16 Core Larrabee ?
- Thread starter Megadrive1988
- Start date
PhilTaylor
Newcomer
To be fair though, that says up-to 32 cores. If the first production runs are 50% 16 core versions I'll leave my silverware dusty and black, if you don't mind.Oh and if they're 24 core only because they disable 8 to improve production yields, I'll give you that one.
I cant comment here in detail, but speculating about a SKU mix based on that single phrase is risky.
You're the one speculating on that - I was suggesting 16 cores, you came up with that link as evidence against it.
Right at this very moment though, I'm thinking that looking at Cell, where SPEs use about as much energy as cores on LRB, 16 cores is already quite a lot. Sure, they can do 32 cores (though then that'll likely be 24, to improve yield?), but will they find a commercially viable application for it? Well, I guess it partly depends on when.
We'll see. In 2010, I guess 32/24 should be possible. But is LRB going to be able to provide anything high-end enough that it can afford to be one of the bigger power burdens in a system?
Right at this very moment though, I'm thinking that looking at Cell, where SPEs use about as much energy as cores on LRB, 16 cores is already quite a lot. Sure, they can do 32 cores (though then that'll likely be 24, to improve yield?), but will they find a commercially viable application for it? Well, I guess it partly depends on when.
We'll see. In 2010, I guess 32/24 should be possible. But is LRB going to be able to provide anything high-end enough that it can afford to be one of the bigger power burdens in a system?
PhilTaylor
Newcomer
You're the one speculating on that - I was suggesting 16 cores, you came up with that link as evidence against it.
Right at this very moment though, I'm thinking that looking at Cell, where SPEs use about as much energy as cores on LRB, 16 cores is already quite a lot. Sure, they can do 32 cores (though then that'll likely be 24, to improve yield?), but will they find a commercially viable application for it? Well, I guess it partly depends on when.
We'll see. In 2010, I guess 32/24 should be possible. But is LRB going to be able to provide anything high-end enough that it can afford to be one of the bigger power burdens in a system?
I can assure you I am not speculating wrt to cores and SKUs. I just cant tell you what I know.
Intel have stated the max cores, as per my link. And provided a wafer shot from which a floorplan was extracted that confirms that count. Intel have made no claim as to whether or not there is a SKU with less cores. Anything except the max count is speculation at the moment.
How could you know what a core on LRB1 uses in terms of power? Or be able to compare that to Cell? I dont believe any informed commentary can be made on the power envelope.
As far as apps and usage - why would you doubt that apps that run standard APIs ( D3D, OGL ) will run on it?
As far as performance - time will tell.
Fair enough.
I *am* speculating! I've never claimed otherwise.
I read a straight number somewhere for a LRB core (30w). It seemed fair, considering, but I could of course be wrong. If they can bring it down to 15w, then it's still a lot (32x15w) ... ! I think they shouldn't want to bring out a chip that goes far beyond 350w, but if they can bring a core down to 15w, a 32 core chip with 24 cores active for yield purposes may be acceptible.
I've not commented on anything like that.
Yes, look forward to it. It's really interesting too, speculation is strong that they're trying very hard to get LRB into one of the consoles. And whatever will happen with just about anything, it will be more cores on a chip, that's for sure.
It's all about power consumption and heat now more than anything else, or so the experts seem to tell me.
PhilTaylor
Newcomer
uh, yes, we are going to ship a part that requires 32*30 watts.
yeah, right. think about it, that just cannot be right.
speculating how close to the existing power envelope of top-end cards we will be, thats informed speculation.
speculating we will be 2-3x existing cards power envelope, that isnt really giving us much credit. we are not crazy, well, at least not that crazy
.you stated:
but will they find a commercially viable application for it?
that seems to indicate some doubt as to what applications LRB1 would be used for.
at a minimum, anything done on an existing graphics part can be done on LRB1.
what beyond that can be done is...interesting...
I think something would have to be fantastically wrong with Larrabee if it drew that much.
There are low-power Core2 chips that draw half that.
Anything double-digits would be unacceptable (and outside PCI-E spec), and I'd be uncomfortable with anything over 5. Given an SPE two process nodes behind on an inferior process drew less than 5, I'd hope Intel hadn't messed up that badly.
uh, yes, we are going to ship a part that requires 32*30 watts.
yeah, right. think about it, that just cannot be right.
speculating how close to the existing power envelope of top-end cards we will be, thats informed speculation.
speculating we will be 2-3x existing cards power envelope, that isnt really giving us much credit. we are not crazy, well, at least not that crazy
Ok, so the cores have to use 10w or less (6w then maybe?). Now that would be interesting ... but you'll forgive my train of thought at least, I hope?
