Official Fury X specs

[DDH]

Damn. Confirmation, just came in tonight:

http://www.twitch.tv/thetechreport/b/670328467

Jump to 36:00...

"It is the case that the display block on this card hasn't changed substantially from the Tonga. As a result, HDMI 2.0 is not supported by this card."

Well shoot, that's it then. Nothing we can do. Fury X is a fantastic card, it deserves all of our support, tremendous value for anyone with a display port on their monitor. Bad mistake about HDMI 2.0, hopefully corrected at some point early next year.

Whoah... make that 4. 1 surely would have been enough?

 

[Fallen Kell]

Damn. Confirmation, just came in tonight:

http://www.twitch.tv/thetechreport/b/670328467

Jump to 36:00...

"It is the case that the display block on this card hasn't changed substantially from the Tonga. As a result, HDMI 2.0 is not supported by this card."

Well shoot, that's it then. Nothing we can do. Fury X is a fantastic card, it deserves all of our support, tremendous value for anyone with a display port on their monitor. Bad mistake about HDMI 2.0, hopefully corrected at some point early next year.

Yeah, that does certainly suck. I don't know if a correction will be that quick though. For that to have happen, they would need to already be prototyping the cards at this point. I think we are at least a year out from a correction because from my understanding, it isn't just a simple "replace this bit with that one", but a full redesign of the output stage/pipeline in order to support the more aggressive copy protections that come with the new port.

 

[flopper]

Yeah, that does certainly suck. I don't know if a correction will be that quick though. For that to have happen, they would need to already be prototyping the cards at this point. I think we are at least a year out from a correction because from my understanding, it isn't just a simple "replace this bit with that one", but a full redesign of the output stage/pipeline in order to support the more aggressive copy protections that come with the new port.

HDMI 2.0 wont be on amd cards in the future either most likely.
they said so 2010 they are phasing out hdmi.
buy a 4k DP tv from panasonic.

 

[Fallen Kell]

Ive looked hard and yet to find one that supports 1440P @ 120hz.... @ 60hz yes - 1080p @ 120 hz yes but never found any that do 1440p at 120hz

please link if you know of one.

Hmmm... you might be right. You need one that will push at least 443MHz. Most I am seeing only support 270MHz.

 

[StereoPixel]

Free DP-to-HDMI 2.0 cable from Fury Retail Box :thumbsup:

 

[Carnage1986]

Fury X has 4096 Shaders and 64 ROP. There's 1/64 ROP/Shader ratio. It's quite unbalanced GPU and it doesn't deserve HBM. HBM enables very high bandwith but the number of ROPs will become a bottleneck.

INT4 fillrate bandwith need:

64*1 Ghz(clock speed)*4=256 GB/sec(ROP bound)

FP16 fillrate bandwith need:

64*1G*8=512 GB/sec

At INT4 the card limited by number of ROPs. Only FP16 is scaled. But with Tonga there comes new color compression techniques and they use memory more efficiently. I wonder why they use HBM for this GPU. GDDR5 give 384 GB/sec bandwidth with 512 Bit Memory Controller at 6 Gbps.(like R9 390X) At 7 Gbps, it gives 448 GB/s. I think HBM is unnecessary for 64 ROPs and Fury X will be a big disappointment.

 

[Pandamonia]

http://www.extremetech.com/gaming/208541-amd-lifts-embargo-on-the-radeon-r9-fury-xs-performance-shows-gpu-competing-neck-and-neck-with-nvidias-gtx-980-ti

Seen this?

 

[Pandamonia]

Fury X has 4096 Shaders and 64 ROP. There's 1/64 ROP/Shader ratio. It's quite unbalanced GPU and it doesn't deserve HBM. HBM enables very high bandwith but the number of ROPs will become a bottleneck.

INT4 fillrate bandwith need:

64*1 Ghz(clock speed)*4=256 GB/sec(ROP bound)

FP16 fillrate bandwith need:

64*1G*8=512 GB/sec

At INT4 the card limited by number of ROPs. Only FP16 is scaled. But with Tonga there comes new color compression techniques and they use memory more efficiently. I wonder why they use HBM for this GPU. GDDR5 give 384 GB/sec bandwidth with 512 Bit Memory Controller at 6 Gbps.(like R9 390X) At 7 Gbps, it gives 448 GB/s. I think HBM is unnecessary for 64 ROPs and Fury X will be a big disappointment.

Turkish armchair GPU architect.

You have just ruined AMD's year with your calculator and rudamentory understanding of GPUs.

Dont say anything else until i unload my AMD stock!

 

[Carnage1986]

http://www.extremetech.com/gaming/208541-amd-lifts-embargo-on-the-radeon-r9-fury-xs-performance-shows-gpu-competing-neck-and-neck-with-nvidias-gtx-980-ti

Seen this?

