Startup Promises Bandwidth Boost

News & Analysis

SAN FRANCISCO. — A math professor’s observation mushroomed into a startup with two demo chips and three standards efforts. Kandou Bus aims to drive a broad range of chip-to-chip and backplane interconnects down to picojoules/bit rates measured in single digits.

The idea for the company came when a student working on ways to get more bandwidth from a DSL line showed Amin Shokrollahi differential pair traces on a circuit board.

“I asked him what’s a differential pair, and he said it’s a way of using complementary signals across two wires. I thought this was so inefficient,” says Shokrollahi, a professor at the Swiss Polytechnic in Lausanne.

“I told him what I would do, and he said that it was so simple but no one had ever looked at the problem that way. I decided it was time for a company.”

So in 2011 with funds from a Swiss agency, Kandou Bus was formed. The name is the term for beehive in Farsi, Shokrollahi’s native tongue. Like a beehive, the Kandou approach involves getting all the links bundled in an interface to cooperate and thus send significantly more bits than is possible on today’s differential pairs.

Kandou presented a paper and demoed its technology at last week’s International Solid-State Circuits Conference (ISSCC) here. The 40nm demo chip sends 12 Gbit/s per wire at less than 4 picojoules/bit, dispersing eight bits across eight wires. Parts of the chip’s technology could be adopted for use in memory interfaces or on 2.5D chip stacks.

The startup hopes to tape out before October, in a 28nm process, a 42 Gbit/s SerDes chip that delivers 21 Gbit/s per wire across four wires. Using 12 taps of decision-feedback equalization, it will drive signals a meter through a Megtron-6 board and connectors, consuming 9 picojoules/bit or less.

By early next year, Kandou hopes to have its technology fully characterized for multiple applications. It expects to deliver hard macros that boost data rates or cut power by up to a factor of four. It will ask for royalties that may vary from 1 to 4 percent of the device’s average selling price, based on the application.

“The idea of a exploiting something across the communications channel is very interesting, but whether system providers will be able to embrace it with minimal perturbation in their systems is an open question,” said Behzad Razavi, a communications expert and professor of electrical engineering at UCLA, speaking in an interview with EE Times at ISSCC.

Next page: Three standards proposals

More Related Links

Re: Your take

2/19/2014 3:41:19 AM

Thanks Bert. Good points! It is  true that Kandou’s scheme is more a modulation scheme than a coding scheme.

The main issues with modulation in chip-to-chip links are power and time (and area of course). Speeds are much higher than encountered in typical communication settings (up to 10’s of Gbps), and the energy used for the complete delivery of a reliable bit is in the low pico-jule regime. This makes the use of ADC’s quite challenging, even if the effective number of bits is low, and hence modulation schemes like n-QAM or n-PAM are out as soon as n is larger than, say, 4 or so. MLT-3 is also out because of the reduced margin (needs 3-PAM detectors but sends as many bits as differential signaling). Moreover, the noise types encountered in this communication setting are different than those encountered in normal comm systems: thermal noise (Gaussian noise) is very low, but ISI, crosstalk, SSO noise, etc are quite dominating. Hence the modulation scheme has to take these into account. Kandou’s modulation scheme is developed to take all of that into account. In that sense, it is quite different from modulation schemes people normally use.

@Bert: Good point. here’s one comparison, another startup pushing bandwidth increases, this one over networks rather than interfaces though, and using a new modulation scheme:

@DrFPGA: Standards essential patents is a huge and controversial area these days.

Re: Your take

2/18/2014 6:50:34 PM

I guess my only comment was that there are multiple ways of cramming more bits per clock period. MLT-3 is one obvious choice, for example. Or in RF modulation, n-QAM. I’m not disputing that the technique used in this case might be clever, I’m just disputing that we’re talking about a revolutionary idea.

You can also use MLT-3, as one example, either to send more bits per clock cycle, or to reduce the clock rate while not reducing the bit rate. All with appropriate advantage or liability to marginal SNR requirement. Ditto with n-QAM. For a given symbol rate, 64-QAM sends 6 bits per symbol, while PSK only sends one bit per symbol. But look at the marginal SNR requirements of each.

Or for DSL lines. If your house is at the distance limit for a given bit rate, it’s because you’ve reached the limit wrt noise. A technique that sends more bits per clock cycle isn’t going to change matters, unless you simultaneously improve other aspects of the system, such as perhaps the FEC, to more closely approach the Shannon limit. And as far as I know, no one has yet violated Shannon’s limit.

Possibly, this Kandou is more “power efficient” than the competition, in practical encoders and decoders. I wouldn’t know. However if that’s the case, it wasn’t clear reading the article. Meaning, no comparisons were made.

Just a Novice Thought.

2/18/2014 6:36:05 PM

Just looking at the trasceivers (line) side, a diff pair consists of lines A, and A*(=B) to transmit 1 equivalent common mode signal. When adding another CM signal to the transmission, what if I only add a signal C which is a differential signal type complemetary of either A or B depend on the logic value of C. Same approach if I wan to add more such as another signal D. The total number of lines in this case would be 4 diff signals that can reference the others for actual 3 CM signals, instead of 6 generally required when using diff pairs. Not sure how it can be implemented in circuits.

I wonder if what I am thinking here is similar to what AminS describes ?

