Support NeoGAF

[dgrdsv]

I was hinting at the new games this fall, which I fear will mosty pump every bit and byte into the VRAM because that seems to be new style of game optimization.

Using VRAM for cache means less hitching and faster loading but it doesn't affect the average performance that much. I've yet to see any game which needs more than 4GB of VRAM even for adjacent frames. Shadow of Mordor is the only prime suspect currently but it's hardly a typical game.

 

[Crisium]

How's crossfire compare to sli? And do you think these new rebreands will be less power intensive than the 290s?

Often xfire has better frametimes now thanks to no bridge needed. Less day 1 support though for new releases.

 

[jfoul]

Sapphire Radeon 300 series listed online

• Sapphire Radeon R9 390X TRI-X OC 8GB GDDR5 (11241-00-20G)
• Sapphire Radeon R9 390 NITRO (TRI-X OC) 8GB GDDR5 (11244-00-20G)
• Sapphire Radeon R9 380 NITRO (DUAL-X OC) 4GB GDDR5 (11242-07-20G)
• Sapphire Radeon R9 380 ITX 2GB GDDR5 (11242-02-20G)
• Sapphire Radeon R7 370 NITRO (DUAL-X OC) 4GB GDDR5 (11240-04-20G)
• Sapphire Radeon R7 370 (DUAL-X OC) 2GB GDDR5 (11240-06-20G)
• Sapphire Radeon R7 360 2GB GDDR5 (11243-00-20G)

 

[Bornstellar]

• 
  [*]Sapphire Radeon R9 390X TRI-X OC 8GB GDDR5 (11241-00-20G)
  [*]Sapphire Radeon R9 390 NITRO (TRI-X OC) 8GB GDDR5 (11244-00-20G)
• Sapphire Radeon R9 380 NITRO (DUAL-X OC) 4GB GDDR5 (11242-07-20G)
• Sapphire Radeon R9 380 ITX 2GB GDDR5 (11242-02-20G)
• Sapphire Radeon R7 370 NITRO (DUAL-X OC) 4GB GDDR5 (11240-04-20G)
• Sapphire Radeon R7 370 (DUAL-X OC) 2GB GDDR5 (11240-06-20G)
• Sapphire Radeon R7 360 2GB GDDR5 (11243-00-20G)

Hello gorgeous.

 

[Shining Sunshine]

June 16.

AMD already confirmed it.

Waaaait... that's like... soon?

But we barely have any information on price or benchmarks.

 

[Coulomb_Barrier]

I wonder whether that Sapphire Radeon R9 380 NITRO is the full Tonga chip.

 

[Unknown Soldier]

Something interesting a person found:
https://compubench.com/result.jsp?b...=true&arch-ARM=true&arch-x86=true&base=device

It's apparently Fiji:
https://compubench.com/device.jsp?b...=AMD+Radeon+Graphics+Processor&testgroup=info

 

[hateradio]

• Sapphire Radeon R9 390X TRI-X OC 8GB GDDR5 (11241-00-20G)
• Sapphire Radeon R9 390 NITRO (TRI-X OC) 8GB GDDR5 (11244-00-20G)
• Sapphire Radeon R9 380 NITRO (DUAL-X OC) 4GB GDDR5 (11242-07-20G)
• Sapphire Radeon R9 380 ITX 2GB GDDR5 (11242-02-20G)
• Sapphire Radeon R7 370 NITRO (DUAL-X OC) 4GB GDDR5 (11240-04-20G)
• Sapphire Radeon R7 370 (DUAL-X OC) 2GB GDDR5 (11240-06-20G)
• Sapphire Radeon R7 360 2GB GDDR5 (11243-00-20G)

Are 2GB cards viable at this point?

 

[kami_sama]

Something interesting a person found:
https://compubench.com/result.jsp?b...=true&arch-ARM=true&arch-x86=true&base=device

It's apparently Fiji:
https://compubench.com/device.jsp?b...=AMD+Radeon+Graphics+Processor&testgroup=info

If it is real, it's extremely good, no?

 

[Seanspeed]

Nitro and Fury? Come on AMD...

 

[ScepticMatt]

If it is real, it's extremely good, no?

290X + 40% or Titan X + 22% according to this
edit: purely compute, i.e. shading performance.

 

[kami_sama]

Nitro and Fury? Come on AMD...

I think Nitro is a Sapphire thing, like TriX.

 

[Coulomb_Barrier]

290X + 40% or Titan X + 22% according to this

22%

How do we know this isn't just an OC'd 295x2?

