Router SoC 101

Status
Not open for further replies.
G

Guest

Guest
Do you ever wonder what goes on inside a router? In this article, we take a close look at the SoCs inside of them, which help us manage and maintain our connected lives.

Router SoC 101 : Read more
 

bwhiten

Distinguished
Sep 6, 2011
48
0
18,530
Uhhhh...Those first pictures are not "schematics". They are CAD renderings of the box and main board at best, but definitely not schematics.
 

EdJulio

Reputable
Jul 7, 2015
10
0
4,510
Uhhhh...Those first pictures are not "schematics". They are CAD renderings of the box and main board at best, but definitely not schematics.

Thanks, bwhiten. Updated the caption...
 

bit_user

Polypheme
Ambassador
Nice article!

Small, irrelevant fact: MIPS was once owned by SGI and used in their servers and workstations. They even used a MIPS CPU in the N64, which they designed for Nintendo. In fact, that was largely the outcome of a previous (if not the first) wave of VR hype. But, I digress...

Also, most people consider ARM to be RISC. Or, at least as much as anything is, these days. Indeed, the name once stood for Advanced RISC Machines.

But I didn't know what MIPS originally stood for, so thanks for that. I wonder whether or how long that remained true of their architectures.
 

EdJulio

Reputable
Jul 7, 2015
10
0
4,510


Thanks! I'll share this with Gene! Cheers!!!
 

Gabriel_1965

Commendable
Feb 27, 2016
2
0
1,510
Question: I've seen a router with 72 cores would that be made to be a 72 core pic and I could use the cores for multi ore computing?
 

GeneFabron

Commendable
Feb 23, 2016
2
0
1,510
Question: I've seen a router with 72 cores would that be made to be a 72 core pic and I could use the cores for multi ore computing?

Networking chips are generally very specialized - they handle certain tasks like I/O and data transport/filtering really, really well. Physics simulations...not so much. So it would really depend on what kind of calculations you were carrying out, how well you can parallelize the operations, and the base architecture you were implementing. If you could *translate* your model problem into something that can be expressed as network I/O data transfer (I can think of a couple of very specialized machine learning algorithms whose math can be expressed this way), then you could take advantage of this kind of system. Either way, I expect you'd have to do the math and all coding from scratch. Or you could brute-force it by (heavily) modifying pre-existing code for Beowulf clusters and the like...

In most cases, you will get much, much better cost:efficiency ratios with a more traditional setup, or an actual cluster.

If this is the TILE-Gx72 Processor you are talking about...it can support up to 1Tb of memory...an interesting problem.

tl;dr yes, you can do anything, but it may not be the most efficient, most cost-effective, or "best" method based on the type of computational problem you are trying to solve.
 

bit_user

Polypheme
Ambassador
No, but it once did. I worded my statement, carefully.

when the company was incorporated in 1990, the acronym was changed to "Advanced RISC Machines", in light of the company's name "Advanced RISC Machines Ltd." At the time of the IPO in 1998, the company name was changed to "ARM Holdings"
 
Status
Not open for further replies.