[BitsAndChips]390X ready for launch - AMD ironing out drivers - Computex launch

[AtenRa]

Top of the line cards aren't marketed towards the 90% nor are they priced for the 90%

So this is likely a card that's priced for the 10% with ram adequate for the other 90 if these rumors are accurate.

90% of the market is up to 1080p, lots and lots of those people buy high end GPUs. So yes, 90% of the people will be just fine with 4GB memory. Those few that have 1440p/4K monitors they can buy the 8GB cards.

 

[2is]

90% of the market is up to 1080p, lots and lots of those people buy high end GPUs. So yes, 90% of the people will be just fine with 4GB memory. Those few that have 1440p/4K monitors they can buy the 8GB cards.

I would just repeat what I said, but you wouldn't understand.

 

[JDG1980]

R9 395X2 with 16 GB for 1299$

source

If the R9 395 X2 is 500W, then it might be reasonable to assume that the single-GPU version will be 250W. AMD might still release it with a reference water cooler simply because they don't have a competent blower, but AIBs should be able to use their existing cooling solutions.

 

[TemjinGold]

Interesting... it does say 2x 8 gb...

 

[sakete]

What do you guys think a 390 and 390X will go for in terms of pricing?

 

[poofyhairguy]

90% of the market is up to 1080p, lots and lots of those people buy high end GPUs. So yes, 90% of the people will be just fine with 4GB memory. Those few that have 1440p/4K monitors they can buy the 8GB cards.

If those 1080p people were buying on need they wouldn't be buying high-end GPUs right? 290x or 980 and call it a day.

That is why this card needs 8GB.

 

[Head1985]

R9 395X2 with 16 GB for 1299$

source

2x FIJI
8192sp
512TMU
128Rops

One FIJI
4096SP
256TMU
64Rops
FIJi have pretty low Rops count but high TMU

 

[AtenRa]

I would just repeat what I said, but you wouldn't understand.

And i would just repeat that High-End GPUs are not only for 1440/4K Users.

 

[AtenRa]

If those 1080p people were buying on need they wouldn't be buying high-end GPUs right? 290x or 980 and call it a day.

That is why this card needs 8GB.

There are more people with 1080p 120/144Hz monitors than 1440/4K, those people need high end GPUs and they will be fine with 4GB variants.

Also, i havent said that 390X wouldnt need 8GB memory.

 

[poofyhairguy]

There are more people with 1080p 120/144Hz monitors than 1440/4K, those people need high end GPUs and they will be fine with 4GB variants.

Also, i havent said that 390X wouldnt need 8GB memory.

Good point.

 

[StereoPixel]

64Rops
FIJi have pretty low Rops count but high TMU

HBM 1 stack 1024-bit 8 Channel = 32 ROP
HBM 4 stacks 4096-bit 32 Channel = 128 ROP

 

[RussianSensation]

HBM 1 stack 1024-bit 8 Channel = 32 ROP
HBM 4 stacks 4096-bit 32 Channel = 128 ROP

I am not buying the idea of 128 ROPs for 2 reasons:

1) R9 285 with only 32 ROPs has higher pixel fill-rate efficiency/throughput in the real world than a 64 ROP R9 290. That means if AMD incorporates the architectural advancements of Tonga into 390X, pixel fill-rate should increase 70-75% on the same 64 ROPs.

*** Diving into history, AMD has improved ROP throughput using various tricks in the past without increasing the number of ROPs.

"Even with the same number of ROPs and a similar theoretical performance limit (29.6 vs 28.16), 7970 is pushing 51% more pixels than 6970 is."

Since we already know that R9 285's 32 ROPs beat R9 290's 64 ROPs, there is no reason for AMD to double the ROPs yet again.

2) 4096 SPs and 256 TMUs is only a 45% increase in those units compared to the R9 290X. If you double the ROPs from 64 to 128, you are wasting them because they'd be bottlenecked by less than a 50% increase in SPs and TMUs.

But referring to #1, when Tonga with only 32 ROPs crushes R9 290's 64 ROPs, imagine 64 "Tonga style ROP pixel fill-rate" on an R9 390X?

R9 285's architecture has nearly a 70% more efficient pixel fill-rate throughput than an R9 280!

 

[StereoPixel]

But referring to #1, when Tonga with only 32 ROPs crushes R9 290's 64 ROPs, imagine 64 "Tonga style ROP pixel fill-rate" on an R9 390X?

R9 285's architecture has nearly a 70% more efficient pixel fill-rate throughput than an R9 280!

