[VC][TT] - [Rumor] Radeon Rx 300: Bermuda, Fiji, Grenada, Tonga and Trinidad

[KaRLiToS]

Speculations ALERT

http://www.tweaktown.com/news/43393...deon-r9-395x2-end-2015-around-1499/index.html

AMD expected to launch Radeon R9 395X2 end of 2015 for around $1499

http://videocardz.com/54858/amd-rad...tion-395x2-bermuda-390x-fiji-and-380x-grenada

Radeon R9 395X2 “Bermuda”

Starting with Bermuda. You probably heard this name before, but no one was certain if Bermuda is the actual flagship GPU or something else. Well as a matter of fact Bermuda could actually be the Vesuvius successor (R9 295X2). Radeon R9 395X2 would be a direct competitor to rumored GTX 990 (likely dual-GM204).

Radeon R9 390(X) “Fiji”

According to 3DCenter, Fiji is “confirmed” as next flagship GPU. The R9 390X will likely be powered by Fiji XT, whereas R9 390 with Fiji PRO. Fiji-powered graphics cards are said to be the first with High-Bandwidth-Memory on board. NVIDIA will possibly adopt HBM with Pascal, so we are looking at months, or maybe even years, of AMD leadership in this technology.

Radeon R9 380(X) “Grenada”

I think it was quite obvious Hawaii, which was released in late 2013, would not be forgotten in upcoming 300 series. The rumor has it that Hawaii will be renamed to Grenada. Does it mean architectural changes? Likely not, but we expect Grenada to be more power efficient, and configured with higher clock speeds.

Radeon R9 370(X) “Tonga”

Tonga was first released with Radeon R9 285. AMD had plans to release R9 285X (this is not a rumor), however due to GTX 970 announcement, those plans had to be changed. The full Tonga is actually equipped with 384-bit interface, so R9 370X should launch with 3GB on board. We are looking at potential GM206 competitor here.

Radeon R7 360(X) “Trinidad”

Trinidad is Curacao (Pitcairn) replacement expected to launch with Radeon R7 360 series. The only information we have is limited to memory configuration of 256-bit and 2GB frame buffer. Additionally, Trinidad may be the first card of Radeon 300 series, as it’s currently planned for March (according to VR-Zone).

 

[Sequences]

It'd be a shame if only the 390 and above has HBM. Makes buying anything below that just like buying what's out on the market right now.

 

[psolord]

Yeah, also 2GB for the 385X but 4GBs for the 380X (rebranded 290X).

What good is that?

 

[Hitman928]

Yeah, also 2GB for the 385X but 4GBs for the 380X (rebranded 290X).

What good is that?

The story doesn't even mention a 385x, or did I miss it?

Edit, nm didn't see the chart at first, hopefully it's just incorrect speculation.

 

[jpiniero]

I suppose the 385X could be an OEM-only type deal although I am a bit skeptical. HBM is going to be expensive so it only being on the top end part makes sense.

 

[tential]

Yeah, also 2GB for the 385X but 4GBs for the 380X (rebranded 290X).

What good is that?

2GB HBM vs 4GB DDR5.

We'll know when we see the benches but I could see some 1080p gamers or lower that prefer higher performing 2GB HBM to DDR5.

I could see the R9 385x (speculated) being a better choice than a R9 380x in some cases.

 

[ShintaiDK]

Some of those speculations are just silly.

The 285 was to replace, not to coexist.

In case anyone forgot:

 

[Blitzvogel]

It'd be a shame if only the 390 and above has HBM. Makes buying anything below that just like buying what's out on the market right now.

Because AMD wants to reuse their older GPU cores. I was hoping for HBM across the board, accept maybe the low-mid range (360X).

 

[el etro]

I don't believe at all in a R9 385x, but it would be a very though GTX 970 competitor. With 3GB of HBM would be a much better pick. Even if not more efficient than Maxwell i would buy.

 

[raghu78]

I don't see AMD renaming Hawaii and sticking with an older architecture when they have already introduced Tonga with GCN 1.2. I expect AMD to have completely new GPUs across the stack. Tonga was a testbed chip which laid the foundations for significant perf improvements and scaling. The tiled GCN architecture will bring better perf and power efficiency. My guess is this is how the stack will resemble.

