AMD: Sea Islands R1100 (8*** series) Speculation/ Rumour Thread

[Love_In_Rio]

I really wonder, why nobody posted the link to the Sea Islands ISA manual so far.

And even in there, Sea Islands is abbreviated C.I.
Found by Locuza from 3DC.

Edit:
I just see that this gets dicussed in th HD7000 thread for some reason. Dave Baumann said, that the currently available "Sea Islands" GPUs (Oland) are actually not C.I., i.e. they are still SI and don't have the new features.

That changes are practically describing the Durango CPU-GPU's behaviour with ESRAM.

 

[Gipsel]

That changes are practically describing the Durango CPU-GPU's behaviour with ESRAM.

That also caught my eye. But it's not necessarily just the behaviour with the eSRAM.

 

[Love_In_Rio]

That also caught my eye. But it's not necessarily just the behaviour with the eSRAM.

What differences can be seen from the architecture overview?. Am i reading it well that LSM has gone from 64KB per CU to 128KB per CU ( now each SIMD has 32KB assigned )?.

 

[Gipsel]

What differences can be seen from the architecture overview?. Am i reading it well that LSM has gone from 64KB per CU to 128KB per CU ( now each SIMD has 32KB assigned )?.

That's actually an error in the manual. That part is copied from a manual of Evergreen or Northern Islands. It's still 64kB. Such errors are unfortunately quite common in the ISA manuals.

 

[boxleitnerb]

Semi Accurates take. Normally I don't really care for his stuff, but this time I completely agree. AMDs lineup 2013 is a chaotic mess without any logic behind it:

There are CI parts in the HD7000 line that was SI only up to this point. Why? That doesn’t even have an explanation much less a good one, it just is. To make matters worse, there are going to be HD8000 parts from the CI family and the SI family, but not a damn bit of rhyme or reason as to what goes where here either. There is also no real way for the buyer to understand which part of which line is from what family, not that there is enough functional difference between SI and CI to actually notice. Even the drivers don’t change, and that is part of what kicked this whole ‘conspiracy’ off in the first place.

AMD has a line that is a mess, two lines that are a mess really, and for some reason seem intent on making things worse by releasing lots of parts across lots more families without any logic behind the moves. So the press on the call were annoyed in short order at the illogic of what was said, the lack of clarity that the few actual data points brought, and worse yet the stonewall on even the most short term details. To deal with the snowballing mess, the presenters seemed to go off script and confused things even more.

So in the end, what do you have? Exactly what SemiAccurate told you was coming in October, basically nothing. It is a ho-hum collection of bug fixes, driver updates, and the odd clock bump that process evolution has gifted AMD with. As we said at the time, there are no changes, think evolution and we barely even got that.

http://semiaccurate.com/2013/02/18/...ounding-amds-sea-islands-ci-end/#.USJA2WdLTpU

 

[no-X]

AMD said, that GPU sales are going up. Would it be more logical to launch a new generation of products if the actual one sells better than expected?

 

[silent_guy]

TBH, other than the way things were communicated (they should have just shut up right from the start, like Nvidia, where all we have is an architecture name and a very vague time frame), I don't see what the problem is. Why should anybody care about GCN2? GCN is very good as it is. If the next big chip fixes the most obvious issue (high power), adds a few units, or makes use of some process tweaks, it should be a very fine product.

So what if they rebadge some older products into a new line? That's exactly what they should do to optimize on R&D.

 

[jimbo75]

It's just typical whining from certain quarters tbh. If AMD says nothing then we all jump to our own conclusions, when they give a roadmap that states is subject to change, then change it, it's all doom and gloom.

Charlie was probably right about Sea Islands being a mediocre upgrade, which is why it's been brought in at certain unglamourous segments instead of near the top end. It fills some gaps in the lineup, strengthens others...but there's no point in releasing it for 7800 and 7900 if it's only 10% faster than what they already have.

 

[UniversalTruth]

but this time I completely agree

Here he wrote so many lines basically saying nothing valuable.

but there's no point in releasing it for 7800 and 7900 if it's only 10% faster than what they already have.

I agree. It would be a mediocre upgrade like the HD 6970 was.
So, AMD will have a card faster than Titan only sometime next year, perhaps in the end of 2014 with the progress to 20 nm.

 

[LittleJ]

Unless they decide to produce something well over 400mm2 and address Tahiti's weaknesses, be it setup rate, ROPs, whatever that might be, then yes it probably won't be until 2014/2015 before either AMD or NV have anything faster than this GK110 Titan.

 

[jimbo75]

So, AMD will have a card faster than Titan only sometime next year, perhaps in the end of 2014 with the progress to 20 nm.

It should always be this way tbh. It's unrealistic to expect AMD to compete on equal terms with a die that is ~40% bigger.

 

[DavidGraham]

AMD said, that GPU sales are going up. Would it be more logical to launch a new generation of products if the actual one sells better than expected?

