AMD GPUs In 2016: Polaris Lights Up The Roadmap

[Joe Black]

Good name

 

[RedJaron]

Tahiti with improvements sounds like more then a rebrand. In fact it's exactly what you described maxwell v2 as. The R9 300 series is power improvements and a few added onboard features. Nvidia and AMD both do this, no need to stint the reality.

Jimmy isn't exaggerating. Yes, NVidia rebrands things too, but AMD has been doing it more often recently. Consider that the 7850/7870, 270/270X, and 370 all use the exact same silicon. The only improvements were in software and power settings. Same went for the 7950/280 and 7970/280X. The 290 and 290X were complete new cards. The 285 was a new card in the sense it was a slight revision of the architecture, kind of a fusion between Tahiti and Hawai'i.

However most of the rebrands continue. The 290s and 390s are the exact same silicon, the 380 is the 285, and I've already mentioned the 370. It's not uncommon for a rebrand to span two generations, but AMD has now done it for three in a row. In some ways it's nice to see that GCN is still this good a performer after so long, but a new architecture from the top down would be helpful, if nothing else than a PR standing.

 

[jimmysmitty]

He's not exaggerating it. Yes, both NVidia rebrands things too, but AMD has been doing it more often recently. Consider that the 7850/7870, 270/270X, and 370 all use the exact same silicon. The only improvements were in software and power settings. Same went for the 7950/280 and 7970/280X. The 290 and 290X were complete new cards. The 285 was a new card in the sense it was a slight revision of the architecture, kind of a fusion between Tahiti and Hawai'i.

However most of the rebrands continue. The 290s and 390s are the exact same silicon, the 380 is the 285, and I've already mentioned the 370. It's not uncommon for a rebrand to span two generations, but AMD has now done it for three in a row. In some ways it's nice to see that GCN is still this good a performer after so long, but a new architecture from the top down would be helpful, if nothing else than a PR standing.

The biggest issue I see with it is that the performance tweaks have been mostly gained from GCN 1.0 and 1.1. 1.2 has some room but only in the Fiji XT. That means they are hitting their performance ceiling sooner rather than what happened with the HD7970, which after time they were able to move to the GHz clock speeds (and higher) stock and even improve tweaks and was still a decent GPU even up until the 2013 launch of Hawai'i in 2013 making it a GPU that lasted 2 years.

Hell I kept mine for 4 years.

Fiji though is almost tapped out clock speed wise and all it can hope for is performance tweaks.

Polaris can't come soon enough, hopefully it is more than just a minor revision to GCN.

 

[xHDx]

The new XB1 model will have Polaris. I think it was leaked already. Pretty exciting if you ask me.

 

[jimmysmitty]

The new XB1 model will have Polaris. I think it was leaked already. Pretty exciting if you ask me.

http://wccftech.com/xbox-one-may-be-getting-new-apu-based-on-amds-polaris-architecture/

Proven to be nothing but a mixup, it is a codename for Netflix for the XB1.

Again that's why reporting every last rumor is bad folks.

 

[Martell1977]

Fiji though is almost tapped out clock speed wise and all it can hope for is performance tweaks.

I'm not sure about this, last I heard, AMD still has Fiji locked down to where OCing is little to nothing. I did see a review that showed that OC'ing the memory a little gave a decent performance boost. I'd love to see what people could do with an unlocked Fiji board.

 

[palladin9479]

"And yet, AMD continues to hold its own in terms of absolute performance. The Radeon R9 Fury X even bests its nemesis, the GeForce GTX 980 Ti, in most benchmarks at 3840x2160."

While that is certainly significant to the fraction of one percent of the market that is at 3840x2160, the fact remains that neither the 980 Ti nor FuryX can provide a satisfactory experience to the gaming enthusiast at this resolution, where even twin cards struggle to top 60 fps in current AA Games. In addition, while 4 GB remains fine for 1440p, at this resolution, 4GB comes up a bit short... not be measuring the amount "allocated" by GPU_Z which should be accepted by now as inaccurate, but by measuring performance impact. For example:

http://www.extremetech.com/gaming/213069-is-4gb-of-vram-enough-amds-fury-x-faces-off-with-nvidias-gtx-980-ti-titan-x/2

In Far Cry 4, the Radeon R9 Fury X is fully playable at 1080p and 1440p, as are the GeForce GTX 980 Ti and the GeForce GTX Titan X. By 4K, with all features maximized, however, only the GTX 980 Ti is managing 30 FPS. The minimum frame times, however, consistently favor Nvidia at every point. We’ve decided to include the 0.1% frame rate ratio as a measure of how high the lowest frame rate was in relation to the highest. This ratio holds steady for every GPU at 1080p and 1440p, but AMD takes a hit at 4K. ...

