwidth limited and therefore perform around 2x 7970 * 1.1 ~= 2.2x HD7970 * GPU clock percent difference or simply 2.2x HD7970 GHz if core is clocked @ 1050Mhz. And jump from r9-290x may be higher than plain shader difference which is 45%.

Lol, when peoples will stop to take rumors written by exemplary jurnalists on none less exemplary sites ( starting by F or W ) as facts.... they will maybe stop to speak about delay, or whatever.. - Computex of Taipei end the 8/6 2015 - E3 2015 start the the 16/6 2015.

This is getting ridiculous.

Yep and now it'll only hurt AMD if it's not performing at 110% over previous generations and makes Titan X look like a 960 (exaggerations aplenty there). The hype and rumours can only support a card for so long and now it'll be hit with all the crazy expectations if it doesn't deliver. I really hope it'll be a good card and force Nvidia to reduce the stupid pricing scheme they've been rustling the last couple of years but I just can't see it otherwise I'm sure they'd be showing this off as much as possible. Typically when things are constantly delayed interest wanes and the expectations increase. Once it fails to deliver these expectations it gets hit harder than it would before. This applies to anything and isn't aimed at AMD or the X390 in particular it's just what I've seen over the years with many different products including games and hardware.

Wasnt this already clear that they will reveal new cards at E3?? If computex was after E3 they could have cards there too but it isnt. (computex 1.june and e3 18.june) More painfull will be the info about availability. Hope they have decent amount of products so I can start blasting with watts soon. 🤓

Wasnt this already clear that they will reveal new cards at E3?? If computex was after E3 they could have cards there too but it isnt. (computex 1.june and e3 18.june) More painfull will be the info about availability. Hope they have decent amount of products so I can start blasting with watts soon. 🤓

GPU's will be at computex as they was on CES 2 months ago for professionals , the interessant dates for us, consumers, is when will they be officially presented. Computex, or E3.. this change nothing for us.. Both are taken place in June with an one week decay. The problem i see, is, in the first post here, you see peoples complaining, doing theory on why they have delay, move it to E3, ( with different theory of what they will write when they was think it was coming to computex ) .. AMD have not give a date, never say it will be at the E3, or computex, or on the moonlight fun party in Goa beach... but peoples seems take rumors from some sites, like precious date. At this point, like with Hawaii, thaiti and older presentation, the date, and place could just be anywhere in the Fiji island, at any time.

Patience is a virtue folks. The new 390X is revolutionary; Would you rather AMD rush the 390X? For those of you who are like "Screw it I am getting a GTX 980" grow up eh?

They are very late to the party. gm200 non titans will probably be out before then.

They are very late to the party. gm200 non titans will probably be out before then.

I want to see them and pricing (GM200-320). I want them at shelves few days before AMD introduces r9-390x. Just for the feeling and amusement.

GPU's will be at computex as they was on CES 2 months ago for professionals , the interessant dates for us, consumers, is when will they be officially presented. Computex, or E3.. this change nothing for us.. Both are taken place in June with an one week decay. The problem i see, is, in the first post here, you see peoples complaining, doing theory on why they have delay, move it to E3, ( with different theory of what they will write when they was think it was coming to computex ) .. AMD have not give a date, never say it will be at the E3, or computex, or on the moonlight fun party in Goa beach... but peoples seems take rumors from some sites, like precious date. At this point, like with Hawaii, thaiti and older presentation, the date, and place could just be anywhere in the Fiji island, at any time.

Lane up to 380x will be launched and released at computex(to build hype) for the big guns the 390 and 390x at e3(windoes 10, directx 12 etc) the world will be eying e3 and that would be a way to debut the titan x killers then those will release. a week or so after.

Will debut June 3rd @ Computex now.