AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread

[Blazkowicz]

1) Hawaii it's GCN 2.0 or 1.1?
2) GCN 1.1 it's second generation GCN architecture or second is GCN 2.0?
3) Rx 2xxxx "2" it's indicate second generation GCN for Hawaii at least, right?

The numbers are in the 200 range for aesthetical reason, maybe.
I guess we have GCN 1.0 : the original chips ; GCN 1.1 the Bonaire and Oland chips, also Kabini and the console APUs ; and GCN 2.0 the HSA compatible chips, so Hawaii and Kaveri.

Or we can say GCN 2.0 are GPU with "hUMA" support, the term for a CPU/GPU integration similar to what modern CPUs do on a multi-socket motherboard.
I expect Hawaii supports this, because Kaveri isn't going able to use any better/newer GPU for a while and it makes sense Kaveri would be able to talk to this big compute-friendly GPU the same way it talks to its own integrated one. (even through the relatively slow, high latency PCIe 16x 3.0 bus)
But that is not confirmed.

This 1.x/2.0 notation is an invention though, never been in any AMD communication, it's like it's been invented in forums like this one, or GPU news articles coining "GCN 1.1" to describe the Bonaire/Oland gen.

 

[Ryan Smith]

Do we actually know if the audio dsp is actually embedded on the same chip as the gpu? Maybe on some models it's a separate chip.

It's on die. It's not a separate chip.

This 1.x/2.0 notation is an invention though, never been in any AMD communication, it's like it's been invented in forums like this one, or GPU news articles coining "GCN 1.1" to describe the Bonaire/Oland gen.

That's primarily my doing, I suspect. I needed a way to quickly label and discuss the architectural differences between Bonaire and the original GCN parts. It's certainly not what AMD uses, but AMD won't use their internal name in public, so we have to call it something.

Oh, and Oland isn't GCN 1.1, it's 1.0.

 

[Blazkowicz]

Well 1.1 is an improved GCN arch. Do you call this second version or rather tweaked first?

It looks similar to going from the geforce GTX 4xx series to the GTX 5xx one, not the exact same thing but it's a low-level rework that gives a perf/watt improvement while the architecture appears _mostly_ unchanged.

 

[Blazkowicz]

Oh, and Oland isn't GCN 1.1, it's 1.0.

Really?, I always thought it was 1.1 and at this point I was thinking it's GCN 1.1 without the audio DSP. But if so be it, no problem.
I remember I thought Kabini was HSA, but I had made that up.

 

[kalelovil]

It looks similar to going from the geforce GTX 4xx series to the GTX 5xx one, not the exact same thing but it's a low-level rework that gives a perf/watt improvement while the architecture appears unchanged.

There are architectural changes between '1.0' and '1.1' (unlike GTX4xx to GTX5xx), although some of them have little relevance to discrete gaming cards.

If we are going to compare it to previous products, It seems like a similar degree of architectural change as between R600/RV630/RV610 and RV670/RV635/RV620.

 

[kalelovil]

That's a good question really. I think most people are expecting it to be somewhat "new" but I haven't actually seen any evidence so far that the "gcn level" is different to GCN 1.1 (so same as Bonaire/Kaveri/Kabini). And obviously I have no idea if it's new if it would be called "2.0" or "1.2".

According to TechReport:
http://techreport.com/review/25473/amd-radeon-r7-260x-graphics-card-reviewed

"AMD's Bonaire graphics processor has been kicking around inside the Radeon HD 7790 since March, and all the while, it's been harboring some secret features. Behind closed doors at the GPU14 event, we learned that Bonaire is based on the same "IP pool" as Hawaii, the next-gen GPU scheduled to premiere inside the R9 290X later this year.

In short, Bonaire has many of the same architectural perks as Hawaii: improved shaders (which also appeared in the Kabini APU), embedded TrueAudio DSP cores, and greater flexibility when it comes to connecting multiple monitors. Bonaire also has the same power management mojo as Hawaii, but unlike the other features, AMD made that functionality public at the 7790's launch.