Re: Your take

2/18/2014 5:29:48 PM

Here is one example of a signaling technique developed by Kandou called “ENRZ”. It uses 4 wires, and the signals that are simultaneously transmitted are either permutaitons of (1,-1/3,-1/3,-1/3) or permutations of the negative of this vector. 8 code-words in total, so 3 bits can be encoded into this codebook. A small digital circuit + a generalization of LVDS across 4 wires jointly drive these signals on the wires. 

One receiver could consist of  three comparators: the first one compares the average of wires 1,3 against the average of wires 2,4. The second comparator compares the average of wires 1,4 against the average of wires 2,3; and the last comparator compares the average of wires 1,2 against the average of wires 3,4. These comparators all reject simultaneous common mode on all 4 wires (but not common mode on adjacent two wires). Moreover, they uniquely determine the codeword sent, provided that the signals have been somewhat equalized before they go into these comparators. 

The interesting thing about this coding is that its resistance to intersymbol interference (ISI) is the same as the resistance of differential signaling to ISI at equal clock rate. But at equal throughput, ENRZ uses a lower clock frequency, hence has the same resistance to ISI as differential signaling at 66% of the frequency. This typically yields much larger margins at same throughput, and can make communication possible in cases where differential signaling would completely run out of steam. 

Re: Your take

2/18/2014 5:18:34 PM

This new coding technique uses the bandwidth more efficiently than differential signaling. The closest way of describing it is that Kandou is using spatial coding, whereas traditional FEC uses temporal coding. That is, Kandou introduces dependencis across the wires of an inteface, whereas channel coding introduces dependencies across time (from one clock cycle to the next). While temporal coding would be possible as well, it comes at the price of higher latency. Spatial coding, however, when properly designed and implemented, has close to no latency. 

Back to the coding part, the goal is to pack more information on the wires than possible with differential signaling, while retaining the properties of differential signaling. The things that make differential signaling robust are common mode resistance (essentially the fact that the receiver rejects noise of equal phase/amplitude on the wires, and the fact that the signals on the wires sum up to zero), there is no simultaneous switching output noise (meaning that current draw from the source does not depend on the particular bits sent), reference-less receivers, and low EMI noise. All these can be captured as mathematical conditions on the codebook used by the communication system, meaning the set of all values that are simultaneously transmitted on the wires. This part requires mathematical analysis. But what is really important and unique about Kandou’s coding techniques is that the concept of an efficient detector is embedded in the definition of the code, and is implemented by a (one-shot) network of generalized comparators — components that are the bread and butter of any SerDes, and can be robustly implemented in any CMOS technology. This puts the design and analysis of the coding system squarely into the mathematical domain without any compromise in implementation and efficiency.

The comparators are such that they reject common mode noise, and are reference-less. The signals put on the wires sum up to zero, and draw the same current regardless of which particular code-word is transmitted. Moreover, the correct design of the codebook reduces EMI noise compared to differential signaling on the same number of wires and equal throughput.

Regarding the Shannon limit: I am not 100% sure what noise types to take into account to compute the mutual information and from that the Shannon capacity, since Shannon’s definitions don’t take into account the efficiency of the implementation, or the power the detector uses. Channels in this environment are largely deterministic, so given enough processing power can be inverted (=fully equalized). This leads to a very large throughput, and the capacity can be calculated easily. However, in practice this is the part that is hardly possible, so the Shannon bounds attained this way are much much better than achievable. 

Re: Your take

2/18/2014 4:46:40 PM

I think the only “differential” involved here is differential signal on two WIRES, not two nets. Here’s the quote:

“I asked him what’s a differential pair, and he said it’s a way of using complementary signals across two wires. I thought this was so inefficient,” says Shokrollahi, a professor at the Swiss Polytechnic in Lausanne.

Differential signals are used to reduce sensititvity to interference, along that link. For example, a noise spike creates a short-term voltage spike on a wire. The same voltage glitch will occur over each wire of the twisted pair. But if the twisted pair is used as a differential pair, where each wire carries the same signal as the other wire, but 180 degrees out of phase, then that “common mode” noise spike will be cancelled. The receiver won’t detect any signal that is in phase, in the twisted pair. The receiver becomes more immune to ambient EMI.

Any number of modulation schemes exist already, that provide greater bandwidth at the expense of higher vulnerability to noise. It’s always a balancing act. So there’s nothing unique in principle, described in the article. The only question should be, how much closer to the Shannon limit is this getting us?

Re: Your take

2/18/2014 4:35:01 PM

SerDes signals sent differentially across two nets will have a communications limit. But, if my understanding of this technology is correct, then one of the limiting factors is that you are only sending differential signals across these nets. You are limiting yourself to an encoding that isn’t necessarily very efficient, yet it is the only one all of us have been using up till now.

Instead of considering it one channel, consider the two nets as two channels that influence each other.  One way of taking advantage of the crosstalk between the two channels is to only send differential signals. But is that the most efficient approach? 

Of course if you send non-differential signals across these nets, you better have a good understanding of the signal integrity effects of the alternately encoded signals, and how pre/post filtering and equalization affects the error rate. It looks like this is the meat of their invention.

Had to laugh at the request that they open source it all. Yeah, and I wish Dell would send me free PCs, and I’ll take a couple of Teslas too, while we’re asking.

Re: Your take

2/18/2014 4:09:31 PM

Anyone have some current data on standards that require royalty payments?

MPEG-2 compression, H.264, ATSC 8T-VSB, for sure. There must be plenty of others too.

Navigate to Related Links Parts Search
185 million searchable parts

(please enter a part number or hit search to begin)

selinz on 7 Presidents Who …