[VC]AMD Fiji XT spotted at Zauba

[Gloomy]

http://www.sisoftware.eu/rank2011d/...dbe6deedd4e7d3f587ba8aacc9ac91a187f4c9f9&l=en

4096 SPs
HBM

this kills the nvidia

 

[Silverforce11]

AMD will have no problems hitting 250W even on 250mm2 20nm GPU.
GCN needs serious perf/W upgrade.

Q2/Q3, or H2 2015 20nm launch suggests >500mm2 28nm design was ditched.

But yeah, I guess (and hope) that hybrid cooling was made for that 28nm design.

28nm is cheap and high yielding.

Why go with 20nm that isn't even designed for HP but for mobile SoCs? Not to mention competing with mega-players (Apple, Qualcomm etc) for wafers.

A 550mm2 refreshed "Hawaii GCN" designed for lower vcore (or rather less pushing of clocks to the max) and tweaked for efficiency should yield a very powerful GPU that won't be a crazy 350W monster. All the way leaving plenty of OC headroom for enthusiasts who don't mind the power consumption to push it.

 

[3DVagabond]

AMD will have no problems hitting 250W even on 250mm2 20nm GPU.
GCN needs serious perf/W upgrade.

Q2/Q3, or H2 2015 20nm launch suggests >500mm2 28nm design was ditched.

But yeah, I guess (and hope) that hybrid cooling was made for that 28nm design.

While nVidia's "boost software" is exceptionally good, they don't have the market cornered on writing good software. AMD can improve their perf/W just like nVidia has.

250W@250mm² is smile worthy, though.

 

[Flapdrol1337]

With a good cooler on top it's no problem having high power/area, and if the chip really has 4096 SP's and is similar to tonga it would be too big on 28nm.

 

[Keysplayr]

While nVidia's "boost software" is exceptionally good, they don't have the market cornered on writing good software. AMD can improve their perf/W just like nVidia has.

250W@250mm² is smile worthy, though.

If you're implying that the secret to Nvidia's power savings in Maxwell is all due to software, I'd be interested in knowing what made you say that.

If not, my bad and I misunderstood.

 

[Erenhardt]

A whisperer told me Fiji is much bigger than Hawaii. That points to 28nm.
Also, how much is "much"?

 

[psolord]

A whisperer told me Fiji is much bigger than Hawaii.

That's not true.

http://en.wikipedia.org/wiki/Fiji

http://en.wikipedia.org/wiki/Hawaii

18,274 km2 vs 28,311 km2

j/k

 

[Keysplayr]

That's not true.

http://en.wikipedia.org/wiki/Fiji

http://en.wikipedia.org/wiki/Hawaii

18,274 km2 vs 28,311 km2

j/k

That was good. :biggrin:

 

[dacostafilipe]

If you're implying that the secret to Nvidia's power savings in Maxwell is all due to software, I'd be interested in knowing what made you say that.

If not, my bad and I misunderstood.

"Software" as in Firmware. Fast response for optimal usage of power gating and clock gating. nVidia did a great job in this area.

 

[Keysplayr]

"Software" as in Firmware. Fast response for optimal usage of power gating and clock gating. nVidia did a great job in this area.

Yes, and they did a great job on the architecture as well. The achievements go hand in hand.

 

[f1sherman]

While nVidia's "boost software" is exceptionally good, they don't have the market cornered on writing good software. AMD can improve their perf/W just like nVidia has.

250W@250mm² is smile worthy, though.

It's not. Take a look at Hawaii 20nm ideal shrink, ie +30% higher clocks option.

200mm2, 1300MHz, TDP remains the same

Add some mm2 to bring this into real world and to save the power.

We've been over this...

 

[SperglordActual]

Can anyone familiar with HBM advise how this will affect GPU performance? I know that the 290x has a wider bus than the 780ti and that contributed to better performance @ higher resolutions... Just wondering at the implications of having so much more bandwidth...

 

[f1sherman]

290x being better performer @ higher resolutions,
or simply not using CPU clocks as efficiently, demanding more CPU power than Nvidia, and therefore being worse at lower resolution... thats debatable

HBM should shave some power. You're not going to have excruciating amount of bandwidth, because that would still be a waste of power

 

[Grooveriding]

Can anyone familiar with HBM advise how this will affect GPU performance? I know that the 290x has a wider bus than the 780ti and that contributed to better performance @ higher resolutions... Just wondering at the implications of having so much more bandwidth...

Yeah I keep seeing this posted. Does HBM provide more/similar memory bandwidth using a smaller bus interface to the GPU than what we have now does ?

edit: Yes. Looks like it's a massive potential increase to bandwidth, lower power use and denser modules.

http://www.skhynix.com/gl/products/graphics/graphics_info.jsp

 

[raghu78]

It's not. Take a look at Hawaii 20nm ideal shrink, ie +30% higher clocks option.

