RUMOR: Intel looking to buy Altera

[miguelvp]

Altera already has hard core ARM processors in their offerings.

The reason for FPGAs on a data center is to lower energy consumption as well as speeding things up to remain competitive. For example Microsoft was brining them to the datacenter.

http://www.datacenterknowledge.com/archives/2014/06/27/programmable-fpga-chips-coming-to-microsoft-data-centers/

Intel already announced a while back they are working of FPGA fabric for their Xeon processors.

http://www.datacenterknowledge.com/archives/2014/06/19/intel-offer-hyper-scale-operators-ability-reconfigure-cpus-dime/

I can't find the reference for Google, but I did read they were also doing that.

But having the fabric on the chip will be even more efficient than going through an external bus. Altera already has that know-how using their Avalon interconnect.

Not sure what progress has been made, but it's coming.

Edit: I forgot to mention, Altera has been using Intel's tech for their higher end FPGAs for a while now.

[Someone]

A co processor fpga communicating via the pcie bus sounds cool, kind of like what a gpu does. But what are the practical uses, and why hasen't it already been done? More and more instruction sets are being added with every generation, seems like it's working pretty well.

It has been done but only huge corporations have the scale to benefit from rolling out FPGA acceleration right now, data centres using FPGAs for packet filtering and queuing is an extension of their extensive use in packet network infrastructure. There is something of a chicken and egg situation where you don't see applications because the hardware is obscure/expensive/fragmented, this changes as soon as a common fabric is available in large numbers of consumer CPUs, just as there was offload to GPUs once they became somewhat uniform and commonplace it just needs an installed base to become viable. An FPGA can change context fast enough for it to be doing many different accelerations in professional workflows, as the user pulls up a tool or picks a command the appropriate acceleration is loaded and fed data. In a common consumer example, rather than needing "hard" accelerators for video encode/decode appropriate accelerators can be loaded as needed for different codecs as the user needs them, power constrained devices still rely on application specific cores to achieve their performance and with desktop computing slowly fading into the minority it will allow the transition of greater flexibility into lower power platforms.

[daqq]

Exactly on topic of FPGA-CPU integration usefulness:
https://forums.xilinx.com/t5/Xcell-Daily-Blog/FPGA-based-Ryft-ONE-search-accelerator-delivers-100x-performance/ba-p/667171

[AndyC_772]

It looks to me as though that product is all about performing specific operations on a large data set independently of any CPU. The relationship between the two looks to be more along the lines of "here's the data, here's a set of rules which define what to do with it, call me when you're done".

[andersm]

AFAIK, no-one is offering products with tighter integration. I don't believe throwing arbitrary FPGA logic into a tightly optimized CPU pipeline would work very well. For applications like that, there are several IP companies offering customizable cores (eg. Cadence Xtensa, Synopsys ARC, MIPS CorExtend).

[daqq]

If Intel decides to do further integration, they will probably do it more vigorously, say, putting PLD core to an addressable memory (just like how Xeon Phi works).

This is pretty much the same as Xilinx do with their ZYNQ line - two ARMs surrounded by a lot of programmable fabric. Both the fabric and the processor can access the memory system through the bus system.

[miguelvp]

Both Google and Microsoft (Bing) use FPGAs on their data centers or at least they are experimenting with that notion.

I could definitely use a Xeon processor with FPGA fabric, might not happen yet, but it's coming.

Think about it, configurable hardware driven protocols on a data center saving some watts.

Maybe Intel is not saying much about it yet, but an FPGA co-processor is pretty cool and I bet it's in the works.

Altera already has hard core ARM processors in their offerings.

The reason for FPGAs on a data center is to lower energy consumption as well as speeding things up to remain competitive. For example Microsoft was brining them to the datacenter.

http://www.datacenterknowledge.com/archives/2014/06/27/programmable-fpga-chips-coming-to-microsoft-data-centers/

Intel already announced a while back they are working of FPGA fabric for their Xeon processors.

http://www.datacenterknowledge.com/archives/2014/06/19/intel-offer-hyper-scale-operators-ability-reconfigure-cpus-dime/

I can't find the reference for Google, but I did read they were also doing that.

But having the fabric on the chip will be even more efficient than going through an external bus. Altera already has that know-how using their Avalon interconnect.

Not sure what progress has been made, but it's coming.

Edit: I forgot to mention, Altera has been using Intel's tech for their higher end FPGAs for a while now.

Well, it is happening and coming in 2017:
https://www.eevblog.com/forum/microcontrollers/intelaltera-brings-fpga-to-xeon-broadwell-chip/