R9 390X - Bermuda XT (4096 sp)
R9 380X - Fiji XT (3072 sp)
R9 370X - Treasure Island (1536 sp)

Both Bermuda and Fiji will sport HBM. I also foresee that contrary to the available information the R9 390X will sport 6 GB HBM and R9 380X will sport 4GB HBM. I do not expect AMD to have less VRAM than GM200 based Nvidia's GTX 980 Ti / 990 when for the last 3 generations they have had more VRAM than Nvidia's competing cards. AMD's HBM bandwidth advantage will allow them to outperform Nvidia at 4k and I doubt AMD will restrict themselves artificially. I am quite sure AMD are going all out for the flagship single GPU crown. an AIO cooling solution is one of the indicators that AMD is pushing the limits of 28nm. HBM gives AMD a significant advantage which I expect them to leverage very well for the next 12 - 18 months.

GTX 570 - 1.25 GB HD 6950 - 2GB
GTX 580 - 1.5 GB HD 6970 - 2GB
GTX 680 - 2 GB HD 7970 - 3GB
GTX 780 - 3GB R9 290 - 4GB
GTX 780 Ti - 3GB R9 290X - 4GB.

Even though Nvidia had GTX 580 3 GB, GTX 680 4GB and GTX Titan 6GB they were not the higher volume products.

 

[garagisti]

Because AMD wants to reuse their older GPU cores. I was hoping for HBM across the board, accept maybe the low-mid range (360X).

I think in the mobile sector, AMD will definitely push HBM. It will have nothing to do with performance, but mostly how efficient that ram is. It's an instant reduction in power. Mobile is usually some mainstream bit slowed down enough to not sip too much juice, so it is not too far fetched to think that if AMD went HBM on mobile, then mainstream part will also use HBM. While temptation rebrand will be high, but let us hope that whatever new GCN cores they cook up are standard in both high-end and mid.

 

[Mondozei]

I don't see AMD renaming Hawaii and sticking with an older architecture when they have already introduced Tonga with GCN 1.2. I expect AMD to have completely new GPUs across the stack. Tonga was a testbed chip which laid the foundations for significant perf improvements and scaling. The tiled GCN architecture will bring better perf and power efficiency. My guess is this is how the stack will resemble.

R9 390X - Bermuda XT (4096 sp)
R9 380X - Fiji XT (3072 sp)
R9 370X - Treasure Island (1536 sp)

Both Bermuda and Fiji will sport HBM. I also foresee that contrary to the available information the R9 390X will sport 6 GB HBM and R9 380X will sport 4GB HBM. I do not expect AMD to have less VRAM than GM200 based Nvidia's GTX 980 Ti / 990 when for the last 3 generations they have had more VRAM than Nvidia's competing cards. AMD's HBM bandwidth advantage will allow them to outperform Nvidia at 4k and I doubt AMD will restrict themselves artificially. I am quite sure AMD are going all out for the flagship single GPU crown. an AIO cooling solution is one of the indicators that AMD is pushing the limits of 28nm. HBM gives AMD a significant advantage which I expect them to leverage very well for the next 12 - 18 months.

GTX 570 - 1.25 GB HD 6950 - 2GB
GTX 580 - 1.5 GB HD 6970 - 2GB
GTX 680 - 2 GB HD 7970 - 3GB
GTX 780 - 3GB R9 290 - 4GB
GTX 780 Ti - 3GB R9 290X - 4GB.

Even though Nvidia had GTX 580 3 GB, GTX 680 4GB and GTX Titan 6GB they were not the higher volume products.

I hope you're right but I think you're wrong. AMD has recycled old archs for a long time. They are still selling some GPUs which are 3 years old.

On the other hand, for those of us with Hawaii, this means that driver support is likely to be quite extended, and this is especially good for Crossfire since a single GPU isn't viable 2-3 years on. Nvidia's Kepler support has been lackluster post-Maxwell.

Nevertheless, we could very well see more efficient Hawaii cards, lower TDP, with higher clock speeds, but still with the same fundamentals.

 

[Stuka87]

I hope you're right but I think you're wrong. AMD has recycled old archs for a long time. They are still selling some GPUs which are 3 years old.

So is nVidia. Low end cards are always old architectures. There is no money is developing a new one for such cheap hardware.

 

[tviceman]

Good news if AMD is actually 1+ year ahead of Nvidia in HBM adoption. AMD needs all the help they can get right now being so far behind in perf/mm^2 and perf/w vs. Maxwell. Maxwell brought gains over Kepler that have been previously only witnessed when changing nodes AND bringing a new architecture with it. It still remains to be seen what kind of real-world game performance HBM will bring over traditional vram.

Tonga might be one of the worst chips ever from an engineering point of view. Same approximate size as Tahiti, slower than Tahiti's 280x, and if it isn't fully functional in the R285 form then it may end up as the longest running chip to have so much functionality cut out of it before getting a "full" release. A mythical 384-bit Tonga should end up quite a bit faster than GM206, given that 256-bit Tonga is essentially tied with GM206's GTX 960, but then again we're talking about a ~225mm^2 chip vs. a ~365mm^2 chip. From that perspective, being only 15% faster and consuming twice as much power is still a disgrace in engineering.