Of course, I was speculating on you not being that crazy, which I why I was suggesting 16 cores in the first place.
What you write however seem to suggest I should stick to PC Watch's 2007 predictions, and I guess I'll do just that for now, as you've no longer given me much reason to think otherwise.
Well, so assuming that the LRB won't be the CPU - it would then be a mix of GPU/PPU type applications, raytracing support, etc. Looking at Caustic's work and hearing John Carmack mentioning that their artists still are held back by the slowness at which their toolset pre-renders their work, I can certainly see the appeal there. I can also see the appeal of being able to flexibly mix compression/decompression work for videostreaming/encoding, graphics processing, image recognition and so on.
I just wasn't sure that people would be ready to abandon a dedicated GPU for something that may be more versatile, but also be behind the curve for current GPU work. So much is going to depend on better libraries, I guess, than the current DXs and Open GLs allow ...
It's going to be exciting. I do hope it works out, because it will force a software revolution that I'm definitely looking forward to.
Think about LRB cores as Atom + VPU. Atom has TDP=2.4W @ 2GHz, so 32 cores would only use 64W (save TDP by eliminating the FSB which is 0.2-0.5W and sse unit). Then you have the VPUs, which are pretty power hungry, but I don't want to wager a guess on the actual numbers. Then you have the memory interface and TMUs.
DK
DK
Megadrive1988
Veteran
In 2010, It's pretty obvious Intel will launch a 32 core Larrabee as the high-end product. I am also expecting, in 2010, the long-awaited 32-SPE Cell from IBM, as per Jim Kahle's comments in late 2006,
and some of the more recent Cell roadmaps:
http://img411.imageshack.us/img411/9205/roadmap1kh0.png
http://img164.imageshack.us/img164/5576/cellroadmapxk3.jpg
i.e. PowerXCell 32iv (4 PPE +32 eSPE), may be the server variant of a PS4 32SPE Cell B.E.
It seems that these chips from Intel and IBM will be the beginning, kick-starting the manycore era, as the entire industry shifts from the present multi-core era of this decade to the upcoming decade of manycore CPUs, GPUs and all kinds of hybrid GPGPU / CGPU, CPU-GPU, GPU-CPU, etc.
and some of the more recent Cell roadmaps:
http://img411.imageshack.us/img411/9205/roadmap1kh0.png
http://img164.imageshack.us/img164/5576/cellroadmapxk3.jpg
i.e. PowerXCell 32iv (4 PPE +32 eSPE), may be the server variant of a PS4 32SPE Cell B.E.
It seems that these chips from Intel and IBM will be the beginning, kick-starting the manycore era, as the entire industry shifts from the present multi-core era of this decade to the upcoming decade of manycore CPUs, GPUs and all kinds of hybrid GPGPU / CGPU, CPU-GPU, GPU-CPU, etc.
Last edited by a moderator:
Depends on when you believe PS4 will debut. If we're talking by 2012 I'd be surprised if it were more than 32 SPEs. We haven't seen anything > 8 yet and we're in the latter half of 2009 so how realistic is it to jump from 1PPE/8SPEs to 4PPE/32SPEs in the space of 1 year, or to expect SPE count to continue to grow given 3 years of identical core counts for Cell?
I actually don't see why you'd have 4 PPEs. I think you can have 8 SPEs on about the same die as 1 PPE? So in this case I'd much sooner go for 2 PPEs max, maybe just even the one.
Arwin, we don't really know how much overhead a PPE can handle, acting as command processor for SPEs. How many threads can it dispatch before it becomes the bottleneck? I'm not saying 4 PPEs is the answer, but my gut tells me 1 is not enough for that many SPEs either.
Well, already in the Cell the PPE couldn't always feed the SPEs. What see there happening instead is that SPEs start feeding SPEs, for instance.
Megadrive1988
Veteran
If PS4 is launched in 2011-2012, then I'd expect 32 SPE.
If PS4 is 2013-2014 then *perhaps* twice as many, 64 SPE, isn't out of the question. Afterall, 64 SPE has been mentioned in at least one article and shown on at least one roadmap, both of which are old.
old article (8+64 configuration):
http://web.archive.org/web/20031208...area.com/mld/mercurynews/business/5311288.htm
old roadmap:
http://techon.nikkeibp.co.jp/article/NEWS/20051027/110194/1027sce_cell_roadmap.jpg
I don't know the overhead involved here (loss of clock cycles for useful work) but I imagine that's not an ideal situation either, since the SPEs tend to do all the heavy lifting. Of course if it's a choice between feeding 4 SPEs with 1 PPE and feeding 6 SPEs with 1 PPE and 1 SPE the latter is preferable. Just saying you generally want all your SPEs performing calculations so perhaps a PPE redesign is in order.
