[VC]AMD Fiji XT spotted at Zauba

[SperglordActual]

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.

I am not familiar enough with the ins and outs of discrete GPU's to know this... Where does memory speed and bus truly come into play vs having a faster chip? I'm assuming higher resolutions are where it is most important? Also doesnt HBM not have to be as "fast" being that there is more bandwidth? This is why I asked.

 

[RussianSensation]

P

Wow 4096 SP with HBM1. thats massive. I hope AMD has made serious architectural efficiency improvements to improve perf/watt. These chips need to be competitive in perf and perf/watt otherwise AMD would just be shooting itself in the foot. I am waiting to see if AMD has improved perf/shader, perf/watt and perf/sq mm. If they have then its going to be one hell of a card. :biggrin:

I had 60-64 CUs pencilled in 1 year ago. 290X increased the number of CUs by 37.5% over 7970 (44 vs. 32). Same increase from 290X, we are already at 60 CUs. It's about time AMD doubled 7970's performance as we are near the 3 year mark since Tahiti launched.

I guess people missed Tonga reviews based on claims that Tonga did nothing for GCN architecture. With Tonga AMD already achieved a 40% increase in memory bandwidth efficiency and nearly doubled the geometry performance of Tahiti. In fact Tonga is faster at tessellation than Hawaii is. These architectural advancements barely show up in games because Tonga is bottlenecked elsewhere - Shader, pixel and texture performance. AMD has quietly used Tonga as a testbed for some architectural improvements that will make their way to 390X. Since 390X will have more TMUs and SPs, these advantages will come into their own.

Right now 980 is about 20% faster than 290X but at 4K in SLI vs. CF, 980's advantage disappears unless you overclock 980s to the max. I find it very hard to believe that 390X will be only 20% faster than a 290X.

If GM200 and 390X are both 50% faster than 980/290X, respectively, GM200 will still have the lead overall over 390X. But the extra memory bandwidth of 390X could make it fairly competitive at 4K. Whatever happens, I have no doubt in the next 6-8 months 980 will become a mid-range product. I am cheering for 390X to blow 980 away so that there is finally competition because $550-600 prices for mid-range chips is laughable, sorry NV.

 

[f1sherman]

@RussianSensation
GM204 has made a little revolution
Analysts reports of Nvidia swamping the channel with 80% at the high-end.

Tonga did pretty much nothing for anyone, except catching up with NV on theoreticals, bit it did go backwards in perf, VRAM and perf/mm2 and nicely so.
Who the heck does that?? Its unheard of

Bottom line: No one wants Tonga except Apple. No one buys it, no one recommends it to anyone. And everyone wants GM204. I am exaggerating but that's the gist of it

To even compare the two, let alone tout Tonga and laugh at 970/980 so called failed launch

(launch which had sent AMD flagships packing all the way to $250/300 ($200-250 at sales) )

to me that suggests heavy lack of realism.
I know you are considered unbiased by many here, and I perhaps am not, but I'll call spade a spade when I see it.

The only thing that can save Tonga from being the most meaningless chip in recent history is supposed 384 bit version, but that remains to be seen.

 

[Fallen Kell]

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

600mm^2 I think was the limit of the manufacturing lines to my knowledge (at least at TSMC).

 

[AtenRa]

Tonga did pretty much nothing for anyone,

Sorry but Tonga brought R9 280/X performance, Mantle and True Audio down to $249 price point, directly attacked NVs GTX760 and completely destroyed it in performance.
Not to mention R9 285 was very close to GTX770 in many games and even faster in some others when GTX770 was at $300+.

http://www.anandtech.com/show/8460/amd-radeon-r9-285-review

 

[f1sherman]

oh come on

lets see a thread here where 285 gets recommended. would you recommended it over 280?

I wanna meet the guy that bought Tonga over R9 280; and then we'll cry together over his 2GB of VRAM and laugh at True Audio

 

[Headfoot]

Tonga is a mobile chip at heart. Moved/moving the ball forward for laptops. If they have the design for mobile, I bet the cost to make a desktop version wasn't all that much more marginally versus the potential extra sales. All of this to say, desktop Tonga is an afterthought that might get interesting at the right price. It doesn't make much sense in a world with $250 290s though.

