Larrabee at GDC 09

[TimothyFarrar]

LRB may be a wonderful dx10 gpu, but do you think it will work well for a dx 11 gpu? I am referring to Tessellation specifically. I can't see how they will be able to run the tessellation efficiently in software. (Though many said the same about rasterization )I guess they'll add some new instructions to the ISA.

Seems like the non-16-aligned number of nodes per patch, and variable number of outputs post tessellation would be challenging to map to a fixed 16-wide SIMD efficiently. If the GT3xx MIMD rumor is true, perhaps this is an area where at least NV's GPUs will have a distinct advantage...

 

[fellix]

http://74.125.77.132/translate_c?hl...wafer/&usg=ALkJrhg3XB_p0k-EbK3dfqBm1W0uZUqXhQ

'First Larrabee wafer' - don't know if something got lost in translation or what. Not sure if I have seen those slides before either, from IDF, might be new.

The picture is quite fuzzy, but I can count ~194 dies in that wafer there.
That would imply no more than 300 mm² of die size.

 

[bowman]

The picture is quite fuzzy, but I can count ~194 dies in that wafer there.
That would imply no more than 300 mm² of die size.

Sorry, I should've edited once it was discredited. Someone else posted a link to the actual Larrabee wafer. The chips are large, don't worry (or worry..).

 

[nAo]

Someone thinks it's too small and someone else thinks it's too big, how funny is that?!

 

[fellix]

Ah, OK.

Looking the other link now, and there is ~ 76 structures, as far as Photoshop can see, so in my book the thing should be ~688 mm²!

 

[Tridam]

Hard to tell the exact die size from the picture I took, that's why I said "seems to be around 600 mm²".

 

[liolio]

I can't remember in which thread I read this but this size is pretty much in line with some early leaks/rumors of a 32 cores larrabee being around 2billions transistors and so that it should in the same ballpark as Intel super high end server processor (don't remember the name it's a 6 cores with bunch of cache).
That's too big 8O

 

[Jawed]

Looking the other link now, and there is ~ 76 structures, as far as Photoshop can see, so in my book the thing should be ~688 mm²!

I got about the same size - it was so ridiculous I didn't bother posting it!

Jawed

 

[fellix]

Yup -- too bad not to be true.

Just think of what monstrous substrate size would be needed to bound on such a die!

I can't remember in which thread I read this but this size is pretty much in line with some early leaks/rumors of a 32 cores larrabee being around 2billions transistors and so that it should in the same ballpark as Intel super high end server processor (don't remember the name it's a 6 cores with bunch of cache).
That's too big 8O

Tukwila is 699 mm² in size, with 2+ billion transistors, so... oh well, Intel can afford such endeavor, anyway.

 

[3dilettante]

Intel would also charge an order of magnitude more for that Tukwila.

Hopefully Intel clarifies some more on just what we're seeing and how it relates to the final product, and which marketing line it's sticking by.
A power-efficient GPU that won't beat standard rasterizers with a die size of an enterprise-class CPU?

 

[Farid]

Let's recycle a meme here

SIX HUNDRED SQUARE MILLIMETRES?

Yeah, this better be a test wafer on a 90nm process or something. That or, they better start looking into some viable business model for GPU cards sold at a loss.

Then again, they could sell that at a high price. And it's not like the chip shouldn't be "binable" with how parallel it is. But still.

Someone thinks it's too small and someone else thinks it's too big, how funny is that?!

The guys who think it's too small are probably big posters in the Console Forum. "More" and "Not enough" are commonplace there.

Not that I disagree with that school thinking when it comes to technology, mind you.

 

[nutball]

Surely the comparison with the die size / price of Tukwila aren't really that relevant. The selling price of Tukwila is set by the size of the market (small) and the price that the market will bear (large). It's not really set by the cost of production.

 

[3dilettante]

A chip that size could really eat into the bill of materials for a graphics board that competes in a sub-500 dollar market.

It would depend on what process it's on, and the yields. Itaniums have been estimated at maybe something over $100 to produce.

That's not a big deal at 2-4K per chip.
That's not so great on a graphics card, unless this is only for professional products.

 

[nutball]

Sure, but I've been told here before that cost of production for Intel is a highly nebulous concept and depends on how they want to to their accounting.

 

[3dilettante]

Intel can't account away the costs associated with all the other components of the graphics board, since many of those would need to be purchased externally.

The consumer graphics field's price brackets only leaves so much revenue that can go towards the chip.

 

[Arnold Beckenbauer]

Some words about Ct:
Intel Gets Ready to Push Ct Out of the Lab

But even the Ct story gets a little murky when you start talking about manycore. Larrabee, Intel's first x86 manycore architecture, which coincidentally provides a lot of data parallel capability, is not the principle target of Ct -- at least not yet. As we reported last year, the first implemention of Larrabee will be targeted to graphics and visual computing applications, not the more general-purpose technical computing applications (seismic analysis, financial analytics, scientific research, high-end imaging, etc.) that Ct is aimed at.

 

[V3]

I always thought Larrabee was going to be on the large size but not that large. Though it's hard to estimate from the picture. It's probably around the size of GT200.

 

[silent_guy]

The consumer graphics field's price brackets only leaves so much revenue that can go towards the chip.

They can claim that the majority of the fab investments were already written off by producing CPU's, a luxury you don't have with an external fab.

If you only take into account the cost of a raw silicon wafer and the cost of operating a fab (electricity, man hours, maintenance etc), your chips will be very cheap indeed.

Reality will probably somewhere in the middle, I suppose, and impossible to estimate for outsiders.

 

[iwod]

90nm would be too good for an Test Case. May be 65nm? If that is the case then 45nm would properly bring it to ~ 400mm2....which seems reasonable enough.

But if this is an 32 Core version @ 400+mm2 die, i think it will have a hard to complete with ATI and Nvidia.......

However i think it wont mean much to us consumers and gamers.... It would have huge impact on Professional Market. According to recent figure this contribute to 40%+ of Nvidia's revenue ( or profit i cant remember which )

 

[DavidC]

Ah, OK.

Looking the other link now, and there is ~ 76 structures, as far as Photoshop can see, so in my book the thing should be ~688 mm²!

Really?? This is supposed to be wafer for Nehalem-EX(Enterprise class Nehalem with 8 cores/24MB L3/4 QPI links/4 channel FB-DIMM): http://blogs.zdnet.com/processors/?p=194

Nehalem-EX looks significantly larger to me.