 

[jpiniero]

Tonga did get a full release (as the M295X in the iMac 5K, but it is 256-bit and 4 GB of memory).

 

[StinkyPinky]

If true, I wonder how much the R390 will cost.

 

[Red Hawk]

So is nVidia. Low end cards are always old architectures. There is no money is developing a new one for such cheap hardware.

This is true. The exception tends to be when a whole product stack needs to be replaced for compatibility, like the jump to DirectX 11. Otherwise, both Nvidia and AMD make use of old products and architectures while still introducing old ones. Most of AMD's 6000 series was VLIW5 even though the 6900 parts went to VLIW4. Nvidia rebranded GK104 under the 700 series. Reusing old parts is nothing new for either company.

 

[CakeMonster]

I'll be very disappointed if they don't update the 290 series to the latest GCN + memory compression + whatever small tweaks they will sure have researched in the mean time. Since it will be the new mid segment they should really include improvements with release almost 2 years later.

 

[Silverforce11]

Good news if AMD is actually 1+ year ahead of Nvidia in HBM adoption. AMD needs all the help they can get right now being so far behind in perf/mm^2 and perf/w vs. Maxwell. Maxwell brought gains over Kepler that have been previously only witnessed when changing nodes AND bringing a new architecture with it. It still remains to be seen what kind of real-world game performance HBM will bring over traditional vram.

Well if the leaks are accurate, Fiji will exceed the 970 on perf/w. That's a pretty big leap. R290X level of power usage, +50% performance, so a 50% efficiency gain. About as good as the Kepler -> Maxwell leap, no?

I suspect that's mostly due to HBM, by reducing the die size and TDP of the memory subsystem (which occupies a large chunk of the die and power usage).

Tonga is a more forward looking architecture than Tahiti so while its not beating it in older games (keeping in mind R285 is castrated, like the 7950 Tahiti SKU), I've seen some recent game benches where Tonga is clearly faster (FC4 with enhanced godrays) thanks to better throughput, tessellation and bandwidth compression. Its more of a stepping stone that doesn't change the status quo at the time, but it will serve them well to build upon it.

If all they did was made Tonga larger to 550mm2, rebuilt the chip for HBM support, that should give it more die area for performance gains (by reducing die space required for a memory subsystem), and with faster latency on the vram along with ridiculous bandwidth, it should be ~50% faster than R290X.

 

[Jodell88]

Trinidad should be the flagship.

<-- Lives in Trinidad.

 

[tviceman]

Well if the leaks are accurate, Fiji will exceed the 970 on perf/w. That's a pretty big leap. R290X level of power usage, +50% performance, so a 50% efficiency gain. About as good as the Kepler -> Maxwell leap, no?

The 970 loses out noticeably on efficiency to the full fat 980. http://www.techpowerup.com/reviews/MSI/GTX_970_Gaming/28.html Color me unimpressed if all GCN 1.3 can do is match a less efficient and castrated GM204 GTX970 with their highly, highly touted HBM.

Tonga is a more forward looking architecture than Tahiti so while its not beating it in older games (keeping in mind R285 is castrated, like the 7950 Tahiti SKU), I've seen some recent game benches where Tonga is clearly faster (FC4 with enhanced godrays) thanks to better throughput, tessellation and bandwidth compression. Its more of a stepping stone that doesn't change the status quo at the time, but it will serve them well to build upon it.

What are you talking about??? Tonga gets destroyed by Tahiti 280x in Far Cry 4 with Godrays and Ultra Settings enabled:

Tonga is crap. You say newer titles, well I see it's slower in Dragon Age: I and Ryse, while way way slower in Shadows of Mordor, Alien Isolation, and Far Cry 4. http://www.techpowerup.com/reviews/MSI/GTX_960_Gaming/18.html That is all there is to it. This talk of it being castrated??? Where is the proof? Real actual PROOF. There is no concrete empirical proof. And what exactly are it's specs if it is castrated? that is all conjecture. What in the world would be the reasoning for it not getting a full release 5 months after being on market? It's not that big of a chip so yields can't be bad. And as the R9 285, it already overlapped the 7950 / 280 so what would it matter if it overlapped the 280x if it could perform better? It. Makes. No. Sense.

The only thing that does make sense is looking at what the situation is right now: Tonga is the worst technological GPU release in a long, long time. Despite the gap in product release between it and Tahiti, it made zero progress in perf/w and perf/mm2 and has become a liability in AMD's ability to generate significant profit being force to compete with a 225 mm^2 chip.