I think Fiji has a 128 ROPs (~160-256 effective "Tonga style ROP pixel fill-rate") to compete with Pascal.

 

[AtenRa]

ROPs are processing units, you can have 128 with less pixel fill-rate than 64 or 32 ROPs with higher fill-rate than 64 etc etc.

So you could have 128 ROPs and only double the pixel fill-rate of your 32 ROPs in your previous architecture.

 

[StereoPixel]

ROPs are processing units, you can have 128 with less pixel fill-rate than 64 or 32 ROPs with higher fill-rate than 64 etc etc.

So you could have 128 ROPs and only double the pixel fill-rate of your 32 ROPs in your previous architecture.

If Fiji will have second generation of Delta Color Compression it will have 128 ROPs with higher fill-rate.

 

[Silverforce11]

I think Fiji has a 128 ROPs (~160-256 effective "Tonga style ROP pixel fill-rate") to compete with Pascal.

You either think Fiji is gonna be extremely late or that Pascal is very early to make them competitors.

 

[RussianSensation]

I think Fiji has a 128 ROPs (~160-256 effective "Tonga style ROP pixel fill-rate") to compete with Pascal.

Fiji is a Maxwell competitor. Arctic Islands is a Pascal competitor. Why would Fiji be competing with 14/16nm GPUs? Also what do you mean 160-256 Tonga style ROPs when I just explained that Fiji needs to have Tonga style pixel fillrate efficiency to begin with.

ROPs are processing units, you can have 128 with less pixel fill-rate than 64 or 32 ROPs with higher fill-rate than 64 etc etc.

It makes no sense to use up transistor and die space with 128 ROPs when 64 Tonga style ROPs are nearly equivalent to 128 R9 290X style. Hence, there is no need for 128 ROPs in R9 390X. Just improve the efficiency based on Tonga's architecture and you got yourself a 70%+ increase in pixel fill-rate over R9 290X without changing the ROP count.

If Fiji will have second generation of Delta Color Compression it will have 128 ROPs with higher fill-rate.

Why? You'd end up with 70% increase in pixel fill-rate over R9 290X from the color compression and then double that again by going from 64 to 128? You'd end up with R9 290X's "ROP-equivalent" of 64 x 1.7 x 2 = 218 ROPs, but you will have only increased shaders and TMUs 45% from 2816/176 to 4096/256? That makes no sense at all. That would be the most unbalanced GPU design of all time.

You either think Fiji is gonna be extremely late or that Pascal is very early to make them competitors.

I don't even understand how he came up with that theory. 28nm R9 390 series needs to combat GM200, not GP200. GP200 will be at least 50% faster than the Titan X. How in the world did he think R9 390X would be able to compete with that? Some of the theories on this forum lately...

 

[StereoPixel]

Fiji is a Maxwell competitor. Arctic Islands is a Pascal competitor. Why would Fiji be competing with 14/16nm GPUs? Also what do you mean 160-256 Tonga style ROPs when I just explained that Fiji needs to have Tonga style pixel fillrate efficiency to begin with.

My guess: Grenada/Bermuda GPU with 64 ROPs and Tonga-style pixel-fillrate efficiency is a Maxwell competitor. Fiji is just a monster.

Why? You'd end up with 70% increase in pixel fill-rate over R9 290X from the color compression and then double that again by going from 64 to 128? You'd end up with R9 290X's "ROP-equivalent" of 64 x 1.7 x 2 = 218 ROPs, but you will have only increased shaders and TMUs 45% from 2816/176 to 4096/256? That makes no sense at all. That would be the most unbalanced GPU design of all time.

Yes, Fiji's 128 ROPs is a 218 ROPs-equivalent or more, because it will have new HBM, Gen2 Color Compression and other features.
I think Bermuda/Grenada with 64 ROPs will have 128 ROPs-equivalent aka 295X2.
Maxwell is "most unbalanced GPU" too -- it has less shaders/TMUs and more ROPs/pixel-fillrate efficiency than 290X/Kepler.

 

[Skurge]

Why? You'd end up with 70% increase in pixel fill-rate over R9 290X from the color compression and then double that again by going from 64 to 128? You'd end up with R9 290X's "ROP-equivalent" of 64 x 1.7 x 2 = 218 ROPs, but you will have only increased shaders and TMUs 45% from 2816/176 to 4096/256? That makes no sense at all. That would be the most unbalanced GPU design of all time.

With all the bandwidth, they could be going for single GPU 4K 60FPS performance. Although 218 seems a bit high. 64 Tonga style ROPs would make more sense.

