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

Put it like this, Kabini is classed as an "APU", when it is an SOC.

 

I see, thanks. Was that a factor in choosing Jaguar instead of, say, Piledriver/Steamroller? Because Jaguar is already designed to be manufactured with a bunch of I/O blocks on the same die with the same process, while PD/SR aren't?

 

I asked this in the PS4 launch thread but it's probably more appropriate for here.

Are you at liberty to tell us which IP's Tehmash, Kabini and Kaveri are running on Dave? They are all described as GCN CU's in the slides but I'm guessing that can mean either Southern or Sea Islands?

I'm going to go with Sea Islands for Kaveri given it's HSA focus which from my laymans pov seems to be were the changes in Sea Islands are focussed.

I know you won't be able to comment on this but that's what I'm guessing the "next generation" architecture described in the PS4 releases is referring to as opposed to any as yet unannounced IP which may power the new high power offerings at the end of 2013.

 

And what's the name of this piece of silicon, apart from "the PS4 APU"?

 

I've not been following this too closely but I *think* it's 'Liverpool'. - which for the record, I love, for obvious reasons!

 

Based on quick googling, the stories talking about "Liverpool APU" are not even remotely close to what we know is in there

 

There have been plenty of names floating around. Pick your fav: Thebe-J, Thebes, Kryptos-O, Kryptos, WANI, Basilisk, Cipher and Carrizo. At least one of them should match the PS4 APU codename. So far it seems that most sources have bet on Thebes.

 

I'm hearing that PS4 GCN has 8 ACE's, each capable of running 8 CL's each. I believe Tahiti is 2 per ACE, 2 ACE's.

 

boooo

 

A 16x increase? Impressive! Do you think this is standard for Sea Islands or specific to the console implementation?

 

The PS4 is supposed to rely on existing IP only (hence the semi-custom description) so this should be standard to the upcoming graphics IP level, or GCN 1.1 as I like to call it. Remember that Sea Islands is not a technical name, it comprises Oland which has the same IP as Cape Verde and Pitcairn. Sea Islands is like Core i5: doesn't mean anything.

 

Yeah I should have used the GCN 1.1 moniker. But for that matter, who's to say Orbis isn't based on GCN 2.0 which I assume we are taking to describe the (possible) new IP launching at the end of 2013 in new high end GPU's?

 

I think the scheduling would have made that tough, but I suppose it's possible.

 

How will this translate into performance?

 

I would assume higher efficiency when swapping between graphics and compute work leading to greater performance in both when both are being run simultaneously. Given the HSA nature of PS4 that makes perfect sense since those are the types of workloads it's likely to be working on.

It's not such a big deal for discrete PC GPU's that wouldn't be dealing with as much compute work anyway due the latency issue.

I certainly wouldn't be surprised to see similar enhancements in AMD's future APU's (Kaveri and so on). Also I'm guessing that if you were to use an APU as a dedicated GPGPU processor and a discrete GPU for graphics this would also negate the need for the additional ACE's.

Perhaps Dave could comment further on that (using ther APU as a dedicated GPGPU processor for future games in combination with a discrete GPU) since that scenario has already been publicly suggested by AMD....

 

Semi-custom means a custom design that leverages existing IP. It doesn't mean there's no custom silicon in addition to existing IP.

 

Yes, but I think it means existing IP for the basic blocks (GPU, CPU cores) and custom silicon for the rest (memory controller, interconnect, etc.).

But of course, that's just a guess on my part.

 

I think it's custom to the PS4, although AMD have a habit of making all their custom work available to all customers (see all the non-x86/AMD64 OPs added to Phenom, Bulldozer etc.).

I don't think this is as much a performance feature as a programmability feature, to allow the ALMOST 2TFLOPS!!!11!!!!1 of GPU compute to be used for combined workloads. With Jaguar cores they're not going for 4GHz clocks and outright CPU performance underneath the AI, physics etc so they've obviously aiming for blended compute; rigid physics on the CPU and particle/soft body on the GPU, AI (remember Froblins?) on the GPU, shadows, AO, GI etc. on the GPU. Having multiple threads both sides of the fence lets more asynchronous processing go on, at least to my basic understanding (Im only a caveman).

 

Did you hear this before or after the Sea Islands ISA document stating there were 8 compute pipelines with 8 user-level queues was released?

 

Before, but not by much. Double confirmation is hard to find for me with my limited contacts.