Arithmatica's CellMath used in Nvidia to reduce die size

[ZenThought]

don't know whether this has been covered.

http://www.siliconstrategies.com/ar...id=N3R5JHFASZR2OQSNDBCSKHQ?articleId=19201721

Arithmatica, Inc., a company that claimed it could improve silicon performance and reduce die area with its innovations in mathematical circuit blocks, has revealed that Xilinx Inc. and Nvidia Corp. are two companies that have used its CellMath libraries.

 

[MfA]

Basically they are outsourcing development of arithmetic macro functions.

 

[moichi]

Now officially revealed "NV4x uses Arithmatica's IP" at arithmatica site.
http://www.arithmatica.com/customers_1.html


And interesting fact.

>The company's chief executive has also told Silicon Strategies that
>Arithmatica is already applying its technology to a 65-nm manufacturing process with a secret partner.
http://www.google.co.jp/search?q=ca...articleId=19201721+arithmatica+65nm&hl=ja


X-bit's article.
http://www.xbitlabs.com/news/video/display/20040428013643.html

 

[Dave B(TotalVR)]

So guesses on the 65nm partner?

I reckon its a CPU manufacturer. Intel has the clout for it but AMD has the get up and get things done attitude for it so I cant guess between either.

 

[loekf2]

So guesses on the 65nm partner?

I reckon its a CPU manufacturer. Intel has the clout for it but AMD has the get up and get things done attitude for it so I cant guess between either.

Maybe I should be impressed, but I'm not. This company as already mentioned before just sells layout optimized standard cell libraries with common used CPU or GPU functions (like floating point).

Usually you code these in VHDL, synthesize, do layout and try to get timing closure. However, usually standard cells are not optimized for high-speed. Even worse in smaller processes the wire delay is what's killing you.

From what I know, GPUs are generally ASIC designs, so unlike Intel and AMD, NVIDIA and ATI use standard off-the-shelf libraries to make their designs. The CPU companies usually take the time to manually optimize the layout. For time to market reasons it makes sense to use standard cell libs, but to get to higher and higher clock speeds it doesn't make life easier.

ATI is making more interesting moves. I read about a cooperation with a company called Intrinsity.

 

[moichi]

I download white paper of CellMath Graphics Library(which NV4x uses)
from Arithmatica's site(Log in tab, need simple registration).
http://www.arithmatica.com/

This interesting paper says "Cell based design using Artisan Sage-X library" at one example.
So it seems that CellMath IP itself is not standard cell libary,
CellMath IP based on standard cell libraries.(CellMath = Cell based designed Math IP?)

 

[JoshMST]

Well, Intrinsity's product does look interesting, but it really makes you wonder when they promise custom cell performance out of bulk 3rd party processes. You can bet that NVIDIA is also looking at Intrinsity, especially if these tools do pan out and can offer some significant performance improvements using standard cells.

There are lots of EDA software developers out there, and all of the IHV's talk to many of them (and use plenty of different software, depending on the application).

 

[MfA]

Intrinsity is pretty unique, in that it offers technology which isnt really relevant to the bread and butter customers of standard EDA tools ... and a lot closer to the tools high end processor manufacturers design in house.

 

[Geo]

Intrinsity was apparently becoming concerned about cash burn rate (ah, yes, all the Internet-bubble reminders that phrase brings!) when the ATI deal really put them on a solid footing. One could wonder if ATI had enough leverage in that deal to demand --and get-- a deal that includes industry (i.e. graphics) exclusivity, at least for some length of time.

 

[Scarlet]

You can only speculate where Intrinsity is being used inside ATI. Supposedly Intrinsity gives you a faster/denser result than Arithmatica. Intrinisty does this by custom design and using dynamic logic. My understanding is Arithmatica uses standard (static) logic and does essentially hand-tuned higher level macros for things such as ALUs. The layout is much tighter and thus the perormance and density is somewhat better than a random tool driven layout.

Maybe nV is using Arithmatica because ATI has Intrinsity tied-up with some sort of exclusivity deal for graphics? (I would if I were ATI...)