News & Analysis
Comment
peter.clarke
Ultimately surely the general purpose model has to be similar to that used by ...
Peleg
Plurality's HyperCore developed a many-core design not only for wireless ...
European startup attempts many-core revolution
Peter Clarke
1/24/2012 7:38 AM EST
Three years and $27 million
The company raised $19.1 million in 2008 and 2009. The company announced it had raised a Series B of $7.8 million in 2011. The company has 50 engineers and is backed by several French investment funds including: ACE Management, Inocap, Eurekap, CEA Investissement, Rhône Alpes Creation, Promelys, local funds and other private investors as well as by the French government agency OSEO, which provides support to startups.
Along with the chip, Kalray is offering development boards and the AccessCore software development environment, the website indicates.
AccessCore supports a C-based programming model but support also varies between different user profiles from Linux support for legacy functions up to a dataflow environment that can maximize use of the array. Standard GCC & GDB technologies are used for compilation and debug.
"Kalray's technology has been developed with many OEM partners across several vertical markets, as well as partnering with third-party software vendors," said Monnier, in a statement issued May 2011. "Our first processor achieves a global processing power of 500 billion operations per second, along with a much lower power consumption than competitive solutions. Embedded designers will get the benefit of high processing power, low power consumption and high level programming to develop innovative applications in the fields of image processing, signal processing, control, communications and data security. The access cost of MPPA processors makes them optimum for all low to medium volume applications."
The 256-processor chip allows array extension by the clustering of several chips and the I/O offering includes generations 2 and 3 of PCI Express, Ethernet able to run at up to 40-Gbits per second and general purpose I/Os. Memory controllers for flash and DDR DRAM allow for external storage of up to 128-Gbytes. And is described as being in a standard 40-mm x 40-mm BGA package.
And, as one might expect, the 256-processor chip is the first in a projected family. The MPPA family scales from 16 to 64 clusters per chip. The 1024-core version delivers up to 2-Tops [tera operations per second]. The MPPA also includes 4 to 8 Interlaken interfaces for multi-MPPA chip systems and connection to FPGAs.
Related links and articles:
www.kalray.eu
News articles:
Defense firm offers many-core runtime support
Many-core Plurality to benchmark on EEMBC
Ten quotes on parallel programming
ARM's 'big-little' strategy and the many-core puzzle
ARM, University of Michigan target many-core R&D
The company raised $19.1 million in 2008 and 2009. The company announced it had raised a Series B of $7.8 million in 2011. The company has 50 engineers and is backed by several French investment funds including: ACE Management, Inocap, Eurekap, CEA Investissement, Rhône Alpes Creation, Promelys, local funds and other private investors as well as by the French government agency OSEO, which provides support to startups.
Along with the chip, Kalray is offering development boards and the AccessCore software development environment, the website indicates.
AccessCore supports a C-based programming model but support also varies between different user profiles from Linux support for legacy functions up to a dataflow environment that can maximize use of the array. Standard GCC & GDB technologies are used for compilation and debug.
"Kalray's technology has been developed with many OEM partners across several vertical markets, as well as partnering with third-party software vendors," said Monnier, in a statement issued May 2011. "Our first processor achieves a global processing power of 500 billion operations per second, along with a much lower power consumption than competitive solutions. Embedded designers will get the benefit of high processing power, low power consumption and high level programming to develop innovative applications in the fields of image processing, signal processing, control, communications and data security. The access cost of MPPA processors makes them optimum for all low to medium volume applications."
The 256-processor chip allows array extension by the clustering of several chips and the I/O offering includes generations 2 and 3 of PCI Express, Ethernet able to run at up to 40-Gbits per second and general purpose I/Os. Memory controllers for flash and DDR DRAM allow for external storage of up to 128-Gbytes. And is described as being in a standard 40-mm x 40-mm BGA package.
And, as one might expect, the 256-processor chip is the first in a projected family. The MPPA family scales from 16 to 64 clusters per chip. The 1024-core version delivers up to 2-Tops [tera operations per second]. The MPPA also includes 4 to 8 Interlaken interfaces for multi-MPPA chip systems and connection to FPGAs.
Related links and articles:
www.kalray.eu
News articles:
Defense firm offers many-core runtime support
Many-core Plurality to benchmark on EEMBC
Ten quotes on parallel programming
ARM's 'big-little' strategy and the many-core puzzle
ARM, University of Michigan target many-core R&D
|
|
|
|
|
Navigate to related information
Jayson Five
1/24/2012 1:30 PM EST
A "general purpose" 5 watt 256 VLIW flow processor for "embedded applications" in the fields of "image processing, signal processing, control (??), communications and data security"?
Doesn't sound too general purpose to me. Am I missing something here?
In the least, the application needs to be amenable to flow processing. That's why there aren't any "general purpose" many-core flow-processors out there.