 

[KingFatty]

AH any more leaks/information regarding when the 3XX series cards will come out? I'm dying here ready to pull the trigger on that Sapphire Tri-x OC 290 for $240 after rebate. I'd like to spend more for more performance, but how long is the wait?

 

[Creig]

AH any more leaks/information regarding when the 3XX series cards will come out? I'm dying here ready to pull the trigger on that Sapphire Tri-x OC 290 for $240 after rebate. I'd like to spend more for more performance, but how long is the wait?

The high end 300 series will undoubtedly be faster than a 290 or 290X. But you'll have to pay a lot more as well. I'd pull the trigger on that 290 Tri-X OC if I was in the market for a new video card right now. For the money, you won't get better performance or better cooling anywhere. And unless you plan to run above 1440P, the performance of that card should be adequate. Especially if you OC it further.

 

[AtenRa]

It makes no sense to use up transistor and die space with 128 ROPs when 64 Tonga style ROPs are nearly equivalent to 128 R9 290X style. Hence, there is no need for 128 ROPs in R9 390X. Just improve the efficiency based on Tonga's architecture and you got yourself a 70%+ increase in pixel fill-rate over R9 290X without changing the ROP count.

If your design has the ROPs connected to the Memory controllers then you will have more "smaller" ROPs. If you decouple your ROPs from the memory controllers like HD79xx, then you can use less but more powerful ROPs.

What im saying is that if ROPs should be connected to the HBM controllers and you need to have 128 ROPs, you dont design them to be that fast as your Tonga ROPs. Thus each ROP will be smaller and less powerful than the ROPs you used in the Tonga design, but Pixel fill-rate will be higher than the one in Tonga simple because you have more throughput.

 

[RussianSensation]

Maxwell is "most unbalanced GPU" too -- it has less shaders/TMUs and more ROPs/pixel-fillrate efficiency than 290X/Kepler.

You cannot compare NV vs. AMD's shaders, textures and ROPs directly. It doesn't work that way. Just by looking at NV's own Kepler vs. Maxwell, we can say it's impossible to compare TMUs, Shaders and ROPs from even the same brand unless NV gives us a point of reference (128 Maxwell CUDA cores provide 90% of the performance of 192 Kepler CUDA cores -- had NV not told us that, we'd need to spend a lot of time looking at benchmarks to try to estimate this ourselves). Finally, theoretical performance and real world efficiency are completely different (see HD6970 vs. 7970 pixel fill-rate despite both having 32 ROPs).

What im saying is that if ROPs should be connected to the HBM controllers and you need to have 128 ROPs, you dont design them to be that fast as your Tonga ROPs. Thus each ROP will be smaller and less powerful than the ROPs you used in the Tonga design, but Pixel feel-rate will be higher than the one in Tonga simple because you have more throughput.

Why would you design ROPs for Tonga that are 70% more efficient and then take that design that cost you millions of dollars and throw it all away and start over? I mean if you are going to go with the argument of using 128 ROPs, might as well use the same ones from Tonga and then you'll have overkill pixel fillrate.

Unless you are a GPU / electrical engineer, I don't see how you can state with such certainty that 128 smaller ROPs are better than "64 or 96 Tonga style ROPs". In fact, right now the amount of ROPs is one of the moving marks, far less certain than the consistenly leaked 4096-bit memory bus width, 4096 SPs and 256 TMUs specs over 1-1.05Ghz clocks. I am not even sure where some people on this board got 128 ROPs from?! I must have missed a rumour that even talked about 128 ROPs. When was this leaked?

Sounds like some of you guys are just chasing "paper specs" aka more is better.

AMD already doubled the number of ROPs with R9 290X but only increased shaders/TMUs by 37.5% over the 7970Ghz. With Tonga's colour compression, there is more than enough ROP performance with just 64 ROPs. AMD needs to focus on the shader, texture, L2 chace and geometry performance since by far those are the 4 biggest bottlenecks in the design. HBM1 takes care of memory bandwidth. ROPs should be the last area of focus for AMD with R9 390 series since there is going to be more than enough of pixel filtrate as is from Tonga's improvements alone. I am not saying I am right but I don't see why 128 ROPs is more credible than 64 or 96 ROPs. Where is the logic behind that claim?

 

[jackstar7]

Kyle at [H] says end of June.

 

[Azix]

Kyle at [H] says end of June.

End of June for what? Better be June first for card release. I'd have to wait 2 months plus and I don't even know if I'd be willing to spend on a r9 390. At least if it releases june 1st I could get a used one end of june or something.

Tell me you aren't talking about release date for these cards.