 

[f1sherman]

Lets cut with this nonsense OK.
I am not going to argue anymore against half-made arguments.
Tonga moves nothing, least of all laptops.

mobile chip at heart $#@&*(


This is the mobile chip
http://www.anandtech.com/show/8694/msi-gt72-dominator-pro-review

AMD making mobile specialized chips, like they can afford it, need it, or like product dispersion is a sound proposition... right

They have around 317986 chips already, lets make another 1

 

[Headfoot]

Lets cut with this nonsense OK.
I am not going to argue anymore against half-made arguments.
Tonga moves nothing, least of all laptops.

mobile chip at heart $#@&*(


This is the mobile chip
http://www.anandtech.com/show/8694/msi-gt72-dominator-pro-review

AMD making mobile specialized chips, like they can afford it, need it, or like product dispersion is a sound proposition... right

You're right buddy. AMD made Tonga purely for the joy of making a chip. They had no target market in mind when they made it. In fact, Tonga was made so the CEO can have a different looking wafer in a shadow box in her office.

 

[f1sherman]

:thumbsup:

haha at least that's funny

lets move on, aight

 

[positivedoppler]

We needed a sticky dedicated GM200 thread and a dedicated 390X Fiji thread.

 

[Enigmoid]

Tonga is a mobile chip at heart. Moved/moving the ball forward for laptops. If they have the design for mobile, I bet the cost to make a desktop version wasn't all that much more marginally versus the potential extra sales. All of this to say, desktop Tonga is an afterthought that might get interesting at the right price. It doesn't make much sense in a world with $250 290s though.

Is Tonga even in a single laptop? Only product I see it in is the iMac.

 

[Silverforce11]

With Tonga AMD already achieved a 40% increase in memory bandwidth efficiency and nearly doubled the geometry performance of Tahiti. In fact Tonga is faster at tessellation than Hawaii is. These architectural advancements barely show up in games because Tonga is bottlenecked elsewhere - Shader, pixel and texture performance.

1. More memory efficiency is good, but given that Hawaii already has plenty to handle 4K, and R390X has stacked memory to nearly double bandwidth, I don't think this is a meaningful improvement in gaming.

2. AMD GPUs have not been limited by tessellation performance in GAMES since Tahiti. Sure in artificial benchmarks you can see the difference, but games that are optimized well or even poorly (Crysis 2/3), GCN keeps up with the competition just fine. I don't think Tessellation will be the bottleneck in games, not when cross-platform devs have to design for consoles running GCN. Maybe in a few Gameworks titles it will make a difference.

On paper, there's just no way for R390X to be 50% faster than R290X. The shader count growth alone discounts that possibility unless Fiji is GCN2.0 with major IPC gains. Somehow I don't think it's AMD's Maxwell.

 

[SunburstLP]

Is Tonga even in a single laptop? Only product I see it in is the iMac.

I have no source, but Occam would seem to indicate that we have 285s because of AMD's contract with Apple. They used their sweet contract money to have a go at some things that they're looking at for the next-gen and sell the dies that didn't make the cut in a stop-gap SKU. Seems fairly straight-forward to me.

 

[Abwx]

On paper, there's just no way for R390X to be 50% faster than R290X. The shader count growth alone discounts that possibility unless Fiji is GCN2.0 with major IPC gains. Somehow I don't think it's AMD's Maxwell.

On paper , at same frequency, a 4096 SPs GCN 2.0 chip would be 60% faster than a Tahiti with 2816 SPs, this should be about the same against a Hawai, pehaps only 50% faster at worst.

http://www.hardware.fr/articles/926-24/tonga-vs-tahiti.html

 

[Silverforce11]

On paper , at same frequency, a 4096 SPs GCN 2.0 chip would be 60% faster than a Tahiti with 2816 SPs, this should be about the same against a Hawai, pehaps only 50% faster at worst.

http://www.hardware.fr/articles/926-24/tonga-vs-tahiti.html

In that test, Tahiti is crippled with much reduced vram clocks.

There is no indication that Hawaii is bandwidth starved, not with its 512 bus.

Thus the bandwidth efficiency gains with Tonga cannot be extrapolated into R390X, certainly not in comparison with a crippled Tahiti.

Total shaders or TMU, ROPs from Hawaii is not a massive leap even when there's 100% scaling. Expecting 50% more performance going from a large Hawaii die to a slightly larger one on Fiji at the same node is not realistic unless you believe Fiji is GCN 2.0 and its greatly improved in IPC.

ps. You also need to re-think your calculation because R290X is a heck of a lot faster than R280X or R285. 60% vs Tahiti is not 50% vs Hawaii.

