Larrabee at Siggraph

Yes.
Native support for int32, float, double.
Load-store conversions to-from unorm8, uint8/int8, int16 (uint16?) and half.
Instructions for conversion to-from more exotic formats (guessing fp10/int10, 5:5:5:1, 5:6:5 etc).

 

I highly doubt that Larrabee will be seen in any consoles. Firstly, console manufacturers are so paranoid of users actually doing things with their consoles besides what they want them to do, and another x86 console would attract the 'homebrew' community beyond anything. Secondly, as someone mentioned, MS wants to 'license' the processors others design for them. MS owns the Xenos, not ATI. Thirdly, Intel has been leaning so much towards open platforms and open source that I doubt they'd dabble in consoles with this. I can more see them creating a Centrino-level platform with it for multimedia and games to rival consoles.

 

The ring bus was disclosed as being 1024 bits wide.
128 bytes per cycle should be enough to fill an entire ring segment, and a few consecutive cycles can fill the entire bus.

There was also the argument that fetching unaligned quads required specialized pipeline logic, and that filtering required excessive register file bandwidth for the VPU.

Multiple fp16 or fp32 samples in flight over the bus would be bandwidth-intensive.

 

I understand that Intel has used this analogy several times now, but that doesn't make it any more correct. Seriously, Larrabee has as much relation to P54c as Conroe does to Katmai.

 

I don't know the exact terms of the deal, but this is false. MS has the right to do with Xenos as they will, but so does ATI as can be seen by it's being the basis of R600 and the "mini-Xenos" mobile core.

Also, MS is in the business of trying to make money so I don't assume past decisions cannot be reversed.

 

I think the way to look at it is that MS own the graphics chip, but they license the IP that goes into it from ATI. They can manufactur it and control that side, but the royalties paid are for the IP of every one sold.

 

On the console front, I heard a couple of months ago that Intel had approached MS with regard to Larrabee usage in the next XBox - now what that means I don't know since I didn't speak to a primary source on the matter. But IMO there would be definite benefits for both parties involved. I don't think it's a matter of MS wanting to own or not own the IP so much as simply having a contract in place that doesn't constrain them. To get Larrabee into the box and have Microsoft themselves working on development tools and DirectX-for-Larrabee I think would be a huge win for Intel, to say nothing of the captive developer market they would then have. Enough of a win that I could see Intel offering Larrabee to Microsoft on some extremely favorable terms, and with their node process advantage, potentially at a price lower than what MS would be paying out to TSMC to begin with for an IP they own.

Now all of that said, if ArmchairArchitect says MS has already decided to continue with POWER, there's no reason to doubt that. One would think that the next consoles would already be shaping up afterall, and this whole thing would be a huge gamble for MS. But I just wanted to point out that the traditional reasons why we think MS wouldn't go back to Intel probably don't apply as such here; Intel is the one that would need MS, not the other way around, and a lot has changed in the industry between the launch of XBox 1 and the ramp-up to XBox [720].

 

"Derived" implies a little bit more than an analogy...

 

I fully understand what the term "derived" (and its derivatives ) mean.

I'm saying they are wrong. Raise your hand if you buy the statement that Larrabee is a P54c derivative. x86-64, SMT, vastly improved branch predication, etc are not things you just slap onto an ANCIENT x86 core. This is why I insist Larrabee has nothing to do with P54c. This *analogy* is just a way to relate the fact that it is a dual-issue, short pipeline architecture.

 

So they wouldn't just have a good GPU, they'd have incredible influence in this new, exciting CGPU programming model.. That would put the definite nail in PC gaming (and Linux gaming obviously). Intel is really good at Linux and open source driver support, they're good at platforms, they could really revitalize PC gaming with this thing. If they pawn it off to MS on a silver platter I will be very sad.

 

I'm curious why you would keep the U and V pipes however, surely the size of them must be minimal now?

 

I don't follow you. I get the reference, but I don't understand your point. Sorry if I'm just being dense here.

 

If its not derived from the P54c then why wouldn't you design the new core with two symmetrical pipelines rather than keeping the complex and simple pipelines?

 

Because it's an in-order core...

 

Excuse my ignorance, I don't see how that affects it.

 

Why would they lie? There's absolutely no motivation for such a statement if it wasn't the truth.

I'm no hardware guy, but it doesn't seem that unreasonable to me. Sure it's an "ANCIENT" core, but who cares if all you're really using/keeping is the stuff that hasn't changed a lot. The key here is that P54c had a really short pipeline which is ideal for the intended application.

Why is it so offensive to you that they may have worked from P54c as a base anyways? Why would they bother to reinvent the wheel as it were when they already had a perfectly suitable basis from which to work?

 

Who said it's a lie? Not me. It's an analogy.

My point is that P54c is so ancient that any "derivative" featuring x86-64, SMT, advanced (modern) branch predication, and a whole host of others would be so far from the original that it could hardly be considered a derivative.

It is an analogy meant to convey design decisions, not a literal etymology.

 

On second thought, if the PS4 guess of Cell 32iv is 1 Tflop machine (4 PPE + 32 eSPE) paired with a somewhat next gen NVidia GPU (1-2 Tflop), a Larrabee alone powered XBox will need something on the order of 2-3 GHz clock and 32-48 cores to match the flops. POWER7 road map is a 4GHz pair of 8 core CPUs, which seems underpowered flops wise compared to the Cell 32iv and Larrabee. Toss in the crazy idea of a pair of Larrabees or Larrabee paired with an ATI GPU and then you have something really interesting (but perhaps too expensive).