​

Like Hawaii, Bonaire has shaders that support flat memory addressing and MQSAD (or masked quad sum of absolute difference) operations. With flat addressing, the idea seems to be to combine system and GPU memory into a single address space. This, among other things, should help facilitate the development of GPU computing applications."

Perhaps there has never been such a thing as 'GCN1.2' or 'GCN 2.0' after all.

 

[Blazkowicz]

Or a wild guess : Kabini is GCN 1.1 (and maybe consoles), Bonaire is GCN 1.2 as well as Hawaii. This is how the above reads to me.

 

[mczak]

There are architectural changes between '1.0' and '1.1' (unlike GTX4xx to GTX5xx), although some of them have little relevance to discrete gaming cards.

If we are going to compare it to previous products, It seems like a similar degree of architectural change as between R600/RV630/RV610 and RV670/RV635/RV620.

The architectural differences between GCN 1.0 and 1.1 cards (with the latter being Kabini/Kaveri/Bonaire) underneath the hood actually seem quite a bit larger. Most of the differences between the r6xx card generations were invisible to a driver even, whereas most of the code for chip initialization etc. (the kernel part of the code) is completely split between "SI" and "CIK" cards in the open source driver ("SI" and "CIK" are the names the driver uses for them).
But I can't tell if Hawaii would be the same generation as CIK really, the code for supporting it hasn't landed yet .

According to TechReport:
http://techreport.com/review/25473/amd-radeon-r7-260x-graphics-card-reviewed

"AMD's Bonaire graphics processor has been kicking around inside the Radeon HD 7790 since March, and all the while, it's been harboring some secret features. Behind closed doors at the GPU14 event, we learned that Bonaire is based on the same "IP pool" as Hawaii, the next-gen GPU scheduled to premiere inside the R9 290X later this year."
...
Perhaps there has never been such a thing as 'GCN1.2' or 'GCN 2.0' after all.

Hmm ok I missed that so maybe it really is exactly the same. Wouldn't be surprised though if it's still _slightly_ different, but I guess it'll most likely end up using the same code as the "CIK" chips then.
For instance the DCE block is probably going to be different (should support now hdmi 2.0 for instance) though obviously I have no clue how much different.

edit: fix up "cik" name - "ci" is also used because some functions (like dpm) are further split between the apu cik chips (kaveri/kabini) and the discrete cik ones (just bonaire), and "ci" denotes the latter.

 

[Dave Baumann]

This 1.x/2.0 notation is an invention though, never been in any AMD communication, it's like it's been invented in forums like this one, or GPU news articles coining "GCN 1.1" to describe the Bonaire/Oland gen.

Correct, there is only GCN. There are roadmap families that are incorrectly used to denote IP, but really there is just graphics (and other blocks) IP versions.

 

[Dave Baumann]

Do we actually know if the audio dsp is actually embedded on the same chip as the gpu? Maybe on some models it's a separate chip.

It is in the chip. That's how 7790 can get the feature 9and why others in the stack do not receive it).

 

[3dilettante]

Or a wild guess : Kabini is GCN 1.1 (and maybe consoles), Bonaire is GCN 1.2 as well as Hawaii. This is how the above reads to me.

If you believe this:

http://cgit.freedesktop.org/~airlied/linux/tree/drivers/gpu/drm/radeon/cik.c?h=drm-next

Kabini, Bonaire, and Kaveri are close enough to be considered the same category.
The description of the consoles and the appearance of various features from each in the Sea Islands docs and the driver description indicates the consoles are also close.

The transitive property applied to the information that Hawaii is of the same lineage as Bonaire would then point out how far AMD's graphics architecture has evolved or not evolved over the muddy Sea Islands launch from earlier this year.

 

[redstrat]

i think now we have four successive iterations of a GCN architecture :
1) original GCN - "GCN 1.0"
2) "GCN 1.1" - Bonaire
3) "GCN 1.2" = new powertune + true audio + new eyefinity + something - Bonaire XTX
4) and finally "GCN 1.3" = new crossfire + something new ( doubling of the geometric performance and a new balance SP/ROPs indicates a change in the architecture i think ) for flagship brand new - Hawaii.
AMD cals it all simply "GCN", because they dont want make a difference in the new R line. What do you think guys?