Both AMD and Nvidia GPUs throw high frames out of band at every resolution, but the AMD Fury X tends to throw more of them, at every resolution. This is particularly noticeable at 4K, which is also where we start seeing spikes at the 4GB node. This looks to be evidence that the GPU is running low on memory, whereas the higher RAM buffers on the 980 Ti and the Titan X have no problem. With the resolution already below 30 FPS in every case, however, it’s hard to argue that the Fury X is uniquely or specifically disadvantaged.....

As in Far Cry 4, AMD takes a much heavier minimum frame rate hit at every resolution, even those that fit well within the 1080p frame buffer. AMD’s low 0.1% frame rates in 1080p and 1440p could be tied to GameWorks-related optimization issues, but the ratio drop in 4K could be evidence of a RAM limitation. Again, however, the GTX 980 Ti and Fury X just don’t do much better. All three cards are stuck below 30 FPS at these settings, which makes the 4GB question less relevant.

Assassin’s Creed Unity shows a similar pattern to Far Cry 4. AMD’s frame timing isn’t as good as Nvidia’s, but we see that issue even below the 4GB limit. The situation gets noticeably worse at 4K, which does imply that Fury X’s memory buffer isn’t large enough to handle the detail settings we chose, but the GTX 980 Ti and Titan X aren’t returning high enough frame rates to qualify as great alternatives. The frame pacing may be better, but all three GPUs are again below the 30 FPS mark.

Not trying to argue the relative strengths of the cards but just making the point that in each instance all 3 cards were below 30 fps making the 4 GB VRAM issue also irrelevant. One could argue that all one needs to do is turn down the quality settings, but when investing $1350 in GPUs, or $450 per year for a typical 3 year system life, that doesn't quite seem "satisfactory".

I guess the point I am addressing is the growth in this segment at this time will be at 1440p, whereas most of the market will remain at 1080p for the immediate future. So I think when most purchasers are looking at Polaris, Pascal or anything else, the great majority of those making purchase decisions, will be basing that decision on 1080p / 1440p performance rather 2160p.

And, to my eyes, most users consider performance per watt a secondary consideration. If two competing cards perform roughly the same and cost roughly the same, only then I think will needing a extra 100 watt PSU and maybe an extra case fan come into play.

The bad news is that new AMD GFX seems at least 6 months away. I don't see nVidia releasing anything, regardless of whether or not it is ready, until AMDs next generation cards drop. If this is accurate, I guess we can expect new AMD cards this summer and new nVidia cards this fall. I don't think either will provide a satisfactory experience at 4k so, while 4k development is certainly exciting, I hope future articles focus more on 1440p / 1080p performance. I think we'll need yet another generation to arrive before single card 4k performance brings a truly enthusiast level experience.

Would love to see some more concrete info on when we'll see the HBM2 cards w/ DP1.3 and how this may impact where monitor manufacturers go w/ regard to refresh rates. I'd be hesitant to invest in an expensive 4k monitor until we see them @ 144 / 165 Hz

4GB is sufficient for 4K gaming, but people need to stop using presets because "Ultra" and "High" turn AA up which is both completely unnecessary at large resolutions and a huge memory hog. There is no need for AA higher then 2X, if that, on a 4K resolution screen. Same is true for 4x at a 1440P resolution. Those higher resolutions are already acting like a form of AA just by sheer size of resolution.

 

[Math Geek]

Launch the Polaris, the end doesn't scare us.

you just made my night

got the song cranked up now cause ...well.... there was really no other option after reading this

 

[zodiacfml]

Mobile first approach is not surprising. Given the honest slide there showing the performance/consumption versus 28nm, it means that clock speeds would be the same as the previous generation only using a lot less power. Chips will be a pack more transistors though to increase performance.

