An LGPL-Licensed, Larrabee-Inspired GPGPU Processor

[phoronix]

Phoronix: An LGPL-Licensed, Larrabee-Inspired GPGPU Processor

A Phoronix reader pointed out this weekend the Nyuzi Processor, an GPL/LGPL-licensed design that's inspired by Intel's failed Larrabee graphics processor design. The Nyuzi Processor is fully open-source with its Verilog, documentation, tests, compiler/tools, etc...

http://www.phoronix.com/scan.php?page=news_item&px=LGPL-GPGPU-NyuziProcessor

[Svartalf]

If it's complete as claimed...

It's a hell of a lot closer than any of the other past attempts, overall. It would still require building up the rest of the infrastructure to make it a "proper" GPU- but much of that's already been done elsewhere in an Open manner. Just tying it all together.

At this point, this is where this is all going and unless the main players continue to work with us, you're going to need something like this. It's promising, if nothing else, as has been the Adaptiva Epiphany cores, for this sort of thing.

[asavah]

Now we would need a 3d printer capable of printing chips/SoCs/electronics at home.
So we could send all vendors flying to hell and make our own hardware from open blueprints.

Sigh ...
Maybe our grandchildren will ...

[profoundWHALE]

Here's a video of the standard teapot (with ~2300 triangles) rendering on a single core on FPGA running at 50 Mhz

Considering that my GPU runs at 947 Mhz with however many cores, this is really impressive.

[Maxim Levitsky]

This might be the start of an revolution....

What we need is indeed an inexpensive Fab. These days Fabs cost unimaginable amount of $$$$ but this is partially because the process is constantly advancing.
Soon it will halt, and we might find that printing an ASIC is as easy as ordering an T-shirt from zazzle.

This will be the true victory of open source and not because we need these verilog sources.
But because hardware vendors are total assholes (see TrustZone) and the only way around this is to build our own silicon.

In fact I suspect that even today one could create an inexpensive Fab using outdated process, specially tailored for small runs... We need someone that has the gut to do this.

[zanny]

Soon it will halt, and we might find that printing an ASIC is as easy as ordering an T-shirt from zazzle.

Why? Because we approaching an arbitrary end of one SI prefix, in this case, the nanometer era?

What happens if instead we called current fab tech 22000 picometers? Is it the end of the line at 14000 picometers? 4000? 1000? 512? 2? Why? Get that small and we can start talking femtometers.

We don't know unless we try, and time and time again it has been demonstrated that perceived physical barriers to smaller fabrication technologies were surmountable, and often those solutions were only known at the preceeding node. We already have fab tech demonstrations and theory approaching 4nm, and it is likely new innovations in the next years will push us smaller.

I don't see reason to believe we are going to run out of options soon. We may exhaust silicon, or need alternate gate composites, but I do not think we are anywhere near the limits of our universe in what we can create at these scales, we just have to find them. The question is more if we find them on time or slow down a bit as the science becomes harder and the research becomes more exhaustive.

[Shnatsel]

What happens if instead we called current fab tech 22000 picometers? Is it the end of the line at 14000 picometers? 4000? 1000? 512? 2? Why? Get that small and we can start talking femtometers.

The smallest known atom is 62 picometers across. You can't have individual parts in a solid object less than that, as far as the current physics theory can tell. And that's helium, which requires temperatures so ridiculously low temperatures combined with high pressures to soldify. And it's not even a semiconductor. So even if you scrap silicon and go for other elements, you're theoretically limited to hundreds of picometers per transistor.

Back to the topic... I'm the guy who submitted this, and I'd appreciate if somebody could actually analyze this thing and tell if has any potential. Or, you know, just asked the author about the state and prospects of the project.

[cb88]

Why? Because we approaching an arbitrary end of one SI prefix, in this case, the nanometer era?

What happens if instead we called current fab tech 22000 picometers? Is it the end of the line at 14000 picometers? 4000? 1000? 512? 2? Why? Get that small and we can start talking femtometers.

We don't know unless we try, and time and time again it has been demonstrated that perceived physical barriers to smaller fabrication technologies were surmountable, and often those solutions were only known at the preceeding node. We already have fab tech demonstrations and theory approaching 4nm, and it is likely new innovations in the next years will push us smaller.

I don't see reason to believe we are going to run out of options soon. We may exhaust silicon, or need alternate gate composites, but I do not think we are anywhere near the limits of our universe in what we can create at these scales, we just have to find them. The question is more if we find them on time or slow down a bit as the science becomes harder and the research becomes more exhaustive.

Because we ARE pushing up against physical limits. A silicon atom is 117.6 pm ... helium is 31pm! Trust me noone truely starts thinking about moving away from silicon untill they realize that the atoms are physically too big. And even then the biggest advances are going to be in electron mobility and power reduction which translate into faster switching gates and thus higher clock rates.

[Hamish Wilson]

It's a hell of a lot closer than any of the other past attempts, overall. It would still require building up the rest of the infrastructure to make it a "proper" GPU- but much of that's already been done elsewhere in an Open manner. Just tying it all together.

At this point, this is where this is all going and unless the main players continue to work with us, you're going to need something like this. It's promising, if nothing else, as has been the Adaptiva Epiphany cores, for this sort of thing.

Sorry for posting something so off topic everyone, but I would like to contact Svartalf regarding a professional matter but his PM inbox on Phoronix is full. I am hoping that he will be notified by email of my reply to his post so that he can make some space free.

[Lennie]

This might be the start of an revolution....

What we need is indeed an inexpensive Fab. These days Fabs cost unimaginable amount of $$$$ but this is partially because the process is constantly advancing.
Soon it will halt, and we might find that printing an ASIC is as easy as ordering an T-shirt from zazzle.

This will be the true victory of open source and not because we need these verilog sources.
But because hardware vendors are total assholes (see TrustZone) and the only way around this is to build our own silicon.

In fact I suspect that even today one could create an inexpensive Fab using outdated process, specially tailored for small runs... We need someone that has the gut to do this.

I don't really know a lot of that industry, but I always got the impression that Fab's with outdated processes don't go away, they are just re-purposed for 'lower-end' chips. Because the new low-end is the old high-end.