R9 380x rumor and speculation thread

[Mondozei]

Kitguru now also running with the story.

Kitguru says it will have "full DX12/Vilkan support". That would make it more futureproof than even the Fury cards, if it means 12.1 feature set support, which the Fury cards lack. This is key as AMD's advantage in DX12 right now is already known.

AMD could have a budget king on their hands if they are smart in their pricing strategy and the GPU is close to GTX 780 in performance levels in 1080p.

 

[Elixer]

Maybe AMD doesn't want to launch with limited numbers of cards again, or they got a big OEM order, and sucked them all up.

Who knows.

I am guessing a Nov. launch as well, would be a nice card (in theory) for the Xmas season.

 

[burninatortech4]

Maybe AMD doesn't want to launch with limited numbers of cards again, or they got a big OEM order, and sucked them all up.

Who knows.

I am guessing a Nov. launch as well, would be a nice card (in theory) for the Xmas season.

It would be great if it had some feature benefit over the 390 series to further differentiate it.

 

[Silverforce11]

Kitguru now also running with the story.

Kitguru says it will have "full DX12/Vilkan support". That would make it more futureproof than even the Fury cards, if it means 12.1 feature set support, which the Fury cards lack. This is key as AMD's advantage in DX12 right now is already known.

AMD could have a budget king on their hands if they are smart in their pricing strategy and the GPU is close to GTX 780 in performance levels in 1080p.

That just means its fully compliant with DX12, not that it supports FL12.1. It's full-die Tonga on a mature process which used to be for Apple only but now I guess they have enough volume for a desktop SKU.

 

[NTMBK]

I suspect that they specced the 384-bit memory bus before the new GCN1.2 memory compression tech was tested. Turns out that with good compression, you can feed a chip that powerful with a 256-bit bus- just like NVidia have been doing since the 680- so they disable a third of the bus, and save a whole bunch of power.

 

[3DVagabond]

I've been wanting to buy a Tonga based GPU for a long time. First AMD only offered a 2GB cut down chip with cut down memory interface, then they drip feed us 4GB almost a year later, and now after another few months later we can finally get a full chip on a cut down memory interface. Yay. I don't think we should be thanking AMD for refusing to sell us this up until now.

I wanted to buy this card a long time ago, and AMD is still refusing to release this GPU in it's full potential. I wanted to upgrade, but I don't need to. Guess I'll be waiting to 14/16nm like everyone else.

There is no reason to believe it has a cut down mem bus.

Kitguru now also running with the story.

Kitguru says it will have "full DX12/Vilkan support". That would make it more futureproof than even the Fury cards, if it means 12.1 feature set support, which the Fury cards lack. This is key as AMD's advantage in DX12 right now is already known.

AMD could have a budget king on their hands if they are smart in their pricing strategy and the GPU is close to GTX 780 in performance levels in 1080p.

Full DX12 doesn't mean DX12_1. As we've found out you can support DX12_1 and not fully support DX12.

 

[Leadbox]

Weren't the XX_1 DX specs added just to make nvidia look good and over-specced for dx 12?

 

[96Firebird]

Weren't the XX_1 DX specs added just to make nvidia look good and over-specced for dx 12?

If you believe what certain forum posters lead you to believe...

 

[Leadbox]

If you believe what certain forum posters lead you to believe...

Do you know better? If so, please elaborate.

 

[Ansau]

I looked at the specs more closely and noticed that the 380x is the same as the 280x without the 384 bit bus, but has 1gb more vram that I doubt you'll ever use.

New guess is 6% faster than a 380 or as fast as a gtx960 max overclock.

A max overclocked 960 is barely better than a stock 380. And a stock 380x will better than a stock 380, since frequencies are bumped quite a bit: 1070/1500 vs 970/1375.

The new 380x will crush the 960, between a 380 and a stock 290, and should be a tad better than a 780.

 

[96Firebird]

Do you know better? If so, please elaborate.

They are just added features. Whether a game uses them or not is up to the developers, just like ASync and any other optional feature added to DX12.

