"Intel already does on-package FPGAs for certain customers, so this would represent the logical extension of that process by bringing the FPGA fully on-die."
Really? I was not aware of this. Anyone know who these customers are?
With their leading Node, and FPGA in place i think Intel has at least secure the future of their DC / Xeon business for another 5 years. Guarding off Potential threats from POWER.
As for ARM i still dont find a it threat to Intel in the server space. The energy and pricing dynamics of server made all the erode all ARM major advantage.
I would have agreed yesterday, but I read a benchmark article on the new HD6200 this morning. Now Intel manage to completely embarrass AMD on the iGPU side of things as well. So their own brand of GPUs, with their advanced manufacturing node, is beginning to sound like an interesting proposal.
Intel cannot be competitive on the discrete GPU market given the very high margins they are used to. They'd have to sell entry level GPUs for 200$ to keep in line with their CPU line.
$200 discrete GPUs will be the only ones around in a year or two. Anything not pushing 150w+ for graphics will be marginally better than iGPUs on-die. The market is changing and GPU prices are already adjusting. Entry-level $200, mid-range $350, flagships $500-650 and halo $999. Remind anyone else of CPU prices? :)
No, IGP is the future. That should be obvious when you look at the track record of IGP vs Discrete. IGP has been growing in leaps and bounds while staying within sane thermal and cost boundaries. Discrete has been prostituting TDP and raping consumer wallets more and more for a smaller and smaller performance advantage. IGP is now competing with 3-4 generation old Discrete, and we're approaching a point where Discrete is going to have to start shipping with water cooling to keep ahead, and radical, costly, new memory technologies to mitigate system bus bottlenecks - by putting 16GB with 1024 bit or wider busses - to compete favourably against IGP. And shortly after that, they will implode.
Would be awesome to have Skylake-E with FPGA... but we all know that with current level of competitions it'll be limited only to few extremely expensive server CPUs.
I hope this move leads to inexpensive reconfigurable processors, especially ones integrated to CPU. That will be one of few ways to improve performance of general applications other than parallelization and process shrinking.
The main problem with reconfigurable processors is that adding a reconfiguration fabric increases latency, which increases pipeline depths, which increases flush costs, which lowers instructions per clock and leads to an overall slower compute performance.
Also, recognise that today's processors are already highly reconfigurable. Between programmable microcode, SIMD, and CISC instruction architecture, there's already tremendous flexibilty in what the core can do in response to a small amount of input code. This is one of the reasons X86 has remained king, and even ARM and MIPS run largely CISC-like code now. But even if you ignore the direction ISA's have taken over the last 2 decades, consider that even Intel - king of the kings - has added specialized GPU cores onto almost all of their current CPU dies. We're already on a road that's littered with heterogenous compute devices. How much more configurable do you really feel our compute devices need to become? LOL
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
20 Comments
Back to Article
Shadow7037932 - Monday, June 1, 2015 - link
"Intel already does on-package FPGAs for certain customers, so this would represent the logical extension of that process by bringing the FPGA fully on-die."Really? I was not aware of this. Anyone know who these customers are?
Ryan Smith - Monday, June 1, 2015 - link
Intel's CPU-FPGA products aren't exactly a secret, but they're not well known, nor have I ever seen specific customers named.https://communities.intel.com/community/itpeernetw...
Kevin G - Monday, June 1, 2015 - link
While they announced this last year, have they actually shipped any Xeon + FPGA products to date?jeffkibuule - Monday, June 1, 2015 - link
Microsoft probably: http://www.enterprisetech.com/2014/09/03/microsoft...Samus - Monday, June 1, 2015 - link
And Apple.phoenix_rizzen - Monday, June 1, 2015 - link
"and alongside rival Xilinx compose the two companies comprise the bulk of the market."There's an extra "compose" in that sentence. :)
Ryan Smith - Monday, June 1, 2015 - link
Two companies, two comprises, no?Thanks for pointing that out. Fixed.
Jambe - Monday, June 1, 2015 - link
"with Intel intending to both continue FPGA development integrating FPGAs into some of their future products."Needs an "and" before integrating, and integrating -> integrate.
iwod - Monday, June 1, 2015 - link
With their leading Node, and FPGA in place i think Intel has at least secure the future of their DC / Xeon business for another 5 years. Guarding off Potential threats from POWER.As for ARM i still dont find a it threat to Intel in the server space. The energy and pricing dynamics of server made all the erode all ARM major advantage.
Heavensrevenge - Monday, June 1, 2015 - link
Honestly... I'd love to see Intel start making discrete GPU's vs just integrated on the CPU die.fluxtatic - Tuesday, June 2, 2015 - link
I think you might be the only person on the planet that feels that way.ShieTar - Tuesday, June 2, 2015 - link
I would have agreed yesterday, but I read a benchmark article on the new HD6200 this morning. Now Intel manage to completely embarrass AMD on the iGPU side of things as well.So their own brand of GPUs, with their advanced manufacturing node, is beginning to sound like an interesting proposal.
benedict - Tuesday, June 2, 2015 - link
Intel cannot be competitive on the discrete GPU market given the very high margins they are used to. They'd have to sell entry level GPUs for 200$ to keep in line with their CPU line.ExarKun333 - Tuesday, June 2, 2015 - link
$200 discrete GPUs will be the only ones around in a year or two. Anything not pushing 150w+ for graphics will be marginally better than iGPUs on-die. The market is changing and GPU prices are already adjusting. Entry-level $200, mid-range $350, flagships $500-650 and halo $999. Remind anyone else of CPU prices? :)iamkyle - Wednesday, June 3, 2015 - link
They already tried this once.Google "Intel i740".
w_barath - Friday, June 5, 2015 - link
No, IGP is the future. That should be obvious when you look at the track record of IGP vs Discrete. IGP has been growing in leaps and bounds while staying within sane thermal and cost boundaries. Discrete has been prostituting TDP and raping consumer wallets more and more for a smaller and smaller performance advantage. IGP is now competing with 3-4 generation old Discrete, and we're approaching a point where Discrete is going to have to start shipping with water cooling to keep ahead, and radical, costly, new memory technologies to mitigate system bus bottlenecks - by putting 16GB with 1024 bit or wider busses - to compete favourably against IGP. And shortly after that, they will implode.Senti - Tuesday, June 2, 2015 - link
Would be awesome to have Skylake-E with FPGA... but we all know that with current level of competitions it'll be limited only to few extremely expensive server CPUs.ZeDestructor - Tuesday, June 2, 2015 - link
So around 4 years delay before I can lay my hands on them... I don't mind...k2_8191 - Tuesday, June 2, 2015 - link
I hope this move leads to inexpensive reconfigurable processors, especially ones integrated to CPU.That will be one of few ways to improve performance of general applications other than parallelization and process shrinking.
w_barath - Friday, June 5, 2015 - link
The main problem with reconfigurable processors is that adding a reconfiguration fabric increases latency, which increases pipeline depths, which increases flush costs, which lowers instructions per clock and leads to an overall slower compute performance.Also, recognise that today's processors are already highly reconfigurable. Between programmable microcode, SIMD, and CISC instruction architecture, there's already tremendous flexibilty in what the core can do in response to a small amount of input code. This is one of the reasons X86 has remained king, and even ARM and MIPS run largely CISC-like code now. But even if you ignore the direction ISA's have taken over the last 2 decades, consider that even Intel - king of the kings - has added specialized GPU cores onto almost all of their current CPU dies. We're already on a road that's littered with heterogenous compute devices. How much more configurable do you really feel our compute devices need to become? LOL