GPU sales have been going up before ? why stop now ?

Also I dont think they are selling better than expected ! either that or their expectations are abnormally low ! they had to compromise several times , once with price reduction and another with crazy unprecedented game packages .. that is not a sign of good corporate economy .

 

[Dave Baumann]

Semi Accurates take. Normally I don't really care for his stuff, but this time I completely agree. AMDs lineup 2013 is a chaotic mess without any logic behind it:

How so? The channel line is very clear; as for anything else in the channel stack it hasn't been announced so how can it be a mess?

Notebook line of 8000M series is in the process of being announced, for the ones that are available. Contrary to Charlies comments they have been fully launched and sites have even reviewed it.

Also I dont think they are selling better than expected ! either that or their expectations are abnormally low ! they had to compromise several times , once with price reduction and another with crazy unprecedented game packages .. that is not a sign of good corporate economy .

Yet, this is the historical precedent from NVIDIA. I fail to see how all of a sudden this is poor corporate economy.

 

[DavidGraham]

Yet, this is the historical precedent from NVIDIA. I fail to see how all of a sudden this is poor corporate economy.

Yep , NVIDIA did the same when it had rough time selling it's GPUs .

 

[Dave Baumann]

Yep , NVIDIA did the same when it had rough time selling it's GPUs .

So Q4 and Q1 was rough for them?

 

[DavidGraham]

So Q4 and Q1 was rough for them?

If they did the same , then why not ! did they do the same ?

 

[Dave Baumann]

Yes they are bundling right now (though the popularity of the bundle doesn't seem to be going down well) and to match Never Settle in Q4 they bundled both Borderlands 2 and Assassin Creed 3 on a wide section of their line; I can quite easliy see a path to that being more costly than Never Settle.

 

[Andrew Lauritzen]

No, not in my opinion. It would be a complete mess if one workgroup or wavefront could write in the LDS region or even the registers of another one. ... Or can one thread on a intel CPU with hyperthreading read the registers of the other thread?

We're talking about memory with addresses here (how would you even access registers OOB on the CPU? heh), so while it may be organized more like registers in GCN, that's not how it looks/works in the API. And yeah, any CPU core can read/write the same memory as any other obviously...

It's about out-of-bounds accesses... the only point where it becomes relevant what they do is where they start to interfere with process isolation. Not running two processes at the same time on the same CU seems like a pretty good solution to the problem to me, especially since you're not going to have multiple DMA buffers running simultaneously with current OSes anyways.

But hey, cool if there's that extra level of safety, just seems unnecessary compared to the simpler solutions. Can't really complain about extra features in any case, I'm just surprised

How so? The channel line is very clear; as for anything else in the channel stack it hasn't been announced so how can it be a mess?

OEM-only rebrandings into the 8xxx series naming isn't a mess? And sure NVIDIA has done similar stuff, but that doesn't exactly justify it as a roadmap plan from day 1.

 

[boxleitnerb]

How so? The channel line is very clear; as for anything else in the channel stack it hasn't been announced so how can it be a mess?

Notebook line of 8000M series is in the process of being announced, for the ones that are available. Contrary to Charlies comments they have been fully launched and sites have even reviewed it.

For starters, Sea Islands has been dubbed a new GPU architecture at the financial analyst day 2012. But from the looks of it, you can barely call it that. Is it *really* new or only rebrands, who knows? Some HD8000M parts are old tech, some aren't. Where Oland/Mars fits in, isn't clear either. Is it a GCN1 GPU or is it "new"?

It would be more logical if there only was one family of GPUs that is used for mobile and desktop. The Pitcairn GPU is used for the 7850/7870 and for the 7970M (yeah, that makes sense...), so why make different families out of it? Just an example. 4-5 GPUs, 4-5 codenames. Add XT, Pro, LE at your leisure, that provides enough combinations. And for each SKU (and family) provide the tech level like GCN x.y.

Done, beautifully structured, logical and easy to understand.

 

[Dave Baumann]

Some HD8000M parts are old tech, some aren't. Where Oland/Mars fits in, isn't clear either. Is it a GCN1 GPU or is it "new"?

Again, that was already made clear. Oland/Mars is the first Sea Islands chip to have been released and it is based on the same IP level as Tahiti/Verde/Pitcairn. This is also the first new product to have been announced.

It would be more logical if there only was one family of GPUs that is used for mobile and desktop. The Pitcairn GPU is used for the 7850/7870 and for the 7970M (yeah, that makes sense...), so why make different families out of it?

Notebook and desktop GPU's have never followed the same naming convention, this is the case for everyone. The naming conventions are also there to dictate segmentation and notebook segmentation is typically one level up from desktop (i.e. a "performance" chip in desktop is an Enthusiast product on notebook, an enthusiast chip for desktop often never makes it to notebook). OEM's making enthusiast level gaming notebooks don't want them being compared to a lower category of product.