 

[3dilettante]

i think now we have four successive iterations of a GCN architecture :
1) original GCN - "GCN 1.0"
2) "GCN 1.1" - Bonaire
3) "GCN 1.2" = new powertune + true audio + new eyefinity + something - Bonaire XTX
4) and finally "GCN 1.3" = new crossfire + something new ( doubling of the geometric performance and a new balance SP/ROPs indicates a change in the architecture i think ) for flagship brand new - Hawaii.
AMD cals it all simply "GCN", because they dont want make a difference in the new R line. What do you think guys?

Per AMD, a driver update should make 2 turn into 3 (edit: mostly?).
Apparently, 3 shares the same pool as 4.

At least the crossfire connector on the 260X differentiates it from the 290, which has dispensed with that.

 

[kalelovil]

Per AMD, a driver update should make 2 turn into 3 (edit: mostly?).
Apparently, 3 shares the same pool as 4.

At least the crossfire connector on the 260X differentiates it from the 290, which has dispensed with that.

The 260X supports it the same new XDMA bridge-less Crossfire technology as the 290X (driver support will probably coincide with the 290X release), according to PC Perspective.
http://www.pcper.com/reviews/Graphics-Cards/AMD-Radeon-R9-280X-R9-270X-and-R7-260X-Review/AMD-Radeon-R7-260X-%E2%80%93-Bonaire-wa

"As for the updated CrossFire technology, currently being dubbed XDMA (External DMA I think), the R7 260X supports it but we are going to wait for the release of the R9 290X (that also integrates it) to see how it changes Eyefinity and 4K CrossFire configurations. More soon!"

The R7 260X boards do still have the crossfire connectors however.

Perhaps XDMA also requires PCB changes which the 260X has but the HD7790 does not, meaning it would be present on 260X cards for backwards compatibility with HD7790 cards for 260X-HD7790 Crossfire.

(doubling of the geometric performance and a new balance SP/ROPs indicates a change in the architecture i think)

Pitcairn has double the geometric perfomance of Cape Verde and Tahiti has a different SP/ROP ratio than Pitcairn, despite them all being 'GCN1.0'.
The GCN architecture is designed to be quite flexible in this manner, so a new ratio would not by itself mean a change in architecture.

 

[redstrat]

Per AMD, a driver update should make 2 turn into 3 (edit: mostly?).
Apparently, 3 shares the same pool as 4.

At least the crossfire connector on the 260X differentiates it from the 290, which has dispensed with that.

yes you right, looks like Bonaire = Bonaire XTX
http://www.sweclockers.com/nyhet/17719-amd-radeon-hd-7790-stoder-trueaudio-med-kommande-drivrutin
Therefore 3 generations GCN if its true

 

[pjbliverpool]

I see no evidence to suggest there are more than 2 levels so far with volcanic islands referring not to an IP level but rather to the Rx 2xx family.

I'm currently expecting R9 290x to be identical IP wise to Bonaire however there are probably a few more HSA enhancements to Bonaire of which we're not yet aware which will hopefully be detailed with the 290x launch.

For what its worth it sounds like both consoles also share the 290x IP just with the audio block stripped out or even just disabled.

 

[Psycho]

Slides (but no comparative performance ones): http://videocardz.com/46571/amd-radeon-r9-290x-confirmed-feature-64-rops

 

[Wynix]

Is that the first confirmation of 64 ROPs?

 

[fellix]

Ha, I knew it!

The 4K craze is evident, but we really know that this will be mostly useful for down-sampling.

 

[pjbliverpool]

Sweeet! So AMD have doubled setup rate and ROP performance while only increasing CU count by around 40%. So perhaps they are expecting games to be less shader/fetch bound this generation compared to other parts of the GPU thanks to the relatively low CU ratio on both next gen consoles. Interesting that they haven't increased bandwidth by much though despite having the obvious opportunity to do so (another 40% bandwidth is possible if they went with the highest speed memory). One can only assume they don't think they need it which raises the question of whether the next gen consoles will have a memory bandwidth surplus. The XB1 GPU which is truly puny compared to this only comes in with 20GB/s less bandwidth!