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

[Enigmoid]

Yes but that is not merit.

R9 285 vs GTX 960 is very interesting because GTX 980 is 90%+ faster than GtX 960 and 3584 GCN 1.2 cores will only be 5% to 10% faster not more than that is what i think base on facts.

1024 Maxwell Core vs 1792 GCN 1.2 Core has some what equal performance.

Clocks.

 

[Azix]

Claimed leaked benchmarks. Looks like same old situation from 290/x launch where it was competing with the titan and Ti, tho this looks better. Still more interested in the 390. Highly unlikely I can justify a 390x purchase at the price it will launch at.

+salt

http://www.overclock3d.net/articles/gpu_displays/amd_r9_390x_vs_gtx_980_ti_performance_leak/1

 

[desprado]

Clocks.

Yes and no

Both have custom cooler and no reference design from the companies.


GCN struggles to go above 1100MHZ.

 

[96Firebird]

Claimed leaked benchmarks. Looks like same old situation from 290/x launch where it was competing with the titan and Ti, tho this looks better. Still more interested in the 390. Highly unlikely I can justify a 390x purchase at the price it will launch at.

+salt

Old benchmarks...

http://forums.anandtech.com/showthread.php?t=2424218

 

[Glo.]

desperado, you completely forgot the fact of new process, regardless if its 20 nm or 28 SHP from Glo-Fo.

And the fact, that Fiji can have higher reference clock rate than Tonga had.

Azix, those benchmarks are from last year from ChipHell.

Titan X here is OCed...

 

[Azix]

sigh.... there's one new image at least

 

[desprado]

desperado, you completely forgot the fact of new process, regardless if its 20 nm or 28 SHP from Glo-Fo.

And the fact, that Fiji can have higher reference clock rate than Tonga had.

Azix, those benchmarks are from last year from ChipHell.

Titan X here is OCed...

it is not 20 nm that is been confirmed and it is also been posted in this topic by someone.

R9 285 is based on GCN 1.2 and it is not new but it has improvements.

 

[Glo.]

Nope, all of them were posted on Chiphell then. Maybe not on this forum, but they were on ChipHell.

 

[desprado]

Nope, all of them were posted on Chiphell then. Maybe not on this forum, but they were on ChipHell.

Lisa said they are focused on 16nm double efficiency next year and skipping 20nm node.U can search on google if u have doubts.

20nm node will only come for Consoles.

 

[Glo.]

I wasn't answering to your post. Azix's that is

 

[desprado]

I wasn't answering to your post. Azix's that is

Sorry my bad....

 

[Cloudfire777]

Claimed leaked benchmarks. Looks like same old situation from 290/x launch where it was competing with the titan and Ti, tho this looks better. Still more interested in the 390. Highly unlikely I can justify a 390x purchase at the price it will launch at.

+salt

http://www.overclock3d.net/articles/gpu_displays/amd_r9_390x_vs_gtx_980_ti_performance_leak/1

Those slides are very old.
I still think its very unlikely that they had 3 unannounced cards: 960Ti, R9 390X and 980Ti, from AMD and Nvidia, at their disposal to test so many games and do power measurements.

Although I wish its true, Im not buying it

 

[beginner99]

Claimed leaked benchmarks. Looks like same old situation from 290/x launch where it was competing with the titan and Ti, tho this looks better. Still more interested in the 390. Highly unlikely I can justify a 390x purchase at the price it will launch at.

What makes me think these are fake is the cut down GM200. just look at the 970 and the VRAM issue. The issue is due to the architecture. Therefore they will have the same problems when cutting downs chips, also with GM200. So either 980TI is full GM200 or it has a similar VRAM issue like the 970 and not all of the 6 GB is actually that fast. Therefore a 4 GB 390x would not be that bad in comparison. Bad would however be if 980ti is full gm200 but then the titan becomes useless.

 

[Elfear]

Yes but that is not merit.

R9 285 vs GTX 960 is very interesting because GTX 980 is 90%+ faster than GtX 960 and 3584 GCN 1.2 cores will only be 5% to 10% faster not more than that is what i think base on facts.

1024 Maxwell Core vs 1792 GCN 1.2 Core has some what equal performance.

The 2816 shader 290X is within ~10% of the 980. Your prediction is that a 27% larger GCN chip (3584 shaders) will only be 5-10% faster than the 980? Even with absolutely no architechural improvements, that seems low.

 

[xthetenth]

What makes me think these are fake is the cut down GM200. just look at the 970 and the VRAM issue. The issue is due to the architecture. Therefore they will have the same problems when cutting downs chips, also with GM200. So either 980TI is full GM200 or it has a similar VRAM issue like the 970 and not all of the 6 GB is actually that fast. Therefore a 4 GB 390x would not be that bad in comparison. Bad would however be if 980ti is full gm200 but then the titan becomes useless.

They don't necessarily have to cut down the GM200 crop in the same way as they did the GTX 970. The architecture gave them the option to do it that way.

 

[Enigmoid]

What makes me think these are fake is the cut down GM200. just look at the 970 and the VRAM issue. The issue is due to the architecture. Therefore they will have the same problems when cutting downs chips, also with GM200. So either 980TI is full GM200 or it has a similar VRAM issue like the 970 and not all of the 6 GB is actually that fast. Therefore a 4 GB 390x would not be that bad in comparison. Bad would however be if 980ti is full gm200 but then the titan becomes useless.

Depends on if they cut the ROP partition or not.

980m is a heavily cut down Gm204 chip yet does not suffer from the memory issue because the ROPs were not touched.

