Intel Announces FPGA on package + Xeon Product

[kimmel]

[FONT=&quot]http://www.theregister.co.uk/2014/06/18/intel_fpga_custom_chip/[/FONT]

The chip company announced on Wednesday at GigaOm Structure in San Francisco that it has started selling a Xeon E5-FPGA hybrid chip to some of its largest customers.

http://new.livestream.com/gigaom/structure2014/videos/54241911

Would be interesting to play with. Probably will be hard to get a hold of even after it does come out though. The article says present tense but the interview doesn't give a release date.

Cache coherent over QPI as well.

 

[sefsefsefsef]

This is excellent news for at least 2 reasons. 1) Intel is moving toward accelerators, which I believe is the future for energy-efficient computing. 2) Intel is selling something that has a programmable non-x86 component. Next up, they need to make an ARM core, and use it to take over the world.

 

[Vesku]

Might be connected to Microsoft's plans to use FPGAs in their Bing infrastructure:

http://arstechnica.com/information-...performance-with-fpgas-will-use-them-in-2015/

Edit: The Register article pursues this line of thinking as well.

 

[Ayah]

major question will be cost. if the fpga portion has to have an Intel-sized markup, it may not be price-competitive against just tacking the fpga onto pcie.

 

[erunion]

Do you think FPGA are cheap?

 

[cytg111]

Sounds exciting. Wonder what it will add to the thermals.

 

[ShintaiDK]

Hopefully this will bring FPGAs closer to the common man as well.

 

[TuxDave]

This is excellent news for at least 2 reasons. 1) Intel is moving toward accelerators, which I believe is the future for energy-efficient computing. 2) Intel is selling something that has a programmable non-x86 component. Next up, they need to make an ARM core, and use it to take over the world.

I'm excited. I know companies are all trying get a competitive advantage by differentiating itself from another. Software is one way and system configurations is another. By putting FPGAs in, it allows Microsoft or Google to go program in very specific functions that differentiate themselves even more. So that's good for them... and it's good for Intel.

This does two things. It sort of jumps ahead of the trend by providing a solution for the customers. That way their eyes don't wander around and start buying other parts. On top of that, it prevents ISA bloat by not needing to put in a lot of hardware for very weird instructions that only 1 or 2 customers really want.

The FPGA implementation probably won't beat out common arithmetic operations on the CPU (or GPU) due to its decrease in density and frequency but it'll allow customers to start putting in weird bit manipulation instructions that no one except them wants.

 

[kimmel]

it'll allow customers to start putting in weird bit manipulation instructions that no one except them wants.

Cryptocoin mining accelerators. :sneaky: Probably not worth it though.

 

[bullzz]

ah didnt think abt the ISA bloat. great explanation TuxDave!

 

[sefsefsefsef]

The FPGA implementation probably won't beat out common arithmetic operations on the CPU (or GPU) due to its decrease in density and frequency but it'll allow customers to start putting in weird bit manipulation instructions that no one except them wants.

Some of my recent research has been into hardware acceleration of sorting functions for big data workloads, and there is a huge amount of performance to be gained even when the accelerator (or FPGA) is running at only a few hundred HMz.

EDIT: here are a couple papers people should look into if they want to know more about what I'm talking about --
http://arcade.cs.columbia.edu/harp-isca13.pdf
http://arcade.cs.columbia.edu/q100-asplos14.pdf

 

[jpiniero]

Yeah, this is a big deal. Keeping Google/Yahoo/Bing/Amazon on x86 (as opposed to doing similar stuff on ARM) is extremely important since that is a high growth area.

 

[Ken g6]

I actually expected FPGAs to be added to CPUs as accelerators as long as ten years ago. Of course, that was before GPGPU acceleration.

I also expected them to be less expensive than they are. I wonder what keeps the prices up? Patents?

 

[ShintaiDK]

I also expected them to be less expensive than they are. I wonder what keeps the prices up? Patents?

Volume.

 

[TuxDave]

Some of my recent research has been into hardware acceleration of sorting functions for big data workloads, and there is a huge amount of performance to be gained even when the accelerator (or FPGA) is running at only a few hundred HMz.

EDIT: here are a couple papers people should look into if they want to know more about what I'm talking about --
http://arcade.cs.columbia.edu/harp-isca13.pdf
http://arcade.cs.columbia.edu/q100-asplos14.pdf

For sure. There's definitely applications which allow significant performance gains even at a density and frequency disadvantage. No doubt about it. Just don't have any wishful thinking about using an FPGA to outperform GPU DGEMM performance.

 

[jpiniero]

I actually expected FPGAs to be added to CPUs as accelerators as long as ten years ago. Of course, that was before GPGPU acceleration.

Intel's rather actively worked against accelerators of any kind; wanting to keep the lock-in to x86 and make the focus of the computer the CPU. They're only doing this now because they fear Google et all will leave them for ARM.

 

[ShintaiDK]

Intel's rather actively worked against accelerators of any kind; wanting to keep the lock-in to x86 and make the focus of the computer the CPU. They're only doing this now because they fear Google et all will leave them for ARM.

Could you provide edvidence of this? Intel even helped FPGA makers to produce FPGAs on the lastest node for years.

So I would really like to see any documentation of this.

 

[jpiniero]

Could you provide edvidence of this? Intel even helped FPGA makers to produce FPGAs on the lastest node for years.

Well, if you are talking about recent history, PCI Express is such a bottleneck that a lot of programs that could use FPGA/GPUs/other accelerators are not able to or the usefulness is limited. Given Intel's history, that's by design. Intel's slowly opening up things only because they have to.

 

[ShintaiDK]

Well, if you are talking about recent history, PCI Express is such a bottleneck that a lot of programs that could use FPGA/GPUs/other accelerators are not able to or the usefulness is limited. Given Intel's history, that's by design. Intel's slowly opening up things only because they have to.

Thats your argument for "Intel working actively against"?

Intel have worked with FPGA designers for quite some years your know.

So lets drop the random FUD and nonsense.

 

[liahos1]

as expected semiaccurate has a negative spin on it.

http://semiaccurate.com/2014/06/20/intels-fpga-announcement-simply-palpable-desperation/

i love how the dude puts up anything that might be remotely interesting behind a paywall. but all his anti nvidia and intel propaganda comes free!

like the brother of daniel nenni

 

[zephyrprime]

Do you think FPGA are cheap?

There is nothing inherently expensive about fpga's. The only reason they are expensive is because they are a niche product. For an equal number of transistors, there is not reason they would be any more expensive to manufacture than a processor. Overall, they would be much cheaper to sell because their design costs are so low because they have such a homogeneous design.

 

[waffleironhead]

Amd's torrenza never really took off. That was an fpga in an opteron socket though. Will be curious to see how well these take off.

 

[ShintaiDK]

Amd's torrenza never really took off. That was an fpga in an opteron socket though. Will be curious to see how well these take off.

You mean Torrenza?

That product never materialzed at all, did it?

 

[crashtech]

as expected semiaccurate has a negative spin on it.

http://semiaccurate.com/2014/06/20/intels-fpga-announcement-simply-palpable-desperation/

i love how the dude puts up anything that might be remotely interesting behind a paywall. but all his anti nvidia and intel propaganda comes free!

like the brother of daniel nenni

I can't tell he doesn't like Intel, not at all, that is totally unbiased coverage.

Wake me up when machines start re-programming FPGAs without human intervention.

 

[witeken]

I can't tell he doesn't like Intel, not at all, that is totally unbiased coverage.

Unbiased?

http://semiaccurate.com/2013/09/18/intel-proves-pcs-coming-back/