There's no big performance difference with GTX 980 Ti. NVIDIA is able to greater GPU with lower chip production cost. All NVIDIA has to do cutting 980 Ti's prices or release a new, TITAN X with lower VRAM size. 12 GB is very unnecessary for games and it causes high graphics card cost.

Turkish armchair GPU architect.

I don't need to be a engineer on NVIDIA/AMD to foresee Fiji's failure. We are going to see results, then we can talk about it this again. I hope i am wrong, because if i am not things are not gonna get easier, at least for me.

 

[RampantAndroid]

http://www.extremetech.com/gaming/208541-amd-lifts-embargo-on-the-radeon-r9-fury-xs-performance-shows-gpu-competing-neck-and-neck-with-nvidias-gtx-980-ti

Seen this?

No reviewer worth anything is releasing details willingly, since they're under NDA. I won't trust anything until the NDA lifts (and really, since AMD is cutting out KitGuru for being too "negative" I will probably take the release day reviews with a grain of salt.

Edit: those images from here:
http://www.guru3d.com/news-story/amd-radeon-r9-fury-x-official-benchmarks.html (and they aren't the source.)

 

[Zstream]

As I've said, $200 is $200 Groover, we all can't light money on fire as you did with those Titan X cards. Some of us actually have a budget and have to be smart about our purchases.





We'll know hopefully in 24 hours whether AMD has made one of the worst mistakes in their history since running ATi or if "AMDMatt" was yet another mouth-breathing PR hack who didn't know what he was talking about.





Until then it's all conjecture.

Damn, you guys have 4k tvs and want the best gpu but hate buying adaptors?

 

[itsmydamnation]

Fury X has 4096 Shaders and 64 ROP. There's 1/64 ROP/Shader ratio. It's quite unbalanced GPU and it doesn't deserve HBM. HBM enables very high bandwith but the number of ROPs will become a bottleneck.

INT4 fillrate bandwith need:

64*1 Ghz(clock speed)*4=256 GB/sec(ROP bound)

FP16 fillrate bandwith need:

64*1G*8=512 GB/sec

At INT4 the card limited by number of ROPs. Only FP16 is scaled. But with Tonga there comes new color compression techniques and they use memory more efficiently. I wonder why they use HBM for this GPU. GDDR5 give 384 GB/sec bandwidth with 512 Bit Memory Controller at 6 Gbps.(like R9 390X) At 7 Gbps, it gives 448 GB/s. I think HBM is unnecessary for 64 ROPs and Fury X will be a big disappointment.

no offense peak number are irreverent, the things that actually are relevant to keep utilization and distribution high are the things no one talks about.

Your completely ignoring stuff like memory access patterns.......

 

[crashtech]

Not to mention the footprint and power consumption of GDDR5. Nvidia is going HBM, too, it's just a question of when, so why bother debating the merits anymore.

 

[xthetenth]

Fury X has 4096 Shaders and 64 ROP. There's 1/64 ROP/Shader ratio. It's quite unbalanced GPU and it doesn't deserve HBM. HBM enables very high bandwith but the number of ROPs will become a bottleneck.

INT4 fillrate bandwith need:

64*1 Ghz(clock speed)*4=256 GB/sec(ROP bound)

FP16 fillrate bandwith need:

64*1G*8=512 GB/sec

At INT4 the card limited by number of ROPs. Only FP16 is scaled. But with Tonga there comes new color compression techniques and they use memory more efficiently. I wonder why they use HBM for this GPU. GDDR5 give 384 GB/sec bandwidth with 512 Bit Memory Controller at 6 Gbps.(like R9 390X) At 7 Gbps, it gives 448 GB/s. I think HBM is unnecessary for 64 ROPs and Fury X will be a big disappointment.

You seem to be working from the assumption that all ROPs are equal, but as the comparison between Tonga and Hawaii shows, some ROPs are over twice as equal as others. Fiji should have approximately 220-250% the pixel fill of Hawaii, considering Tonga beat Hawaii in synthetics.

 

[garagisti]

No reviewer worth anything is releasing details willingly, since they're under NDA. I won't trust anything until the NDA lifts (and really, since AMD is cutting out KitGuru for being too "negative" I will probably take the release day reviews with a grain of salt.

Edit: those images from here:
http://www.guru3d.com/news-story/amd-radeon-r9-fury-x-official-benchmarks.html (and they aren't the source.)

If you care about it so much, then you must be aware that Nvidia's reviewers guide allows very little to no flexibility? If you defer much, then no more samples for you (this was a year or so ago, don't know if much has changed...). Just saying that you should save some stick for the other guys too

 

[garagisti]

Not to mention the footprint and power consumption of GDDR5. Nvidia is going HBM, too, it's just a question of when, so why bother debating the merits anymore.