 

[Mondozei]

Weren't the XX_1 DX specs added just to make nvidia look good and over-specced for dx 12?

Of course. It's well known that NV and MS are closely collarborating on PR strategy. There are no coincidences as Kojak all taught us.

 

[Headfoot]

12.1 is largely irrelevant much like how 11.1 was irrelevant and 10.1 was irrelevant. They're just marketing bullet points that show up in maybe 1 or 2 games. GPU PhysX still shows up in more games than XX.1 features do. For a number of reasons, the market chooses not to spend dev time on xx.1 features.

 

[LTC8K6]

It would be great if it had some feature benefit over the 390 series to further differentiate it.

It does, it's GCN 1.2, same as Fiji/Fury. 390 is only 1.1

https://en.wikipedia.org/wiki/Graphics_Core_Next#1.2

 

[Headfoot]

It'll probably be a full Tonga GPU, I doubt it will have the extra features from Fiji back ported to it

 

[Techhog]

It'll probably be a full Tonga GPU, I doubt it will have the extra features from Fiji back ported to it

What features? Both Tonga and Fiji are GCN 1.2.

 

[Headfoot]

What features? Both Tonga and Fiji are GCN 1.2.

Tonga is GCN 1.2, Fiji is GCN 1.2 with some extras. Probably better to call it GCN 1.2.1 or GCN 1.3

One feature is improved geometry front end load balancing.

More specific to Fiji’s incarnation of GCN is how distribution is handled. Load balancing and distribution among the geometry frontends is improved overall, including some low-level optimizations to how primitives generated from tessellation are distributed. Generally speaking distribution is a means to improve performance by removing bottlenecks, however AMD is now catching a specific edge case where small amplification factors don’t generate a lot of primitives, and in those cases they’re now skipping distribution since the gains are minimal, and more likely than not the cost from the bus traffic is greater than the benefits of distribution.

. . . AMD has put in quite a bit of effort in to improving how geometry data moves around within the chip and how it’s used, on the basis that at this point the limitations aren’t in raw geometry performance, but rather the difficulties in achieving that performance.

As with our R9 285 review, I took the time to quickly run TessMark across the x8/x16/x32/x64 tessellation factors just to see how tessellation and geometry performance scales on AMD’s cards as the tessellation factor increases. Keeping in mind that all of the parts here have a 4-wide geometry front-end, the R9 285, R9 290X, and R9 Fury X all have the same geometry throughput on paper, give or take 10% for clockspeeds. What we find is that Fury X shows significant performance improvements at all levels, beating not only the Hawaii based R9 290X, but even the Tonga based R9 285. Tessellation performance is consistently 33% ahead of the R9 290X, while against Tonga it’s anywhere between a 33% lead at high factors to a 130% lead at low tessellation factors, showing the influence of AMD’s changes to how tessellation is handled with low factors.

The feature/ability to use HBM is a pretty big one. An entirely different memory controller seems like enough to call it a different version of GCN alone, all else being equal

 

[richaron]

There is no reason to believe it has a cut down mem bus.

There are multiple reasons to suspect it has a cut down memory bus... Where have you been?

I suspect that they specced the 384-bit memory bus before the new GCN1.2 memory compression tech was tested. Turns out that with good compression, you can feed a chip that powerful with a 256-bit bus- just like NVidia have been doing since the 680- so they disable a third of the bus, and save a whole bunch of power.

I suppose it comes down to what you believe is good enough to "feed" this particular GPU. It sounds like you're suggesting there may be only tiny performance improvements 384 vs 256bit, but 50% more memory bandwidth must equate to at least a few percent performance across the board.

I believe they are crippling this GPU with "good enough" performance to fill a certain price/performance level. They have done it from the start because it has the potential to put pressure on 290/390, chips which AMD must move because of their pro gpus.

Of course I'm not ignorant of the idea it's genuinely "not worth" enabling the entire 384bit interface due to power/performance trade offs... But given AMD's track record of managing this chips release I call BS. Even if it's true why not give us the option? That's the point of multiple skus.