 

[hannibal]

Mobile first approach is not surprising. Given the honest slide there showing the performance/consumption versus 28nm, it means that clock speeds would be the same as the previous generation only using a lot less power. Chips will be a pack more transistors though to increase performance.

Yep! Same as Intel CPUs have been during the last years. You don't get much faster processor, but it use less power.

 

[jimmysmitty]

Fiji though is almost tapped out clock speed wise and all it can hope for is performance tweaks.

I'm not sure about this, last I heard, AMD still has Fiji locked down to where OCing is little to nothing. I did see a review that showed that OC'ing the memory a little gave a decent performance boost. I'd love to see what people could do with an unlocked Fiji board.

The core will help more than the memory and it is not locked. If you look at most reviews, which normally get the cherry picked samples, they were able to push maybe 1100MHz. That is compared to a 980Ti which stock is 1000MHz and most manufactures to end OCs hit 1200-1300MHz stock with 1300-1400MHz boost on high end air cooling.

The biggest issue for Fiji is power and heat, it is a big chip and you hit limits which it has.

 

[NightAntilli]

Tahiti with improvements sounds like more then a rebrand. In fact it's exactly what you described maxwell v2 as. The R9 300 series is power improvements and a few added onboard features. Nvidia and AMD both do this, no need to stint the reality.

Jimmy isn't exaggerating. Yes, NVidia rebrands things too, but AMD has been doing it more often recently. Consider that the 7850/7870, 270/270X, and 370 all use the exact same silicon. The only improvements were in software and power settings. Same went for the 7950/280 and 7970/280X. The 290 and 290X were complete new cards. The 285 was a new card in the sense it was a slight revision of the architecture, kind of a fusion between Tahiti and Hawai'i.

However most of the rebrands continue. The 290s and 390s are the exact same silicon, the 380 is the 285, and I've already mentioned the 370. It's not uncommon for a rebrand to span two generations, but AMD has now done it for three in a row. In some ways it's nice to see that GCN is still this good a performer after so long, but a new architecture from the top down would be helpful, if nothing else than a PR standing.

I can agree with your last sentence. It would be nothing more than a PR standing, which would actually be nonsense for AMD to invest time and money in, considering they're already short of it as it is.

To put things into perspective using DX12 as a reference, we have the following;...

GCN 1.0 supports FL11_1
GCN 1.1 supports FL12_0
GCN 1.2 supports FL12_0
Fermi supports FL11_0
Kepler supports FL11_0
Maxwell supports FL11_0
Maxwell 2 supports FL12_1 (sort of)

For some reason, multiple people have argued in here that the jump from GCN 1.1 to 1.2 was bigger than GCN 1.0 to 1.1, and said that GCN 1.0 to GCN 1.1 can be considered a rebrand. The above alone indicates otherwise.

Note that GCN 1.1 and GCN 1.2 support FL12_0. This means that since 2013, AMD has GPUs on the market supporting pretty much all DX12 features. On top of that, they support the majority of these features on the highest tiers, including the 'hidden' 11_2 feature level. They are only missing conservative rasterization and ROV. Compare that to nVidia. Maxwell which was released in 2014, were still only capable of FL11_0!!! In fact, AMD's GCN 1.0 from 2011 has a higher rank in feature level support than 2014's Maxwell... Let that sink in for a moment...
With Maxwell 2, they included the two missing features of AMD, conservative rasterization and ROV, making them capable of advertising FL12_1. They also have a higher tier in tiled resources. nVidia didn't rebrand because they could not afford to rebrand. They had to include these features in their cards (and fast) for them to be remotely competitive to the features that GCN was offering.

However, despite all this;
- AMD's is still a higher tier in resource binding, even GCN 1.0 which is from 2011.
- Stencil reference value from pixel shader is still only supported by GCN cards, again starting from GCN 1.0.
- All GCN cards have the full heap available for UAV slots for all stages, Maxwell 2 cards are limited to 64.
- GCN 1.0 cards have two asynchronous compute engines with two queues per unit (total of 4), which allow concurrent calculations of graphics + compute. Maxwell 2 still can't do this since they're limited by their required context switch. They can do asynchronous computing, but they can't do concurrent graphics + compute. GCN doesn't have this limit.GCN 1.1 increased the compute units from two to eight compared to GCN 1.0, and the queues from 2 to 8 per unit also (total of 64). And this isn't even being used yet. Developers are starting to experiment with it.
- GCN 1.2 is the only card (including among GCN) that has minimum float precision. Maxwell 2 doesn't have it at all.