 

[ScepticMatt]

22%

How do we know this isn't just an OC'd 295x2?

We don't, could be anything

 

[Seanspeed]

I think Nitro is a Sapphire thing, like TriX.

Ah ok. lol

 

[DBT85]

If it is real, it's extremely good, no?

It appears in a few other tests too and wins some loses some.

I wouldn't take one chart as meaning anything.

 

[jfoul]

Ah ok. lol

RADEON WCW NITRO MAXX FURY ROAD X2

 

[DBT85]

Fudzilla say that AMD are going to bundle Star Wars Battlefront

http://www.fudzilla.com/news/games/37952-amd-to-bundle-star-wars-battlefront

Hope its not like HL2 where I buy a 9800xt and had to wait a bloody year for the game lol.

 

[Unknown Soldier]

22%

How do we know this isn't just an OC'd 295x2?

It has 4096 shader cores, it's Fiji

 

[Coulomb_Barrier]

It has 4096 shader cores, it's Fiji

Ah ok, that's logical then.

 

[dgrdsv]

290X + 40% or Titan X + 22% according to this
edit: purely compute, i.e. shading performance.

...In ocean surface simulation. There are other tests in this benchmark which show a different picture - as it usually is with pure compute.
And compute isn't the same as shading either.

 

[jfoul]

Fudzilla say that AMD are going to bundle Star Wars Battlefront

http://www.fudzilla.com/news/games/37952-amd-to-bundle-star-wars-battlefront

Hope its not like HL2 where I buy a 9800xt and had to wait a bloody year for the game lol.

This is probably happening, AMD is tight with Dice, and BF4 was bundled with the R9 270 - 290X.

[url=http://imgur.com/45okS7q][img]http://i.imgur.com/45okS7q.png[/img][/url]
[url=http://imgur.com/H6s70wo][img]http://i.imgur.com/H6s70wo.png[/img][/url]

 

[wachie]

Leaks are happening? There is a roadmap posted on Videocardz but they rightfully say it doesnt look genuine.

 

[King Gilga]

If that render I'd legit that's most aesthetically please graphics card I've seen. Not tacky at all like most.

Often xfire has better frametimes now thanks to no bridge needed. Less day 1 support though for new releases.

I see thanks, tbh the last non wow expansion game I played on release was skyrim. So its not really an issue.

 

[Dictator93]

290X + 40% or Titan X + 22% according to this
edit: purely compute, i.e. shading performance.

Not exactly shading performance, but still extremely useful.

 

[Deleteeeeee]

If that render I'd legit that's most aesthetically please graphics card I've seen. Not tacky at all like most.

I could do without the MSI in the grill but otherwise I would agree. It almost looks like comic sans.

 

[michaelius]

290X + 40% or Titan X + 22% according to this
edit: purely compute, i.e. shading performance.

Which should tell You from the start that this result can't be used to compare performance in games.

 

[Nachtmaer]

Not exactly shading performance, but still extremely useful.

At least it sort of confirms that Fiji has 64 CUs. So on paper that's 290X + 45% performance which puts it damn close to the Titan X. I know shaders themselves don't paint the whole picture and they don't always scale linearly. Also we still don't know the exact frequency either.

It's kind of crazy how a long time ago the 64 CU number seemed nothing more than a pipe dream and it turns out to most likely being true.

 

[Coulomb_Barrier]

At least it sort of confirms that Fiji has 64 CUs. So on paper that's 290X + 45% performance which puts it damn close to the Titan X. I know shaders themselves don't paint the whole picture and they don't always scale linearly. Also we still don't know the exact frequency either.

It's kind of crazy how a long time ago the 64 CU number seemed nothing more than a pipe dream and it turns out to most likely being true.

Well the unknown really is what performance gain the massive memory bandwidth will offer.

I think the card will be clocked at at least 1Ghz (rumours say 1.05Ghz) but this may have been bumped up after the 980 Ti reveal.

 

[tuxfool]

I'm not so sure how good an indicator those compute tests are. Nvidia has been dragging its heels on OpenCL and that is reflected in the tests. The TitanX tests show it is using OpenCL 1.2, where as that unnamed AMD card is using OpenCL 2.0.

 

[Renekton]

I'm not so sure how good an indicator those compute tests are. Nvidia has been dragging its heels on OpenCL and that is reflected in the tests. The TitanX tests show it is using OpenCL 1.2, where as that unnamed AMD card is using OpenCL 2.0.