 

[NTMBK]

I suppose it comes down to what you believe is good enough to "feed" this particular GPU. It sounds like you're suggesting there may be only tiny performance improvements 384 vs 256bit, but 50% more memory bandwidth must equate to at least a few percent performance across the board.

It could be a 100% increase for all the difference makes- if bandwidth isn't the bottleneck, then it won't improve performance. If e.g. triangle setup, or shader performance was the thing that limited performance gains, they could throw bandwidth at the problem with little effect. Look at the 390X->Fury X comparisons for a demonstration- the front-end is the same between both cards, and when triangle setup is the limiting factor the performance is almost identical.

I don't buy into the conspiracy theory for one simple reason- AMD are desperate for money at the moment. If they thought they could put out a significantly more competitive part simply by enabling the 384-bit bus, then they would. They're in no position to play market segmentation games when the bottom is falling out of their market share. The fact that they haven't done it tells me that there isn't a significant gain to be had.

 

[richaron]

It could be a 100% increase for all the difference makes- if bandwidth isn't the bottleneck, then it won't improve performance. If e.g. triangle setup, or shader performance was the thing that limited performance gains, they could throw bandwidth at the problem with little effect. Look at the 390X->Fury X comparisons for a demonstration- the front-end is the same between both cards, and when triangle setup is the limiting factor the performance is almost identical.

I don't buy into the conspiracy theory for one simple reason- AMD are desperate for money at the moment. If they thought they could put out a significantly more competitive part simply by enabling the 384-bit bus, then they would. They're in no position to play market segmentation games when the bottom is falling out of their market share. The fact that they haven't done it tells me that there isn't a significant gain to be had.

It's not a conspiracy mate, it's simple business. AMD must move their harvested Hawaii chips because AMD must continue to produce these chips for the Pro sector. These chips are their DP/ECC memory designs.

Tonga when operating at full potential will put pressure on the price/performance sectors of 290/390; which they can't have. Simple business, and you start throwing around accusations of conspiracy theories. Have you every wondered why there are so few performance VW cars around? Same company owns Audi & Porsche.

The idea is that you don't compete with yourself, & right now it's much more important to keep moving Hawaii chips. This idea may work in theory, and even be the best method going forwards... But it relies on brand segregation and the perceived differences in value of the products. We know full Tonga exists. We know how they've refused to sell us even 4gb versions for so long (are you going to say I'm imagining that too? please explain). And we know from die shots and basic design cues (i.e. updated Tahiti) that it's got a 384bit bus.

Tonga is the chip I was most excited for in the last few years, and I would have happily bought. But AMD has put a sour taste in the mouth of those who have actually watched this extended release.

 

[3DVagabond]

There are multiple reasons to suspect it has a cut down memory bus... Where have you been?

List them, please? I haven't seen any.

Just because it has the same shader count as Tahiti doesn't mean they share the same mem bus. Actually, the 512 bit bus in Hawaii is smaller and uses less power than the 384 in Tahiti. There is no logical reason to use a huge (relatively speaking) high draw mem bus and then cut it down to 256 when you have an actual 256 mem bus you can use.

 

[richaron]

There is no reason to believe it has a cut down mem bus.

List them, please? I haven't seen any.

And we know from die shots and basic design cues (i.e. updated Tahiti) that it's got a 384bit bus.

Easily more than "no reason". Whether you subscribe or not, you were way off base with your original statement; no matter how much of a keyboard warrior you may be.

 

[3DVagabond]

Easily more than "no reason". Whether you subscribe or not, you were way off base with your original statement; no matter how much of a keyboard warrior you may be.

So, nothing but a personal jab? I thought maybe there were multiple reasons, as you said. Cheers.

 

[AtenRa]

AMD Radeon R9 380X to cost $249

 

[happy medium]

AMD Radeon R9 380X to cost $249

Mabe that's why I see gtx970's so cheap now?

237$ shipped no rebate.
http://forums.anandtech.com/showthread.php?t=2453583

Is this going to be as fast as a gtx970?