Has any one herd of The TILE-Gx CPUs?

Has any one else herd of The TILE-Gx CPUs?

I just came across it when I was reading about the new Linux 2.6.36 kernel (see: http://www.linuxfordevices.com/c/a/News/Linux-2636-kernel-released/ ). Apparently they have added native support for the Tilera chip architecture to the kernel.

Here is the manufacture website: http://www.tilera.com/products/processors/TILE-Gx_Family

Any one got any performance benches,etc? Are there any other stand alone devices except for the 1U server that they have ( http://www.tilera.com/products/platforms/TILEmpower_platform )?

A few things I notices:
1. They already have Quad Channel DDR3 support!
2. 16-100 cores/tiles. If the apps are extremely multi threaded this may be a better choice over a i7
3. Linux supports it now (native) and GCC works on it. Has Eclipse IDE support.
4. 10-55W @ 1-1.5Ghz


What do you think? Do you think this will replace current Intel,TI,etc SoC used in routers,etc?
 
Solution


It's not very impressive compared to a GPU or to the server platforms used in high-performance computing setups. Consumers aren't the target of this chip at the moment, HPC users and such apparently are. A high-end GPU has 256-bit (quad-channel) GDDR5 running in the neighborhood of about 4000 MHz effective, which is nearly twice the bandwidth of the quad-channel DDR3-2133 Tilera uses in their top model. Both Intel and AMD have quad-channel memory controllers in their server CPUs already, and can link together multiple CPUs in NUMA to...


I've heard of them before. Their products seem somewhere in between a normal desktop/server CPU and a GPU or FPGA in structure, function, and capabilities. I don't know much about the actual ISA of their cores, other than they are 64 bits wide. I'd expect these to work well for things that are highly multithreaded but need a large memory capacity and a bit more general-purpose function than a GPU can deliver. Applications that are embarassingly parallel and consist of pretty much all FPU/SIMD work will still run better on GPUs as GPUs have even more memory bandwidth and FLOPS than the Tilera chips. Applications that need more clock speed, cache size, or memory bandwidth will still perform better on normal server CPUs. The 100-core model has four channels of DDR3-2133, which is pretty fast. However, two Opteron 6100s have even more with their 8 channels of DDR3-1333, and that is only shared with 16-24 cores.

I think that this may have some promise as an application accelerator to use in place of GPUs where GPUs may not quite have enough general-purpose functions to effectively do the task at hand. It probably won't replace too many SoCs unless Tilera scales this way down to a small handful of cores as many SoCs are pretty basic units. Maybe it would make a dent in some of the biggest SoCs like those in big-iron routers, but that remains to be seen. The fact that Linux will run on it is definitely a plus for that case.
 
Applications that are embarassingly parallel and consist of pretty much all FPU/SIMD work will still run better on GPUs as GPUs have even more memory bandwidth and FLOPS than the Tilera chips.
True, but these can actually support an entire OS on top of it. And these have much lower power requirement.

Maybe it would make a dent in some of the biggest SoCs like those in big-iron routers, but that remains to be seen. The fact that Linux will run on it is definitely a plus for that case.
Exactly. That's what I was thinking. I think, Broadcom is one of their VCs/Investors.

As for the DDR3 Channels, it's quite impressive considering the consumer line up still doesn't use it yet. AFAIK, Intel/AMD won't use Quad Channel on the consumer front until LGA2011 and Bulldozer.
 


It's not very impressive compared to a GPU or to the server platforms used in high-performance computing setups. Consumers aren't the target of this chip at the moment, HPC users and such apparently are. A high-end GPU has 256-bit (quad-channel) GDDR5 running in the neighborhood of about 4000 MHz effective, which is nearly twice the bandwidth of the quad-channel DDR3-2133 Tilera uses in their top model. Both Intel and AMD have quad-channel memory controllers in their server CPUs already, and can link together multiple CPUs in NUMA to get staggering bandwidth numbers. Consumer CPUs don't have quad-channel memory controllers because the existing three-channel setups are more than enough even at low RAM speeds, as a recent test here showed. Little besides synthetic RAM benchmarks show much of a difference in LGA1366 Core i7s running with two or three channels of memory as xtreview found, so we're not going to see four channels anytime soon unless there's an on-package GPU also sitting behind those memory controllers.
 
Solution