200mm2, 1300MHz, TDP remains the same

Add some mm2 to bring this into real world and to save the power.

We've been over this...

You make it seem as though AMD cannot make architectural efficiency improvements and improve perf / sq mm and perf/watt. btw indications are that AMD has gone 28nm with HBM and massive die (> 500 sq mm) for R9 390X. It makes sense to use a mature process for a new architecture with a new memory standard and given that 2.5D stacking brings its own set of challenges with yields as its a brand new technology. Just hold on till the final product is out. Don't rush into conclusions or pre judge a product. Wait until it has been released. :whiste:

 

[f1sherman]

You make it seem as though AMD cannot make architectural efficiency improvements and improve perf / sq mm and perf/watt.

Well yes, direct shrink is what we have been discussing.

Far from me saying that they won't improve GCN

 

[NostaSeronx]

AMD's upcoming HBM GPU processor is operating in HBM2 pseudo-channel mode.

HBM2 pseudo-channel => 64-bit data width, rather than 128-bit data width.

Think unganged versus ganged. (16*64 for unganged/pseudo channel or 8*128 for ganged/full channel)

HBM1 => 1/2th tFAW of DDR3
HBM2 => 1/3rd tFAW of DDR3 // 1/2th tFAW of DDR4


Quit spewing information without any sources. This is your last warning until I start infracting upon it.

-Moderator Rvenger

 

[CakeMonster]

We should expect some memory compression scheme similar to NV with the next AMD release too, right?

 

[monstercameron]

We should expect some memory compression scheme similar to NV with the next AMD release too, right?

like the kind that tonga uses?

 

[KaRLiToS]

http://wccftech.com/amd-fiji-r9-390x-specs-leak/

This leak indicates that the HBM modules on the R9 390X have been clocked slightly above the default 1.2Ghz frequency to 1.25Ghz. Which should drive overall bandwidth up to ~533.3GB/s. That’s more than double the memory bandwidth available to the GTX 980 which is 224GB/s. And easily surpasses the 320GB/s bandwidth of the R9 290 series and the 336GB/s data rate of the GTX 780 Ti. The R9 390X has by far the greatest amount of memory bandwidth we’ve seen on any graphics card to date.

Read more: http://wccftech.com/amd-fiji-r9-390x-specs-leak/#ixzz3ImFcRdJp

 

[ShintaiDK]

Can anyone familiar with HBM advise how this will affect GPU performance? I know that the 290x has a wider bus than the 780ti and that contributed to better performance @ higher resolutions... Just wondering at the implications of having so much more bandwidth...

You dont need to be familiar with HBM to determine that. Just try to change the memory clocks on your current GPU and see the implications.

 

[raghu78]

Yeah I keep seeing this posted. Does HBM provide more/similar memory bandwidth using a smaller bus interface to the GPU than what we have now does ?

edit: Yes. Looks like it's a massive potential increase to bandwidth, lower power use and denser modules.

http://www.skhynix.com/gl/products/graphics/graphics_info.jsp

a 4 stack 4 Hi HBM config provides 512 GB bandwidth at 1/3rd the power it takes with a 512 bit GDDR5 memory bus running 8 Ghz memory. see slide 45 from the below presentation .

http://www.microarch.org/micro46/files/keynote1.pdf

HBM uses a ultra wide bus width of 1024 bits per HBM stack. for a 4 stack config where each stack has 4 chips stacked on another has 4 x 1024 = 4096 bit memory bus width. The big power reduction happens as the GPU and the HBM memory stack are on a single silicon substrate and thus the wiring does not have to leave the package and so the power required to drive signals is reduced in a big way. :thumbsup:

 

[raghu78]

http://wccftech.com/amd-fiji-r9-390x-specs-leak/

That bandwidth calculation is rubbish. a single stack of HBM in a 4 Hi config provides 128 GB/s at 1 Gbps. So If the clocks are 1.25 Ghz the bandwidth will be 640 GB/s . But the funny part is Hynix is only in production of 1Gbps HBM memory. So that 1.25 Gbps rumour is wrong. read the page 2 of Databook_Q4'2014_Graphics.pdf

https://www.skhynix.com/products/support/databook.jsp

 

[AnandThenMan]

That's not true.

http://en.wikipedia.org/wiki/Fiji

http://en.wikipedia.org/wiki/Hawaii

18,274 km2 vs 28,311 km2

j/k

lol you win the Internet for today.

 

[raghu78]

Well yes, direct shrink is what we have been discussing.

Far from me saying that they won't improve GCN

From your statements its clear that you don't have much hopes for the R9 3xx series and GCN architectural improvements. Thats your view. There are others who want to see the final product before making a conclusion.

imo Tonga is not representative of the upcoming GCN architecture. I expect AMD to make significant improvements to perf/shader, perf/sq mm and perf/watt with R9 390x. I also expect R9 390X to be 28nm , HBM and > 500 sq mm. As always the final product will tell the facts. We have to be patient till that.