You can just exclude Nvidia and safely focus on where it compares within its own AMD lineup.

(aaaand it shows only 18% improvement over 290X, nevermind let's wait for official benches <_<" )

 

[Nachtmaer]

Well the unknown really is what performance gain the massive memory bandwidth will offer.

I think the card will be clocked at at least 1Ghz (rumours say 1.05Ghz) but this may have been bumped up after the 980 Ti reveal.

I think it will be hard to tell. Obviously a beefier chip requires more bandwidth to keep it fed, but once you're past that bottleneck, throwing more at it won't add more performance. This is the reason why OCing a 290X's memory doesn't do much either. Besides a smaller package and lower power consumption, it will also be hard to say what other gains HBM brings vs GDDR5 since we don't have two versions to compare. Like let's say Fiji with HBM and a Fiji with GDDR5 clocked high enough to have the same bandwidth. This probably sounds nitpicky, but I assume you get my point.

And yeah, I've seen that ~1GHz figure being thrown around too. I'd say Fiji will probably be about 40 to 50% faster than a 290X. Anything higher than that would be a nice surprise though.

 

[Souldestroyer Reborn]

Something interesting a person found:
https://compubench.com/result.jsp?b...=true&arch-ARM=true&arch-x86=true&base=device

It's apparently Fiji:
https://compubench.com/device.jsp?b...=AMD+Radeon+Graphics+Processor&testgroup=info

Holy shit. That's impressive.

I don't know why I even read these threads. I don't even have a PC >.>

Last GPU I had was a HD6670...(Didn't even know if it was good, bad or whatever)

 

[mjontrix]

Holy shit. That's impressive.

I don't know why I even read these threads. I don't even have a PC >.>

Last GPU I had was a HD6670...(Didn't even know if it was good, bad or whatever)

The 3xx series will run rings around it. Should consider saving up for a gaming pc - the 4xx series is the one to get.

 

[Renekton]

I think it will be hard to tell. Obviously a beefier chip requires more bandwidth to keep it fed, but once you're past that bottleneck, throwing more at it won't add more performance. This is the reason why OCing a 290X's memory doesn't do much either.

Would Titan X be underserved by its 384-bit bus (336GBs) if it wasn't for the compression tech?

I'm trying to get a sense of where this bandwidth scenario would kick in.

 

[ohyouagain]

I hope the 360 is a serious competitor to the 750 Ti in the SFF, PCI-E-only, market. I am baffled that in one and half year we haven't seen any development in this segment.

 

[Nachtmaer]

Would Titan X be underserved by its 384-bit bus (336GBs) if it wasn't for the compression tech?

I'm trying to get a sense of where this bandwidth scenario would kick in.

I can't say for sure since my knowledge doesn't go that low level. Nvidia's memory controllers the past few generations have been able to deal with really high clocked memory which compensates the narrower bus. The compression is probably another reason why they can get by with less bandwidth compared to AMD. Putting it simply, the bus width in itself doesn't mean much because bandwidth = bus width * frequency. Then there's also cache performance and size (which I think Nvidia has improved a lot in Maxwell).

Being able to achieve higher frequencies requires beefier circruity afaik though. This is why AMD went with a 512-bit wide one with slower clocked memory for the 290s. They noticed that a simpler but wider memory controller would take up less space than Tahiti's (7970) 384-bit wide one.

It just goes to show that building a chip is a balancing act of loads of factors like physical size, bandwidth, shaders and other "core" blocks and whatnot.

 

[Coulomb_Barrier]

I think it will be hard to tell. Obviously a beefier chip requires more bandwidth to keep it fed, but once you're past that bottleneck, throwing more at it won't add more performance. This is the reason why OCing a 290X's memory doesn't do much either. Besides a smaller package and lower power consumption, it will also be hard to say what other gains HBM brings vs GDDR5 since we don't have two versions to compare. Like let's say Fiji with HBM and a Fiji with GDDR5 clocked high enough to have the same bandwidth. This probably sounds nitpicky, but I assume you get my point.

And yeah, I've seen that ~1GHz figure being thrown around too. I'd say Fiji will probably be about 40 to 50% faster than a 290X. Anything higher than that would be a nice surprise though.

Tbh, given both AMD and Nvidia have HBM2 on their GPU roadmap, which offers even more massive bandwidth than HBM Gen 1, I would assume any bandwidth bottleneck won't be fully removed with these upcoming cards.

