Marvell Semiconductor had a less than perfect Xmas last year, having had a $1.17 billion judgement against them for infringement of two patents held by Carnegie Mellon University, after a jury trial in Pittsburg.
The award was in relation to two patents that referred to complex algorithms used to convert signals into data, for example when reading the data on hard disc drives. When reading such data, the computer sees a stream of data - but which of this signal is the underlying information and which is noise?
To answer this question, the Italian born American engineer Andrew Viterbi developed the Viterbi algorithm, which is a means of determining the most likely explanation, i.e. the underlying information, that underpins the signal. And then went on to receive a national medal of science for developing this algorithm and co-found Qualcomm, among other companies.
The Viterbi algorithm works in part by forming a structure with branches. Initially it was assumed that the noise in the different branches was independent, but in some circumstances this noise is correlated and this can introduce errors in the final outputs. The two Carnegie Mellon patents claim a means of compensating for this correlated noise.
Claim 4 of US patent 6201839, (Filing year 1998, priority date May 1997, and which we will 'CM1' because it is the earlier patent ) covering:
A method of determining branch metric values for branches of a trellis for a Viterbi-like detector, comprising:
selecting a branch metric function for each of the branches at a certain time index from a set of signal-dependent branch metric functions; and
applying each of said selected functions to a plurality of signal samples to determine the metric value corresponding to the branch for which the applied branch metric function was selected, wherein each sample corresponds to a different sampling time instant.
Claim 2 of US patent 6438180 (Filing year 1999, same priority date as CM1, hereafter 'CM2'), covers (when combined with its dependent claim 1)
A method of determining branch metric values in a detector, comprising:
receiving a plurality of time variant signal samples, the signal samples having one of signal-dependent noise, correlated noise, and both signal dependent and correlated noise associated therewith;
selecting a branch metric function at a certain time index; and
applying the selected function to the signal samples to determine the metric values.
wherein the branch metric function is selected from a set of signal-dependent branch metric functions.
In response, Marvel claimed that these patents were anticipated by US patent 6282251 to Seagate, which also discloses the use of a Viterbi algorithm to read the data on a spinning hard disc drive, and where the noise on the different branches in the multi path is correlated, and not random. However Seagate, a competitor to Marvell, was apparently happy to support Carnegie Mellon in this dispute, providing an email from the inventor of it patent noting that the CMU invention went well beyond the current state of the art.
Whether the Seagate patent is valid prior art for two Canargie Mellon patents is a matter for the appeal for this case - and there is some commentary on the web which expresses strong opinions on this. However, we were intrigued to see what AmberScope patent searching could add to this case, starting with the links between Carnegie Mellon patents and Marvell.
Citation links to Marvell Semiconductor
Carnegie would not had to look far for suggestions that Marvell may need to take a license to their two patents. As a simple example of this, the website Patentbuddy offers the ability to identify and rank companies that have forward citation links to their patents. For the CM1 patent, the leading owner of forward citations was Marvell with 13 forward citations, while CM2 had 24 forward citations from Marvell. Forward citation by themselves are not proof of infringement, but does suggest that the citing company is interested in and has invested in developing similar technology.
It is also worthwhile considering how the Marvel patents are connected to the Carnegie patents. The figure below shows an AmberScope network of the patents connected to the CM2 patent. Patents belonging to Marvell have been marked with green circles using the word 'filter'.
As you can see, the patents connected by (mostly forward citations lines, which are green as opposed to the blue backward citation connections) to the CM2 patent fall into three main clusters. A cluster means that the patents have interconnections between them, as well as to the CM2 focus patent.
You can also see the Marvel patents are distributed mainly in the bottom right cluster, although some patents are found in the cluster on the left. This cluster of Marvel patents would suggest a number of patents filed for similar inventions by Marvel, which in turn suggests a technology investment by Marvel in this technology area. So an investigator in this area would be well advised to review this patents to see if they described a technology, which if being commercialised by the patent owner, may infringe other patents in this area.
Prior art search - the easy way?
Given the natural interest in this billion dollar duo of patents, we were intrigued to see what the citation data could tell us if there were other patents out there that:
- Discussed the treatment of correlated noise in a Viterbi algorithm using a time delay
- Had filing dates prior to 1998
- Had not been previously identified as prior art for either of the Carnegie Mellon patents by the USPTO.
We have done this search in three ways:
1) Using line thicknesses within AmberScope to suggest similar inventions - and 'exploring' onwards from there
The complexity of inventions such as this one puts a big onus on searchers, who may not be fully familiar with the underling technology. Which raises an interesting question - what if AmberScope could analyse the citation information to lead searchers to a potentially relevant, purely in a objective way? Which may never replace the hard and necessary work of drawing final subjective conclusions from reviewing a patent, but could suggest which patents most deserve further analysis.