 

[3DVagabond]

The 970 loses out noticeably on efficiency to the full fat 980. http://www.techpowerup.com/reviews/MSI/GTX_970_Gaming/28.html Color me unimpressed if all GCN 1.3 can do is match a less efficient and castrated GM204 GTX970 with their highly, highly touted HBM.




What are you talking about??? Tonga gets destroyed by Tahiti 280x in Far Cry 4 with Godrays and Ultra Settings enabled:

Tonga is crap. You say newer titles, well I see it's slower in Dragon Age: I and Ryse, while way way slower in Shadows of Mordor, Alien Isolation, and Far Cry 4. http://www.techpowerup.com/reviews/MSI/GTX_960_Gaming/18.html That is all there is to it. This talk of it being castrated??? Where is the proof? Real actual PROOF. There is no concrete empirical proof. And what exactly are it's specs if it is castrated? that is all conjecture. What in the world would be the reasoning for it not getting a full release 5 months after being on market? It's not that big of a chip so yields can't be bad. And as the R9 285, it already overlapped the 7950 / 280 so what would it matter if it overlapped the 280x if it could perform better? It. Makes. No. Sense.

The only thing that does make sense is looking at what the situation is right now: Tonga is the worst technological GPU release in a long, long time. Despite the gap in product release between it and Tahiti, it made zero progress in perf/w and perf/mm2 and has become a liability in AMD's ability to generate significant profit being force to compete with a 225 mm^2 chip.

Any idea why the 280X is destroying every other card in the comparison?

 

[Silverforce11]

The 970 loses out noticeably on efficiency to the full fat 980. http://www.techpowerup.com/reviews/MSI/GTX_970_Gaming/28.html Color me unimpressed if all GCN 1.3 can do is match a less efficient and castrated GM204 GTX970 with their highly, highly touted HBM.

What are you talking about??? Tonga gets destroyed by Tahiti 280x in Far Cry 4 with Godrays and Ultra Settings enabled:

You are talking about 10-20% efficiency differences here (970 is within that range vs 980 from multiple reviews). On early ES sample with most definitely pre-beta drivers. The fact it comes that close to GM204 on efficiency is pretty damn good. We've yet to see how GM200 pans out as well.

Also, don't confuse volumetric godrays in FC4 with enhanced godrays.

http://hardocp.com/article/2015/01/12/far_cry_4_graphics_features_performance_review/2#.VNgOX_mUefU

Note Tonga doesn't take a major performance hit with it, it behaves like Kepler/Maxwell.

Whereas older GCN take a major hit. This setting in FC4 is why Maxwell takes a major lead in this game compared to R290/X.

Quoting from the review:
"This graph could not make it anymore clear. The AMD Radeon R9 285 barely takes a performance hit turning on Enhanced Godrays, just like the GeForce GTX 980!

The reason for this is the fact that the Radeon R9 285 GPU (Tonga) is a newer version of GCN than the AMD Radeon R9 290/X GPUs. Tonga GCN version received improvements to tessellation, and other factors, that architecturally improve performance in a feature like this. The video card itself is slower, underpowered, gimped by memory bus and speed. However, the tessellation improvements are paying off big time in terms of Enhanced Godray performance in Far Cry 4. This shows how much one GCN version can vary from another."

 

[Silverforce11]

Any idea why the 280X is destroying every other card in the comparison?

Because GameWorks crap over all Kepler. We've seen it where Maxwell just pulls ahead so much in comparison.

 

[tviceman]

A) With respect to 390x and efficiency, you are talking about unconfirmed leaks. Neato. But I'll go with it.

B) You are trying to compare a full fledged chip's (Fiji 390x) expected efficiency to a cut-down chip's efficiency (GM204 970). Niiiiiice. That makes absolutely no sense given that the 390x should trounce the 970 AND especially (more logically) when I very easily pointed out that 980 is about 15-20% more efficient than the 970 since both end up using about the same power and have a 15-20% performance difference.

C) Tonga is still noticeably slower in every game both new and old vs. the 280x by 20% or more except Wolfenstein and Watch Dogs - but Watch Dogs doesn't count because you said as much in the other thread going on. Wait a second, you didn't think Far Cry 4 should count either because GW and Ubisoft and Hardocp sucks but here you are grasping. Doesn't matter. Tonga is still crap. CRAP. CRAP. CRAP. You've found one solitary scenario where it does well. That doesn't change anything. It's a complete dud for an engineering standpoint and AMD would have been better off not spending the R&D bringing it out in it's current state on 28nm. You conveniently ignored the fact where I proved you wrong with "it performs better with newer games" citing multiple newer games where it still just sucks and are focusing on GODRAYS. Gameworks Godrays at that.