It is no more when. Pascal will be out with HBM 2.

 

[RoarTiger]

Havent seen this mentioned yet:
"AMD will not allow you to overclock the memory when the first Fiji cards are released" source

Seems reasonable considering it is new tech, but not exactly what I expected when I heard "overclockers dream" at E3. Disappointed.

As for HDMI 2.0 nonsense, it is a terrible input which will hopefully be replaced before 4k becomes relevant to the majority of users. Early adoptors suffer lots of penatlies and this standard is one of them for 4k users. Manufacturers should have realized 4k adoption is a perfect time to abandon HDMI since HDCP 2.2 can work with DP as well, but instead they keep trying to breath life into a bad input which still doesnt have the bandwidth that will be needed going forward.

 

[Carnage1986]

You seem to be working from the assumption that all ROPs are equal, but as the comparison between Tonga and Hawaii shows, some ROPs are over twice as equal as others. Fiji should have approximately 220-250% the pixel fill of Hawaii, considering Tonga beat Hawaii in synthetics.

Tonga's raw fillrate performance worse than Hawaii:

But in this test Tonga beats his big brother:

The only reason of Tonga's success on 3DMark Pixel Fill the delta color compression bandwith efficiency. Tonga's theoritical bandwidth is 176 GB/sec and it has 32 ROPs. Now, Fiji has HBM (512 GB/sec). It doesn't scale HBM. It could have more performance with more ROPs. Yes, HBM has another benefits(for decreasing power consumption dramatically, not just because memory modules' s power needs) but it can be more useful with more raw pixel fillrate.

 

[jpiniero]

Fury X supports DP at 1/16 rate, so the 390X should be faster on DP. Still much better than Titan X though.

 

[happy medium]

Damn. Confirmation, just came in tonight:

http://www.twitch.tv/thetechreport/b/670328467

Jump to 36:00...

"It is the case that the display block on this card hasn't changed substantially from the Tonga. As a result, HDMI 2.0 is not supported by this card."

Well shoot, that's it then. Nothing we can do. Fury X is a fantastic card, it deserves all of our support, tremendous value for anyone with a display port on their monitor. Bad mistake about HDMI 2.0, hopefully corrected at some point early next year.

I watched most of this video as I fell asleep last night. Towards the end they talk about Fury and why the interposer made the AMD engineers make a smaller less powerfull Fury chip that does not completely utilize the HBM memory. To me it seems like the Fury was not quite ready for prime time but was released anyway. I think the next gpu's from Nvidia and AMD are gonna be much faster than what we have now and possibly as early as next summer.

Edit: they also talked about why one game might run better on fury and others on the 980ti. Not just like today but much differently because the fury chips are made to do one thing great and the 980 was made to be faster in another area. .

 

[AnandThenMan]

Havent seen this mentioned yet:
"AMD will not allow you to overclock the memory when the first Fiji cards are released" source

If there is one thing Fiji won't have to worry about is memory bandwidth so it's a non issue.

 

[raghu78]

If there is one thing Fiji won't have to worry about is memory bandwidth so it's a non issue.

exactly. in fact with tonga color compression which brings 40% more bandwidth efficiency and the 60% higher bandwidth over R9 290X you are looking at 1.6 x 1.4 = 2.24 times effective bandwidth of R9 290X for just 45% increase in shaders. The bandwidth per sp is up by > 50% . So yeah Fiji is not going to need memory overclocking as it already has excess bandwidth. It makes sense to be cautious when releasing such bleeding edge memory technology. So AMD is playing it safe and I don't think we can blame them.

 

[Erenhardt]

I thought the reason is thermal microbumps that could be damaged by overclocking and overvolting memory.
Breaking the microbump could fry a HBM quite easy.

Also, fiji has insane amount of bandwidth, I don;t think overclocking hbm will bring any noticable improvements

 

[swilli89]

Thanks for the info. My other question was, between the 7500 and 7100, are you ONLY losing the curve? Or is there something else.

I wouldn't say that. I would say you are gaining a proper flat screen

 

[Vesku]

I watched most of this video as I fell asleep last night. Towards the end they talk about Fury and why the interposer made the AMD engineers make a smaller less powerfull Fury chip that does not completely utilize the HBM memory. To me it seems like the Fury was not quite ready for prime time but was released anyway. I think the next gpu's from Nvidia and AMD are gonna be much faster than what we have now and possibly as early as next summer.

Edit: they also talked about why one game might run better on fury and others on the 980ti. Not just like today but much differently because the fury chips are made to do one thing great and the 980 was made to be faster in another area. .

Smaller? It's 598mm2 which is near the limit for viable dies, Nvidia's Titan X chip is 601mm2.