People complaining about GCN being outdated and the architecture needing to be redesigned have no idea what they're talking about. In fact, from this perspective, Fermi, Kepler and Maxwell look more like rebrands than the GCN cards. The GCN architecture is one of the best that has ever been designed in terms of longevity and being future proof. It supported so many features since 2011 that are still not being used nowadays. Why redesign something that A, is only now starting to be used, B, has not been optimized in its use yet, and C, has useful and relevant features for graphics that are still not used?

It has already been stated that Polaris is GCN again. No surprise, for the reasons mentioned above. Now I'll be waiting for the moment everyone is saying that Polaris is also just a rebrand that coincided with a die-shrink.

 

[jimmysmitty]

NightAntilli, you are missing the point. It is not so much in what it supports, in fact no one has really brought up the DX support level since most of the levels beyond the first are not performance related. The biggest issue with GCN right now is in its efficiency and currently they have reached the limit as to how many SPUs and clock speed.

And no one will say Polaris is a rebrand, a die shrink is more than just a rebrand. However short of Fiji most of the other AMD GPUs use the exact same core as previous models and before that only Hawaii was a new GPU. The changes are normally maybe faster VRAM, higher core clock speed and power improvements, mostly in software.

And why redesign? Because it is the only proper way to get real performance gains. A lot of the features you are talking about will mostly not benefit gamers performance wise. Trying to hold onto the same uArch base for too long is never a good thing. Even Intel gave up on the original Core design because they were not getting the gains they needed anymore.

 

[Vogner16]

We’re cautiously optimistic. A mobile-first approach makes sense for AMD, given the potential to reclaim lost market share with compelling hardware and an early disclosure. And while the company chose to demo a processor designed for thin-and-light form factors, hoping to catch its OEM customers’ eyes early in their design cycles, representatives admit that AMD will also have a “larger” version ready close to the first GPU’s launch, presumably for the enthusiast crowd.

couldn't have worded it better myself.

@jimmy this was my point on the 300 series thread.

 

[Vogner16]

I can agree with your last sentence. It would be nothing more than a PR standing, which would actually be nonsense for AMD to invest time and money in, considering they're already short of it as it is.

To put things into perspective using DX12 as a reference, we have the following;...

GCN 1.0 supports FL11_1
GCN 1.1 supports FL12_0
GCN 1.2 supports FL12_0
Fermi supports FL11_0
Kepler supports FL11_0
Maxwell supports FL11_0
Maxwell 2 supports FL12_1 (sort of)

For some reason, multiple people have argued in here that the jump from GCN 1.1 to 1.2 was bigger than GCN 1.0 to 1.1, and said that GCN 1.0 to GCN 1.1 can be considered a rebrand. The above alone indicates otherwise.

Note that GCN 1.1 and GCN 1.2 support FL12_0. This means that since 2013, AMD has GPUs on the market supporting pretty much all DX12 features. On top of that, they support the majority of these features on the highest tiers, including the 'hidden' 11_2 feature level. They are only missing conservative rasterization and ROV. Compare that to nVidia. Maxwell which was released in 2014, were still only capable of FL11_0!!! In fact, AMD's GCN 1.0 from 2011 has a higher rank in feature level support than 2014's Maxwell... Let that sink in for a moment...
With Maxwell 2, they included the two missing features of AMD, conservative rasterization and ROV, making them capable of advertising FL12_1. They also have a higher tier in tiled resources. nVidia didn't rebrand because they could not afford to rebrand. They had to include these features in their cards (and fast) for them to be remotely competitive to the features that GCN was offering.

However, despite all this;
- AMD's is still a higher tier in resource binding, even GCN 1.0 which is from 2011.
- Stencil reference value from pixel shader is still only supported by GCN cards, again starting from GCN 1.0.
- All GCN cards have the full heap available for UAV slots for all stages, Maxwell 2 cards are limited to 64.
- GCN 1.0 cards have two asynchronous compute engines with two queues per unit (total of 4), which allow concurrent calculations of graphics + compute. Maxwell 2 still can't do this since they're limited by their required context switch. They can do asynchronous computing, but they can't do concurrent graphics + compute. GCN doesn't have this limit.GCN 1.1 increased the compute units from two to eight compared to GCN 1.0, and the queues from 2 to 8 per unit also (total of 64). And this isn't even being used yet. Developers are starting to experiment with it.
- GCN 1.2 is the only card (including among GCN) that has minimum float precision. Maxwell 2 doesn't have it at all.