 

[Abwx]

In that test, Tahiti is crippled with much reduced vram clocks.

It is not crippled for the sake of being crippled, the reviewer tested with equal memory bandwith, this way neither chip get advantaged by this parameter and only architectural differences will have an influence on the results, so his numbers are quite relevant.

As for calculations it is based on 4096 SPs and 8-11% better architectural efficency than Tahiti, the number more or less holds.

 

[Silverforce11]

It is not crippled for the sake of being crippled, the reviewer tested with equal memory bandwith, this way neither chip get advantaged by this parameter and only architectural differences will have an influence on the results, so his numbers are quite relevant.

As for calculations it is based on 4096 SPs and 8-11% better architectural efficency than Tahiti, the number more or less holds.

The improvement is in bandwidth efficiency. That would only lead to a performance gain IF bandwidth is limited.

 

[Abwx]

Agree on the bandwith issue, he should had rather overclocked the RAM of the 285X than downclocking the 280 RAM, this would had eliminated this unknown, and amplified the architectural difference influence in his percentages.

 

[raghu78]

The improvement is in bandwidth efficiency. That would only lead to a performance gain IF bandwidth is limited.

Sorry but you completely ignore the 50% fill rate improvement, 90% tesselation improvement and new instructions in Tonga alongwith improved UVD and VCE. Actually Tonga can gain even more perf if it had more bandwidth.

http://techreport.com/review/26997/amd-radeon-r9-285-graphics-card-reviewed/2

To give you a few examples of games where R9 285 beats R9 280X / R9 280 due to improved architecture - Bioshock Infinite, Thief, AC Black Flag while still being bandwidth constrained

Bioshock infinite

http://www.anandtech.com/show/8460/amd-radeon-r9-285-review/9
http://www.techpowerup.com/reviews/Sapphire/R9_285_Dual-X_OC/10.html
http://www.guru3d.com/articles_pages/powercolor_radeon_r9_285_turboduo_review,13.html
http://www.bit-tech.net/hardware/graphics/2014/09/02/amd-radeon-r9-285-review/4
http://ht4u.net/reviews/2014/amds_tonga-gpu_-_radeon_r9_285_im_test/index27.php
http://www.techspot.com/review/873-amd-radeon-r9-285/page3.html

Thief
http://www.anandtech.com/show/8460/amd-radeon-r9-285-review/14
http://www.guru3d.com/articles-pages/amd-radeon-r9-285-review,10.html
http://techreport.com/review/26997/amd-radeon-r9-285-graphics-card-reviewed/7
http://www.tomshardware.com/reviews/amd-radeon-r9-285-tonga,3925-7.html
http://www.techpowerup.com/reviews/Sapphire/R9_285_Dual-X_OC/18.html

In Assassins Creed Black Flag you can see R9 285 beat R9 280X when FXAA is used . But fall behind when MSAA 4x is used

AC Black Flag (FXAA)

http://www.tomshardware.com/reviews/amd-radeon-r9-285-tonga,3925-8.html

http://www.purepc.pl/karty_graficzn...remierowy_test_sapphire_r9_285_dualx?page=0,4

AC Black Flag (MSAA)

http://ht4u.net/reviews/2014/sapphire_r9_285_itx_compact-edition_im_test/index23.php

http://www.techpowerup.com/reviews/Sapphire/R9_285_Dual-X_OC/6.html

Even in games like Crysis 3 the R9 285 opens up a larger gap against R9 280 with SMAA 2TX or FXAA as compared to MSAA.

http://www.techspot.com/review/873-amd-radeon-r9-285/page3.html
http://hexus.net/tech/reviews/graphics/74033-sapphire-radeon-r9-285-dual-x-oc-28nm-tonga/?page=6
http://www.bit-tech.net/hardware/graphics/2014/09/02/amd-radeon-r9-285-review/5
http://www.hardwareluxx.de/index.ph...delle-der-radeon-r9-285-im-test.html?start=10
http://techreport.com/review/26997/amd-radeon-r9-285-graphics-card-reviewed/5

So yeah even with the memory efficiency improvements R9 285 can do more if it had more bandwidth. R9 390X is going to come with 512 GB/s. This is 60% more bandwidth for 45% more stream cores compared to R9 290X. I definitely expect perf/shader to go even higher than what we saw with R9 285 vs R9 280. With higher bandwidth per shader, higher memory bandwidth efficiency, architectural improvements we saw in Tonga and more which we are yet to see, R9 390X has a good chance at being close to 15 - 20% higher perf/shader than R9 290X. I can see a > 50% perf increase at 4K, close to 50% increase at 1440p/1600p and slightly lesser increase at 1080p wrt R9 290X.