Or more simply put, I think we won't see a scenario where these cards have got too much memory bandwidth that any performance gains plateau after a certain point.

 

[Nachtmaer]

Tbh, given both AMD and Nvidia have HBM2 on their GPU roadmap, which offers even more massive bandwidth than HBM Gen 1, I would assume any bandwidth bottleneck won't be fully removed with these upcoming cards.

Or more simply put, I think we won't see a scenario where these cards have got too much memory bandwidth that any performance gains plateau after a certain point.

Oh yeah definitely. I didn't mean that there'll never be any need for higher bandwidth. I was just pointing out that once you have enough for one certain chip, adding more to it won't make much difference.

When we finally go beyond 28nm, bandwidth requirements will go up again and something like HBM will be needed anyway. GDDR5 still has some life left, but once HBM becomes cheap enough, it will probably start to show up in the lower end too.

 

[Crisium]

Something interesting a person found:
https://compubench.com/result.jsp?b...=true&arch-ARM=true&arch-x86=true&base=device

It's apparently Fiji:
https://compubench.com/device.jsp?b...=AMD+Radeon+Graphics+Processor&testgroup=info

Seems legit. The 290X has max_compute_units of 44 and this Fiji has 64. Same ratio as 2816 to 4096 shaders.

I doubt we can draw overall performance just from compute though, unfortunately.

 

[riflen]

Would Titan X be underserved by its 384-bit bus (336GBs) if it wasn't for the compression tech?

I'm trying to get a sense of where this bandwidth scenario would kick in.

It's very tough to find benchmarks that expose memory bandwidth limits in recent GPUs. These are from around the time of the 980 launch:

 

[Sir Abacus]

Would Titan X be underserved by its 384-bit bus (336GBs) if it wasn't for the compression tech?

I'm trying to get a sense of where this bandwidth scenario would kick in.

Textures are already compressed loaded into VRAM and go from VRAM to the L1 texture cache compressed. As the textures are sent to the TMU they're decompressed by a special circuit in the L1 texture cache.

So if you have 12GB of assets in a scene, even compressed, the Titan will only be able to render at a theoretical max of 28 frames per second (although it would probably be well below that because of overhead). But since all the assets in VRAM are (almost?) never used in a scene this doesn't often become a limiting factor. You could deliberately chunk up VRAM with shit and render it to make a point but that doesn't really happen in gaming situations.

 

[Abilidebob]

What am I going to lose if going with a 390 or 390x this gen, as they probably use GCN 1.1 and not 1.2?
Some people told me it's not a big deal, some said it's something big(regarding tessellation and some DX12 features).

 

[Meh3D]

I'm liking the teasers they've been releasing. It certainly effective in building up excitement for me.

Think this is the latest one.

https://m.youtube.com/watch?v=kp1GrjcJSxo

 

[opticalmace]

Supposedly confirmed specs for Fury and Fury X:

http://wccftech.com/amd-radeon-fury-x-specs-fiji/

 

[tuxfool]

What am I going to lose if going with a 390 or 390x this gen, as they probably use GCN 1.1 and not 1.2?
Some people told me it's not a big deal, some said it's something big(regarding tessellation and some DX12 features).

To me it seems unlikely that they will stick with GCN 1.1 and charge those prices for the cards. Either way, even with 1.2 you're not going to get dx12_1 features. We'll just have to wait and see.

 

[Barzul]

Any word on price compared to the 980ti?

 

[Nikodemos]

Any word on price compared to the 980ti?

According to rumours they still haven't decided yet.

 

[Crisium]

Supposedly confirmed specs for Fury and Fury X:

http://wccftech.com/amd-radeon-fury-x-specs-fiji/

"Supposedly confirmed". Well said!

If true, Fury (Fiji Pro) will be as attractive as I thought. Like the 5850, 7950, 670 and 290, it is the same as the bigger card just fewer shaders and TMUs. The equal ROPs and Memory Bandwidth mean they should be within 5-7% at equal clocks. It'll be a bit more gap with the 50Mhz extra clocks but like the aforementioned cards they will all OC about the same. If priced right, it'll be the go-to card while the XT commands the performance tax for the halo effect. We shall see.

 

[Unknown Soldier]

Supposedly confirmed specs for Fury and Fury X:

http://wccftech.com/amd-radeon-fury-x-specs-fiji/

If 128 ROPs is true, this thing will certainly have brute force in it's favor over the Titan X and 980 Ti which have 96 ROPs.