People complaining about GCN being outdated and the architecture needing to be redesigned have no idea what they're talking about. In fact, from this perspective, Fermi, Kepler and Maxwell look more like rebrands than the GCN cards. The GCN architecture is one of the best that has ever been designed in terms of longevity and being future proof. It supported so many features since 2011 that are still not being used nowadays. Why redesign something that A, is only now starting to be used, B, has not been optimized in its use yet, and C, has useful and relevant features for graphics that are still not used?

It has already been stated that Polaris is GCN again. No surprise, for the reasons mentioned above. Now I'll be waiting for the moment everyone is saying that Polaris is also just a rebrand that coincided with a die-shrink.

good points!

a 7970 is still very relevant as a card capable of playing most to all games on ultra at 60 fps in 1080p. the latest Radeon software update gave it another fps boost nobody expected.

Jimmy is right for the efficiency though. GCN uses more power for the same fps in games and therefor misses out on the entire mobility market after APU's.

GCN is great feature wise and as fast as Nvidia, and hopefully with Polaris and this "mobile first approach" they will close that gap. gain some laptop sales put the amd logo on more computers

 

[RedJaron]

People complaining about GCN being outdated and the architecture needing to be redesigned have no idea what they're talking about.

I am not saying AMD needs to radically redesign their mircoArch. I'm also not saying GCN needs to go. So either you're misunderstanding me or you're trying to twist words. As I said, the fact that GCN cards ( especially GCN 1.0 ) are still around and viable is a testament to its good overall design. When I say they need a new mArch from the top down, I mean their portfolio needs to be refreshed from the top to the bottom, not that GCN needs to be abandoned. Had I meant that I probably would have said from the bottom up. I can see if that wasn't terribly clear. However, you're still missing the point. Supporting features is one thing. Supporting those features well and being readily available is something else entirely.

The reason it needs to be a "top-to-bottom" refresh is because all those advancements you just listed are spread out over their portfolio and aren't available to everyone. GCN 1.0 cards include the vast majority of the HD 7000 and Rx 200 series and cover the whole gamut. GCN 1.1 was only two chips: Bonaire ( 260, 7790/260X ) and Hawai'i ( 290, 290X ). So, while AMD might have had a DX12 card on the market since 2013, that doesn't mean a whole lot when it didn't cover the vast majority of gamers in the mid-level GPU market.

GCN 1.2 was again only two chips: Tonga ( 285 ) and Fiji ( Fury ). Tonga might have been more meaningful than Bonaire because the higher performance level put it in the mid-level market, but the older 7950/280 and 7970/280X cards out performed it at more demanding settings. Fiji is impressive, but again the vast majority of gamers don't buy $600+ GPUs so they don't get the benefit of the new features.

Maxwell may be a bit behind supporting all the DX functions, but you can get a new Maxwell card at any price point. Conversely AMD only offers one "new" chip in the mainstream space and that's the somewhat disappointing 285 ( now 380 and 380X ). AMD needs to refresh their portfolio so they have the same generation of products readily available at each major price point. I don't much care whether that's a GCN revision or complete rebuild. I have no problem with mArch revisions. Revision is what got us from the Pentium M to Core to Sandy Bridge to Skylake. But imagine if Intel had released their updates only to certain product families.

Let's say Intel SB is AMD GCN 1.0. It was a big step over the previous arch and performance gains were great. You can get a SB CPU at every performance level and price point. Next comes IB, but it's only available for i7 and i3 chips. All Pentium, Celeron, and i5-2000 chips are actually rebadged SB. Next is HW, but that only hits the i7-4770K and i5-4460 SKUs alone. Everything else is a rebadged mix of SB and IB. Now overall, your performance levels are pretty good even with an older SB chip. But you're tired of the re-releasing and hope Skylake releases something in your budget so you can finally get a meaningful upgrade.

