another multi-core architecture or interesting ?

[jlambrecht]

Hello,

http://tilera.com offers impressive chips with a highly optimized version of Linux on it. Warning, the following will sound silly.

As i'm not a full fledged hardware expert i'd like to know what the odds are to run a server or desktop on this kind of chip. Given the fact it has 32-64 cores running on 600-900 Mhz and only consuming ~13 watts (typically). I do realize the obvious answer is no but 32*600 = 19200 Mhz with 3 instructions per cycle. That should do to run a common desktop, not ? Even with all kinds of obvious restrictions in place.

Regards,

J.

http://tilera.com/products/TILEPro64.php is the most powerfull one.

[987687]

You can't just multiply like that. That chip would probably not be ideal for a desktop because most applications are not multi-threaded well enough to use all that power.
Where it would be good is on, say, a production server that is compiling things with gcc. gcc is very well multi-threaded and would take advantage of the power very well.

[jlambrecht]

Indeed. Also java server or any other server relying on multi-threading like heavy loaded web-servers ? Multi-Media workstations would do as well i figure.

My question popped when i read this chip is also geared toward more generic applications so not only signal-processing. Or for where it's used, network sniffing etc.

price said to be around 450$ in quantities of 10.000

[plugwash]

I don't think gcc itself is multithreaded. You can use make to run multiple gcc runs in paralell ( -j switch ) but that doesn't work with all makefiles and can also eat up a lot of ram ( especially if it's g++ rather than plain gcc that is involved. I've seen g++ get well over a gig of ram usage)

Massively parallel processing is great for some applications but often you are better off with a few fast cores that have plenty of cache.

[jlambrecht]

I don't think gcc itself is multithreaded. You can use make to run multiple gcc runs in paralell ( -j switch ) but that doesn't work with all makefiles and can also eat up a lot of ram ( especially if it's g++ rather than plain gcc that is involved. I've seen g++ get well over a gig of ram usage)

Massively parallel processing is great for some applications but often you are better off with a few fast cores that have plenty of cache.

AFAIK tilera provides a custom linux distro for their chips, though i don't know what price (if any) they're asking.

http://tilera.com/products/software.php

Tilera provides several operating environments to accommodate the different application needs.

Standard SMP Linux Environment
• Standard operating environment with a full SMP Linux OS, a hypervisor, and application libraries
• For applications and control plane code requiring operating system services

Zero Overhead Linux (ZOL)
• Similar to standard SMP environment
• Avoid Linux overhead: OS interrupts, timer ticks, TLB shoot downs, etc.
• Transparent to programmer - no software change required
• For high performance real-time dataplane applications

Bare Metal Environment
• Full control of the hardware
• Rich set of libraries for memory, communication, and I/O
• For embedded applications requiring fine grain control of memory and I/O
• For customers who want to port their own OS

Hybrid Environment
• Using two or all three of the above models
• Each environment can be run on one or more tiles
• Ideal for customers aggregating dataplane and control plane

[jlambrecht]

YEAH.

it's good to be ahead ... this IS going to prove rather nifty cornerstone for a next-gen desktops

http://www.linuxfordevices.com/c/a/News/Tilera-TileGx/

http://www.theregister.co.uk/2009/10/26 ... esh_chips/

[BioTube]

As an expansion card, maybe, but the PC architecture is designed with x86 in mind.

[jlambrecht]

As an expansion card, maybe, but the PC architecture is designed with x86 in mind.

i'm by no means qualified to understand what this CPU *REALLY* can do , but to my understanding this is NOT the same as so many other dsp-centric cpu designs. This is novel and flexible, the newest releases introducing more feature that allow to mix generic and signal processing.

I hope someone would take a serious look at post a good read on this cpu. Realising there's no holy grail for cheap but this one does seem to be very flexible with regards to how it could be implemented and used.

[aspnair]

number of cores could increase software lockout also.
http://en.wikipedia.org/wiki/Software_lockout
http://web.mit.edu/smadnick/www/papers/P001.pdf

[jlambrecht]

Feature Enables

Massively Scalable Performance

• Array of 16 to 100 general-purpose processor cores (tiles)
• 64-bit VLIW processors with 64-bit instruction bundle
• 3-deep pipeline with up to 3 instructions per cycle
• 32K L1i cache, 32K L1d cache, 256K L2 cache per tile
• Up to 750 billion operations per second (BOPS)
• Up to 200 Tbps of on-chip mesh interconnect
• Over 500 Gbps memory bandwidth with four 64-bit DDR3 controllers


• 40 - 80 Gbps Snort® processing
• 40 - 80 Gbps nProbe
• H.264 HD video encode: dozens of streams of 1080p (baseline profile)
• 64+ channels of OFDM baseband receiver processing (wi

[BioTube]

i'm by no means qualified to understand what this CPU *REALLY* can do , but to my understanding this is NOT the same as so many other dsp-centric cpu designs. This is novel and flexible, the newest releases introducing more feature that allow to mix generic and signal processing.

I hope someone would take a serious look at post a good read on this cpu. Realising there's no holy grail for cheap but this one does seem to be very flexible with regards to how it could be implemented and used.

You called it a "rather nifty cornerstone for a next-gen desktops[sic]"; my point was merely that the computer architecture associated with desktops was designed with the x86 CPU ISA in mind and might not be workable with processors that don't use it.

