Breaking News
News & Analysis

MIT Puts 36-Core Internet on a Chip

Uses twin meshes to scale to any number of cores
7/8/2014 07:25 AM EDT
3 comments
NO RATINGS
More Related Links
View Comments: Threaded | Newest First | Oldest First
EmbeddedSteve718
User Rank
Author
Nothing really new. . .
EmbeddedSteve718   7/8/2014 10:42:24 AM
NO RATINGS
Just an updated variation of a theme:

http://en.wikipedia.org/wiki/Network_on_a_chip

3 companies listed here:  They've been around for years. . . .

ARM's Latest:  http://www.arm.com/products/system-ip/interconnect/corelink-ccn-508.php

IBM's been doing this for years:  http://en.wikipedia.org/wiki/Blue_Gene

Anyone know of Tilera (http://www.tilera.com) recently purchased by EZChip?  Another multi-core MIT spin-off.

HW is good, but being able to exploit this "correctly" requires rather smart Software that doesn't exist.  Tilera and other multi-core folks have been playing with this for years, still not there. . .  Even today, with ARM interconnects, SW can only do so much to exploit the HW.  Until HW and SW get together (i.e. make a hole in the fence) - we'll never fully utilize all the CPU power we have.  We'll just continue to waste it with more cores running crappy software. . . . 

As the Saying Goes:  "It's the software, stuipd". . . .

R_Colin_Johnson
User Rank
Author
Re: Nothing really new. . .
R_Colin_Johnson   7/8/2014 11:14:42 AM
NO RATINGS
Your'e right if they fail, but just like the other guys , these guys think they have the software right this time. If they do have siomething new and validate their software through testing, then they will release the Verilog hardware description, so at least they are doing things in the right order!

markhahn0
User Rank
Author
Re: Nothing really new. . .
markhahn0   7/13/2014 2:07:53 AM
NO RATINGS
It's funny that people always whinge about how software is the problem, that sw isn't taking advantage of the hw.  I'd like to propose the opposite: very little sw needs highly parallel hw.  As a person who works with highly parallel systems every day, I always cringe when someone asks why their desktop system doesn't run 4x faster on a 4-core system.  The answer is mainly "because you don't need it to".  Or a little less sarcastically: you have tiny sections of workload that may be inherently parallel enought to take advantage of highly parallel hardware, but most of your time is spent waiting on the user, or IO or memory.

There are plenty of workloads that can effectively utilize a 36-core chip today, AND DO SO.  Heck, 36c isn't even a bit system these days - it's doable in a simple commodity dual-socket server.  Not to mention Phi or GPUs.

Most Recent Comments
michigan0
 
SteveHarris0
 
realjjj
 
SteveHarris0
 
SteveHarris0
 
VicVat
 
Les_Slater
 
SSDWEM
 
witeken
Most Recent Messages
9/25/2016
4:48:30 PM
michigan0 Sang Kim First, 28nm bulk is in volume manufacturing for several years by the major semiconductor companies but not 28nm FDSOI today yet. Why not? Simply because unlike 28nm bulk the LDD(Lightly Doped Drain) to minimize hot carrier generation can't be implemented in 28nm FDSOI. Furthermore, hot carrier reliability becomes worse with scaling, That is the major reason why 28nm FDSOI is not manufacturable today and will not be. Second, how can you suppress the leakage currents from such ultra short 7nm due to the short channel effects? How thin SOI thickness is required to prevent punch-through of un-dopped 7nm FDSOI? Possibly less than 4nm. Depositing such an ultra thin film less then 4nm filum uniformly and reliably over 12" wafers at the manufacturing line is extremely difficult or not even manufacturable. If not manufacturable, the 7nm FDSOI debate is over!Third, what happens when hot carriers are generated near the drain at normal operation of 7nm FDSOI? Electrons go to the positively biased drain with no harm but where the holes to go? The holes can't go to the substrate because of the thin BOX layer. Some holes may become trapped at the BOX layer causing Vt shift. However, the vast majority of holes drift through the the un-dopped SOI channel toward the N+Source,...

Datasheets.com Parts Search

185 million searchable parts
(please enter a part number or hit search to begin)
Like Us on Facebook
EE Times on Twitter
EE Times Twitter Feed