 

[visz963]

Could someone please define what is meant by transistor performance? something like clock speed?

 

[blazorthon]

Another thing I was thinking about is that use of GDDR5 is a smart move. It OC's far better than HBM, and you can use more of it at this stage than just 4GB. Also, the raw frequency is more, which I think should help in lower resolutions with FPS.

Youre making some rather pointless assumptions. We have little to no idea how HBM will perform in the next generation because as it is, you can't really overclock it due to software limitations. Furthermore, the 4GB limit was only for the first generation implementation utilized by Fiji. The next generation will scale higher.

Frequency alone is irrelevant for memory. The width of the bus and the latency timings are equally important for determining performance. Furthermore, GDDR5 is already often near its limits at reference clocks for current hardware.

Where GDDR5 has an advantage is in cost for lower tier cards. HBM needs a large interposer, probably even with a small GPU, and GDDR5 does not. Higher tier cards need so many GDDR5 chips that the costs may favor HBM, especially with the smaller PCB and power circuitry but the lower tier cards are far less effected by these problems.

 

[jimmysmitty]

Another thing I was thinking about is that use of GDDR5 is a smart move. It OC's far better than HBM, and you can use more of it at this stage than just 4GB. Also, the raw frequency is more, which I think should help in lower resolutions with FPS.

Youre making some rather pointless assumptions. We have little to no idea how HBM will perform in the next generation because as it is, you can't really overclock it due to software limitations. Furthermore, the 4GB limit was only for the first generation implementation utilized by Fiji. The next generation will scale higher.

Frequency alone is irrelevant for memory. The width of the bus and the latency timings are equally important for determining performance. Furthermore, GDDR5 is already often near its limits at reference clocks for current hardware.

The overclocking limits in HBM is not software limitations, it is more due to the fact that it is new and how it works is very different. The software is mostly used to just keep it from being tweaked to keep stability.

We do know that HBM 2 will perform 2x as fast. Thing is all that memory bandwidth would go to waste on a weak GPU core that can't keep it fed.

 

[blazorthon]

The overclocking limits in HBM is not software limitations, it is more due to the fact that it is new and how it works is very different. The software is mostly used to just keep it from being tweaked to keep stability.

We do know that HBM 2 will perform 2x as fast. Thing is all that memory bandwidth would go to waste on a weak GPU core that can't keep it fed.

It really isn't that different on a level relevant to overclocking. It still boils down to bus width x, frequency y, timings z. We just can't properly control these things yet because, as you said, its new. We don't really know if it needs to run at 500MHz for stability because we aren't given the option to properly control it. Considering everything that went into it that is similar to PCIe in how it is multiple lanes instead of a single wide interface, I'd be surprised if the frequency didn't actually have some headroom if we had full control over it. That it is new doesn't tell us how overclocking would affect it if we had full control because those software limitations halt our study of this.

I wasn't saying that AMD was wrong to use GDDR5 in this lower tier graphics card that was demonstrated. I simply pointed out that the supposed issues with HBM stated by cst1992 are not really known to be issues at this time. Besides, why couldn't AMD just use fewer dies for a smaller interface instead of using GDDR5? HBM could do the trick if desired. It might be more expensive, but it would still reduce power consumption a little.

Also, we don't really know that HBM2 will perform twice as fast. I've yet to see definitive proof either way, only data slides that make claims. It might be twice as fast, it might be slower, it might be faster. We won't know until it hits the market.

 

[alextheblue]

Short of the R9 Fury everything else in the R9 300 lineup is a rebrand of the previous generation and before that Only Hawaii XT was a "new" uArch.

Maxwell V2 is actually more like GCN 1.1 to 1.2, GCN 1.2 (Fiji) was all about power improvements more than anything which is why they are not quite rebrands.

Even the newest R9 300, the 380X, is just Tahiti with the improvements to Tonga although I would say their biggest power savings came from dropping the 384bit memory controller for a 256bit memory controller and faster clocked VRAM than anything else they did.

Tahiti and Tonga are somewhat different. They're both GCN and have shared/similar components, but Tahiti is oldschool GCN 1.0 and Tonga is GCN 1.2. They use different memory controllers and have other changes even above GCN 1.1 chips (Tonga uses delta compression for example). Frankly I don't know why you'd even bother bringing this up in a discussion about rebrands.