[jlambrecht]

i'm by no means qualified to understand what this CPU *REALLY* can do , but to my understanding this is NOT the same as so many other dsp-centric cpu designs. This is novel and flexible, the newest releases introducing more feature that allow to mix generic and signal processing.

I hope someone would take a serious look at post a good read on this cpu. Realising there's no holy grail for cheap but this one does seem to be very flexible with regards to how it could be implemented and used.

You called it a "rather nifty cornerstone for a next-gen desktops[sic]"; my point was merely that the computer architecture associated with desktops was designed with the x86 CPU ISA in mind and might not be workable with processors that don't use it.

Hmmm ... curious remark(s).

So given the fact there's Linux running on this cpu allready and it's supporting PCIe did not make an impression

I still can't get my head around the statement 16 x 1.25Ghz of generic computing power will result in unuseable desktop machine. Most obvious reason will be pricing in I figure .. though that would count less with the 16-core version ... x86 competitors will IMHO not be able to ramp up the number of cores that fast, for a while at least.

[jlambrecht]

This will not come cheap but hey ... who needs packet processing accelerators on a desktop anyway

In addition to supporting a greater number of cores, the Tile-Gx is said to offer several major enhancements compared to the TilePro, including:

* Next-generation 64-bit core -- A new three-issue 64-bit core offers a full virtual memory system. Each core includes 32KB L1 I-cache, 32KB L1 D-cache, and 256KB L2 cache, with up to 26MB total L3 coherent cache available across the device, says Tilera.

* Enhanced SIMD instruction extensions -- A 4-MAC/cycle multiplier unit delivers up to 600 billion MACs per second signal processing performance, claimed to be more than 12 times the fastest commercial DSP.

* DDR3 memory controllers -- Two or four 72-bit controllers are available, running at up to 2133MHz speeds with ECC support. The chips support up to 1TB total DDR3 capacity, as well as "powerful memory striping modes for maximum utilization," says the company.

* Hardware acceleration engines -- An on-chip MiCA (Multistream iMesh Crypto Accelerator) system is touted as delivering up to 40Gbps encryption and 20Gbps full duplex compression processing, "tightly coupled to the iMesh for extremely low latency and wire-speed small packet throughput." Meanwhile, a true random number generator (RNG) and public key accelerator are said to enable up to 50,000 RSA handshakes per second.

* mPIPE packet processing accelerator -- The mPIPE (multicore Programmable Intelligent Packet Engine) system offers wire-speed packet classification, load balancing, and buffer management functions, says Tilera. The C-programmable engine is said to delivers 80Gbps and 120 million packets-per-second of throughput for packets with multiple layers of encapsulation.

Peripheral support is said to include up to eight 10GbE XAUI interfaces, two Interlaken interfaces, and three "Gen2" PCI-Express interfaces, each selectable as endpoint or root complex. The Tile-Gx processors are also said to offer up to 32 gigabit Ethernet MAC interfaces. The two higher-end versions come in 34 x 45mm BGA packages, while the two lower-end models ship with 35 x 35mm BGA packages.

[BioTube]

Hmmm ... curious remark(s).

So given the fact there's Linux running on this cpu allready and it's supporting PCIe did not make an impression

I still can't get my head around the statement 16 x 1.25Ghz of generic computing power will result in unuseable desktop machine. Most obvious reason will be pricing in I figure .. though that would count less with the 16-core version ... x86 competitors will IMHO not be able to ramp up the number of cores that fast, for a while at least.

Linux can run on all sorts of configurations, many of which have no relation to the IBM PC architecture. What I was saying was that the large number of standards in such a computer were made with x86 in mind and might not be compatible with other processors.

[jlambrecht]

Hmmm ... curious remark(s).

So given the fact there's Linux running on this cpu allready and it's supporting PCIe did not make an impression

I still can't get my head around the statement 16 x 1.25Ghz of generic computing power will result in unuseable desktop machine. Most obvious reason will be pricing in I figure .. though that would count less with the 16-core version ... x86 competitors will IMHO not be able to ramp up the number of cores that fast, for a while at least.

Linux can run on all sorts of configurations, many of which have no relation to the IBM PC architecture. What I was saying was that the large number of standards in such a computer were made with x86 in mind and might not be compatible with other processors.

uhm ... well ... okay then

[jlambrecht]

Hi,

Interesting response as it actually contains an example of how to put this to good use. At the time i wrote this article the sum of all cores in what seemed to be the standard-chip was aroun 19Ghz, i pressume one could get decent desktop performance on that system regardless of the unoptimized nature for this processor.

http://www.tomshardware.co.uk/forum/236 ... ware-linux here's an old list of multi-threaded applications ...

This processor actually has it's own Linux distro with compilersupport in place for it's design ... they've been running a recompiled but othwerise unmodified snort on this monster with great success.

Did you check http://tilera.com/development_tools ?

Personally i kind of hope these cpu's wil end up on a motherboard from a major vendor and (kind of) become an alternative to x86 based systems. As many malware will not have any chance of passing thru anymore. Unless they'd be recompiled for this platform, haha.

Greetings,

J.

[jlambrecht]

This cpu is now used by Facebook, Google etc. See also http://www.tilera.com/about_tilera/in_the_news

Great stuff for a cross-compiler set-up. Shame it comes bundled with CentOS.