So what's the angle here?
Sign in to Reply
david.may
1/24/2012 4:14 PM EST
"startup attempts many-core revolution"
i have to agree jayson, infact its already been done before a long time ago now April 4, 2006 in the so called Rapport Inc's "kilocore"
http://en.wikipedia.org/wiki/Kilocore
"Kilocore, from Rapport Inc. and IBM, is a high-performance, low-power multi-core microprocessor that has 1,025 cores. It contains a single PowerPC processing core, and 1,024 eight-bit Processing Elements running at 125 MHz each, which can be dynamically reconfigured, connected by a shared interconnect. It allows high performance parallel processing."
and OC lets not forget to give other credit wher eits due too, that being dave may (not related :) of XMOS many multicore fame.....
sure this France pretenders version is using more modern process and adds current busses etc but lets not forget the past and still living broard shoulders Joel Monnier is standing on...
Sign in to Reply
Jayson Five
1/25/2012 12:55 AM EST
Riddle: What do you call a beast configured like a kilocore only it's a MEGAcore?
A single Power PC processing core hooked up to an FPGA.
Which requires the hapless software programmer to also be a logic designer. Which degenerates to a situation where you have no software traction, and therefore no apps, and therefore no customers, and therefore no real product. Just a really cool idea that doesn't quite fly.
Sign in to Reply
david.may
1/24/2012 4:57 PM EST
on a side note i really think its the right time for "kilocore" Michael O'Brien, tranputer/XMOS pioneer dave may, and ARM Inc to get together with some VC and make a mass produced ARM A15/ Mali T600 based Achronix FPGA on the 22nm or lower Intel Process ASAP ;)
Sign in to Reply
entity279
1/25/2012 5:56 AM EST
Perhaps the difference between this processor and others before it comes down to peformance.
The promised 200 GFLOPS at 5W is quite an achievement, I would say.
Sign in to Reply
adapteva
1/25/2012 11:48 AM EST
Oh,the lure of programmable massively parallel architectures... Here are the players I know that have tried similar things. Let me know if I missed someone.
Adapteva, Ambric, AMD,Asocs, Cavium, Cradle, Clearspeed, Chameleon, Coherent Logix, CPUtech, ElementCXI, EZChip, Freescale, Inmos, IPFlex, Greenararys,Intel, Kalray, Mathstar, Morpho, Monarch, Netlogic, Nvidia, Octasic, Picochip, Plurality, Panave, PACT, Recore, Rapport, Sandbridge, Silicon Hive, Sound Design, Stretch, Stream Processors, ST, Tilera, Transputer, SiCortex, SpiralGateway, Venray, XMOS
Sign in to Reply
dkp
1/25/2012 12:20 PM EST
That's a pretty good list!
Sign in to Reply
Athlor
1/25/2012 7:34 PM EST
Way to go Adapteva! I could only remember a half a dozen of them but you hit it out of the park, lol.
Still, 200 Gflops at 5W, if to be believed, is alluring.
Sign in to Reply
peter.clarke
1/26/2012 5:53 AM EST
It is always about keeping the processors busy with USEFUL work. And that is problem of software, programming, distributed memory and I/O bandwidth.
Sign in to Reply
peter.clarke
1/26/2012 5:51 AM EST
Good list...I think you got most of them.
Although the lines get blurry as you go to more application-specific processors.
Sign in to Reply
peter.clarke
1/26/2012 5:52 AM EST
You missed Siroyan of Reading UK ....which just came back in to my head.
Sign in to Reply
DrQuine
1/26/2012 2:22 PM EST
Don't forget Martin Marietta with their Geometric Arithmetic Parallel Processor (GAPP) in 1988. A large circuit board held 32,000 processors as I recall. It supported a Single Instruction Multiple Processor (SIMP) approach. Great for working on images after the initial overhead of clocking in the data from an edge.
Sign in to Reply
david.may
1/25/2012 3:02 PM EST
well , Inmos,Transputer, and XMOS are in effect the same company/person at least pioneer dave may :) he does some nice FPGA files but then he has been at it longest...
Sign in to Reply
Peleg
9/12/2012 3:23 AM EDT
Plurality's HyperCore developed a many-core design not only for wireless inftastructure.
Although not successful yet in producing a maketable product, it proposed a holistic solution for programming model and hardware that programmers liked and silicon could handle.
I believe message passing for massively parallel machines is not usefull. Here is another one who tries.
Sign in to Reply
peter.clarke
9/12/2012 6:36 AM EDT
Ultimately surely the general purpose model has to be similar to that used by multiple computing nodes hanging off the Internet, but writ small on a single die.
For now the likes of Intel and ARM want to keep everything coherent and synchronized but my instinct tells me that cant work as you go to scores of core except in certain very tightly controlled applications.
Sign in to Reply