Cutting out shaders is independent of L2/ROP/memory I/O.

 

[gamervivek]

If R9 290X with old technology is right now 10-20% slower than GTX980, then bigger die, with higher core count and newer technology should be faster than GTX980.

3584 GCN cores with 4 GB of HBM should be 20% faster than GTX980.
4096 GCN core with 4 GB of RAM and 1050 MHz core clock, air cooled should be 10% slower than Titan X
4096 GCN core, with 8 GB of HBM, 1200 MHz of core clock, and water cooled should be faster than Titan X.


That is my prediction.

For a 980 that boosts to 10% higher clocks than a TriX 290X, it's only ahead by about 4% at 1440p.

http://www.hardware.fr/articles/933-6/specifications-gpu-boost-pratique.html

http://www.hardware.fr/articles/933-24/recapitulatif-performances.html

290X has 40% more shaders than 980.

So if a gtx980 with similar per shader efficiency as 290X and clocks was standard at 100, then a gtx980 with 10% higher clocks but same shader efficiency would be at 110, while 290X would be at 140 and the real gtx980 at 145.6.

So the shader efficiency advantage of maxwell over hawaii GCN would be,

145.6/110 = 1.32

At same clocks, the 3584 GCN part would be, 3584/(2048*1.32) = 1.33 times as fast as gtx980. At the same clock difference as with the above 290X and 980, it would be 1.33/1.10 = 1.21 times as fast as gtx980.

So your figure does turn out be right after all! :whiste:

edit: for some reason I was thinking that 290X had 2880 shaders. But it's only 2816, so 1.375 times shaders than gtx980. Maxwell shader efficiency over hawaii, 1.3. Then 22.4% faster than gtx980.

 

[Glo.]

http://community.cadence.com/cadenc...sing-finfet-fd-soi-or-bulk-planar-fets-part-2

This is interesting from technology perspective. If Fiji and upcoming lineup of AMD GPUs is made on FD-SOI process, that would mean gigantic leap in efficiency of AMD GPUs.

 

[RussianSensation]

The 2816 shader 290X is within ~10% of the 980. Your prediction is that a 27% larger GCN chip (3584 shaders) will only be 5-10% faster than the 980? Even with absolutely no architechural improvements, that seems low.

His assessment is way off.

4K pushes GPUs to their limits and 980's lead over 290X is about 6-9%.

TPU = 8.6% (76% / 70%)
Sweclockers = 6% (106% / 100%)

There are also OC vs. OC results at 4K (ComputerBase):

980 OC = 14% faster than 290X OC
290X OC = 8% faster than 970 OC

A 3840 or a 4096 SP R9 300 series card would be pretty fast.

Since many sites like AT or TPU or TechReport generally tested reference 290X cards, and these cards still thermal throttled, it's highly possible 290X didn't even run at full 1Ghz all the time in their benchmarks.

Let's say if 1.05Ghz 390X 4096 SPs is always maintaining these clocks, but a 290X reference only averages 950mhz on a reference cooler, we get:

1.05Ghz x 4096 / (0.95Ghz x 2816) = 61% faster at 4K.

That's why those early leaks showed 390X outperforming 290X by up to 65%. AMD's reference blower on a 290X means 390X will get a 'hidden' advantage of constant clocks, something that penalizes a reference 290X in many reviews. Also, the scaling from 280X to 290X is almost linear.

 

[Cloudfire777]

2 more Fiji`s have shipped. YAY. lol
Can`t say a lot of samples of AMD`s Fiji are going back and forth exactly. Last one was in February

The value of just one card though: $1229...
R9 290X for example was worth $590 back in 2004.

Is this evidence that Fiji/390X will be very expensive?
:'(

 

[gamervivek]

$626 and $560 for 290X and 290 respectively. Might be the Fiji is getting a markup for being an unreleased card and is going by the Indian price which is usually a bit higher than the US ones. Titan X is around $1350 at an online shop that sells them for a bit less than amazon where it is at $1496.

http://www.amazon.in/GeForce-Maxwel...&ie=UTF8&qid=1431870964&sr=1-2&keywords=titan

Or Fiji is quite the monster. Another distant possibility could be a dual GPU card.

 

[parvadomus]

Do anyone have a 285 to OC memory like crazy? Its the only way to get a solid performance base to define 390X performance.

IMO getting color compression + better tessellation performance on 290X would get a solid performer vs GM104. But I still think they will use HBM to replace 290X, not only for fiji, because its the easier way to gain efficiency to compete in the mobile field.

 

[PPB]

Yes and no

Both have custom cooler and no reference design from the companies.


GCN struggles to go above 1100MHZ.

It struggles so much that we have been seeing tahiti chips reach 1.2ghz since their launch almost 4 years ago.

Yeah, GCN totally has clocking issues.

 

[Cloudfire777]

$626 and $560 for 290X and 290 respectively. Might be the Fiji is getting a markup for being an unreleased card and is going by the Indian price which is usually a bit higher than the US ones.
Or Fiji is quite the monster. Another distant possibility could be a dual GPU card.

Here is GTX 980 before it was released.
42000 INR = $662

Fiji in comparison:
78000 INR = $1229

I`d say we now have good sign that Fiji will be a good deal more expensive than GTX 980.
Could def be a dual GPU card based on the price yes---

 

[Azix]

It struggles so much that we have been seeing tahiti chips reach 1.2ghz since their launch almost 4 years ago.

Yeah, GCN totally has clocking issues.

It matters if you're looking for records. clocks don't need to be high but I assume its because of power consumption and heat.