Anyway, 380 *is* a rebrand of the 285, Clocks are higher in general for the 380 cards which is nice, but it's got the same configuration. However I disagree with Chris on the 380X - this is not a rebrand. The silicon itself might very well be the same, but the configuration is different and hasn't been used prior to the 380X. It's a Tonga XT with 2048 shaders (vs 1792 in previous Tonga cards), no 200 series card used this setup.

The 380X is priced too high, however, relative to the 380. Especially if you factor in MIR, you can get a really nice XFX 380 with 4GB (to match the 380X) for $170 (after $20 MIR). That's about $50 less than an equivalent 380X, too big of a jump considering the mild increase in performance.

 

[jimmysmitty]

Short of the R9 Fury everything else in the R9 300 lineup is a rebrand of the previous generation and before that Only Hawaii XT was a "new" uArch.

Maxwell V2 is actually more like GCN 1.1 to 1.2, GCN 1.2 (Fiji) was all about power improvements more than anything which is why they are not quite rebrands.

Even the newest R9 300, the 380X, is just Tahiti with the improvements to Tonga although I would say their biggest power savings came from dropping the 384bit memory controller for a 256bit memory controller and faster clocked VRAM than anything else they did.

Tahiti and Tonga are somewhat different. They're both GCN and have shared/similar components, but Tahiti is oldschool GCN 1.0 and Tonga is GCN 1.2. They use different memory controllers and have other changes even above GCN 1.1 chips (Tonga uses delta compression for example). Frankly I don't know why you'd even bother bringing this up in a discussion about rebrands.

Anyway, 380 *is* a rebrand of the 285, Clocks are higher in general for the 380 cards which is nice, but it's got the same configuration. However I disagree with Chris on the 380X - this is not a rebrand. The silicon itself might very well be the same, but the configuration is different and hasn't been used prior to the 380X. It's a Tonga XT with 2048 shaders (vs 1792 in previous Tonga cards), no 200 series card used this setup.

The 380X is priced too high, however, relative to the 380. Especially if you factor in MIR, you can get a really nice XFX 380 with 4GB (to match the 380X) for $170 (after $20 MIR). That's about $50 less than an equivalent 380X, too big of a jump considering the mild increase in performance.

Short of the 256bit memory controller the HD 7970 and RR9 280X used the exact same core configuration (2048 SPUs, , 128 TMUs and 32 ROPs) as the R9 380X just like Tonga Pro was a HD 7950 with a smaller memory bus and the GCN 1.2 additions.

It is not the first time AMD has made subtle changes to a GPU and sent it out as a new model. The HD 3870 was a HD 2900 with slight enhancements and only came in a 256bit bus instead of the 512bit bus. Saved tons of power but didn't increase performance much.

 

[gmayol]

The fact AMD used intel CPUs, the 14nm / 16nm duality, the memory, showing a mobile device... what about CPU - APU ?
Perhaps AMD wants to show a lot without letting see anything...
It seems AMD is struggling to keep fighting battles it can win while its long term bets materialize.

 

[Vogner16]

eodeo you would buy a 970 over a 290x just because it uses more power during video idle and multi mon?

it would take years of heavy use to have that total cost be more than that of an identical system with a 970 and at that point you would have upgraded to a new gpu.

electric is cheap m8.

don't ever use power efficiency as a reason to buy one gpu over another as the math will prove you wrong. only reason to get a power efficient gpu is for people on laptops where battery life is at stake.

having said that the fix to multi mon power draw is in gpu mem clock speeds. driver clocks gddr5 to 100% when multi moniters are plugged in, and it doesn't need to.

 

[RedJaron]

Electricity is cheap only is some areas of the world. Some people pay dearly for it and as such it impacts them more than others.

And can we please lay to rest the notion that Tonga is the same thing as Tahiti? It's not. We've been over this many times. While many paper specs are the same ( shader count, texture units, and ROPs ), Tonga shares a lot with Hawai'i. It has 8 async compute engines compared to Tahiti's two, quad-shaders instead of dual-shaders, and bridgeless CFX over XDMA. It may be a small revision, but it's a revision nonetheless and is its own silicon. If you say Tonga is the exact same as Tahiti, then you should say Haswell is the exact same as Ivy Bridge.