In fact this is possible by using line thicknesses in Amberscope networks. Consider the figure below, which shows the network associated with CM1. This network shows forward citations, backward citations, and ghost patents, which are indirectly connected patents that may be similar to the focus patent.
If we zoom into the area around the focus patent CM1 (which is easy to do using a mouse wheel), it can be seen that the lines joining the patents are of different colours and thicknesses. The colours refer to the citation direction with respect to the patent being reviewed, in this case CM1. Green lines point to forward citations, and blue lines to backward citations.
The line thickness refers to the predicted similarity of the two patents, which is calculated from other citation relationships in the wider network. In this, the thickest line is to the other Carnegie Mellon patent, CM2. This should not be surprising, as CM2 is a continuation in part patent of CM1.
But how can this be used to find prior art patents? CM1 has a filing date of 1998, and so we will filter out all all patents later than 1998 using the filing year circled below. Only a few patents are left, and all are backward citations. One line is slightly thicker than others, see below.
This patent is US5689532 (filed 1996), filed by Quantum Corporation and now owned by Seagate. This patent does not specifically mention correlation of white noise and does not appear to be a direct disclosure of CM1 - this is not surprising, as being a known backward citation for CM1, CM1 may not have been granted if the Quantum patent was a direct disclosure.
However the value of the Quantum patent is it too has connections, in this case to some ghost patents to the CM1 patent. The strongest connection (by line thickness) is to US5949831 (filed 1997) filed by IBM and now owned by HGST Netherlands, a subsidiary of Western Digital.
This discloses a similar technology, namely that "Colored (non-random) noise at the input of a PRML Viterbi sequence detector results in sub-optimal performance." using a matching delay circuit to provide a delayed PR4 Viterbi output signal.
2) Searching using ghost patents
There are various means of using AmberScope effectively. One of the more useful and innovative features of AmberScope is it ability to find ghost patent, which are defined by our system as 'second order' patents (patents connected to patents connected to a focus patent, or 'friends of friends') that have a series of good connections to first order (directly connected) patents. We think they are useful because they are 'new' information, and so possibly can be used an opposing party to request re-examination of a patent. Ghost patents can be easily identified in an AmberScope network as they are faded out and so 'ghost like' in their appearance, as shown in the figure below, and highlighted by a red circle.
Although there are two Carnegie Mellon patents, we can start our analysis with CM1. So what the ghost patents for this patent?
In the example below, all of these ghost patents have have given a relevancy score of 0 (purely to colour them green) and been added into the table below, which has been downloaded and formatted below. For comparison, we have also added the Seagate patent referred to above in this table.
One of these patents is the US5949831 patent identified earlier. As this technology is quite complex, technical specialists in this area may be required to confirm the ultimate relevance of this potential prior art to the two Carnegie Mellon patents.
As you can tell from Table 1, some of these ghost patent appear to disclose some of the key elements of Carnegie Mellon patents, yet have not been cited by the patent examiner. It is of up to the courts to decide whether these other citations are relevant to Carnegie Mellon patents, but this shows how quickly AmberScope can identify potentially relevant prior art.
A further observation from this table is that the two Carnegie Mellon patents do not have particularly high AmberScore values. CM2 has an "AmberScore" (a measure of network connectiveness) of 4.4, which is 4.4 times the average AmberScore for US patents granted in the last 20 years. 4.4 is obviously better than average, but would not ordinarily predict a billion dollar infringement.
Marvell have made a similar point on their website, noting that the 50 cent per chip royalty failed to consider the:
'value attributable to that (patented by Carnegie Mellon) functionality, as opposed to other improvements.....more than 80 additional features were added at the same time that the (patented) feature was added.'
Obviously any such statements released by a litigant in the middle of litigation are likely to reflect the view and interests of the litigant. However this simple analysis would suggest that the Carnegie Mellon patents are not the most important in this 'neighbourhood', either in absolute (4.4 is not that high an AmberScore value) or relative terms (there were more dominant patents in the immediate neighbourhood).
$1.1 billion seems to me to be a very high royalty payable on a duo of patents that are far from dominant in their space. It will be interesting to see how the appeal plays out.
3) Using the already identified Seagate patent as the basis of an AmberScope search
Another option for the prior art search could be to start an AmberScope search from the already identifed Seagate patent US6282251. A preliminary search on this patent admittedly failed to identify anything directly relevant, but we did not fully explore this area. This still remains an option for searchers closer to this case to pursue.
This blog has reviewed a pair of Carnegie Mellon patents that have been litigated against Marvell Semiconductor. AmberScope was used to identify some potential prior art, and provide some perspective on the relative value of these patents. Overall, notwithstanding the complexity of the technology and value of hard disc drives, this analysis has suggested that the billion dollar infringement cannot be supported by the relative dominance of these patents. We await the appeal with interest