 

[III-V]

In that test, Tahiti is crippled with much reduced vram clocks.

There is no indication that Hawaii is bandwidth starved, not with its 512 bus.

Thus the bandwidth efficiency gains with Tonga cannot be extrapolated into R390X, certainly not in comparison with a crippled Tahiti.

Total shaders or TMU, ROPs from Hawaii is not a massive leap even when there's 100% scaling. Expecting 50% more performance going from a large Hawaii die to a slightly larger one on Fiji at the same node is not realistic unless you believe Fiji is GCN 2.0 and its greatly improved in IPC.

ps. You also need to re-think your calculation because R290X is a heck of a lot faster than R280X or R285. 60% vs Tahiti is not 50% vs Hawaii.

The important thing you need to recognize is that the extra bandwidth enables the addition of more hardware. Hawaii's may have been fine with 48 ROPs, but with 96, it'd have been massively bandwidth starved. Today's workloads are apparently favoring pixel fillrate, and I'd imagine that's especially the case with higher resolutions being pushed on consumers.

Of course, they could go the (IMO) optimal route and have awesome compression algorithms like Nvidia, who's managing very well with a twice-than-average ROP:IMC ratio on GM204, despite being on a 256-bit bus. That way, when disruptive memory technologies surface, they'd be better prepared to take full advantage of them.

 

[AtenRa]

oh come on

lets see a thread here where 285 gets recommended. would you recommended it over 280?

I wanna meet the guy that bought Tonga over R9 280; and then we'll cry together over his 2GB of VRAM and laugh at True Audio

R9 285 is not recommended because its price is too close to R9 290. So everyone recommends the cheaper 280 or the little more expensive 290. R9 285 should be at $200 or less price point not higher at this time.

And yes i would recommend R9 285 over R9 280 if the price was close and R9 290 was at $300 or more. R9 285 is way better card than R9 280, even with 2GB of ram.

In Europe you can find the ASUS STRIX R9 280 from 190.00 euros, ASUS STRIX R9 285 from 250.00 euros and ASUS DC II R9 290 from 270-280.00 euros.

No wonder why it is not recommended, for 20.00euro more you get the 4GB R9 290 or for 60.00euro less you get almost the same performance and 3GB memory.

If 280 and 285 had the same price i believe the majority would recommend the 285.

 

[garagisti]

R9 285 is not recommended because its price is too close to R9 290. So everyone recommends the cheaper 280 or the little more expensive 290. R9 285 should be at $200 or less price point not higher at this time.

And yes i would recommend R9 285 over R9 280 if the price was close and R9 290 was at $300 or more. R9 285 is way better card than R9 280, even with 2GB of ram.

In Europe you can find the ASUS STRIX R9 280 from 190.00 euros, ASUS STRIX R9 285 from 250.00 euros and ASUS DC II R9 290 from 270-280.00 euros.

No wonder why it is not recommended, for 20.00euro more you get the 4GB R9 290 or for 60.00euro less you get almost the same performance and 3GB memory.

If 280 and 285 had the same price i believe the majority would recommend the 285.

That's the rub... even though it has more features, and is a properly better card, save for issue already pointed out, it is not easy to recommend it.

On the other hand, i fully expect to see people complaining about something or the other, when the new cards are launched and they once again can't fault the in game performance much. I think one chap who was employed with AMD, still is iirc, was confident about 295x2 and well, it wasn't misplaced. Thought i read him being confident about whatever's to come next. Let us see what comes...

 

[Erenhardt]

4GB 285 for current 2GB prices wouldn't be a bad buy.

But buying card now is a bad idea in general. With the leaks coming from all directions it is better to wait and see if new cards bring more performance at lower prices or push older cards to lower price brackets

 

[garagisti]

4GB 285 for current 2GB prices wouldn't be a bad buy.

But buying card now is a bad idea in general. With the leaks coming from all directions it is better to wait and see if new cards bring more performance at lower prices or push older cards to lower price brackets

I would second it! God knows i'm waiting to build a system, and i'm waiting for the card launch.