RVN! 26 Banner revised (800 x 100 px) (600 x 100 px)

Intel 10nm Yield Issues

Intel 10nm Yield Issues
by Scotten Jones on 04-29-2018 at 4:00 pm

On their first quarter earnings call Intel announced that volume production of 10nm has been moved from the second half of 2018 to 2019 due to yield issues. Specifically, they are shipping 10nm in low volume now, but yield improvement has been slower than anticipated. They report that they understand the yield issues but that improvements will take time to implement and qualify.

During the question and answer segment they said it is “really tied to this being the last technology tied to not having EUV and the amount of multi patterning and the effects of that on defects”.

This has led to a lot of speculation about what is causing the yield issues. I have seen some comments that everyone is doing multi-patterning Intel’s explanation doesn’t make sense and there is some speculation that the yield issues are related to cobalt. I thought it would be useful to explore multi-patterning usage and cobalt usage and how these differ between companies and what the impact may be on yields.

There are four companies currently pursuing leading edge logic: GLOBALFOUNDRIES (GF), Intel, Samsung, and TSMC. The following will explore multi-patterning and cobalt usage by company.

Multi-Patterning
In the front end of the line all four companies are using Self Aligned Quadruple Patterning (SAQP) with multiple cut masks for Fin formation and Self Aligned Double Patterning (SADP) for gate formation. At contact some versions of Litho-Etch are used, either Litho-Etch-Litho-Etch (LE2), or Litho-Etch-Litho-Etch-Litho-Etch (LE3) or even LE4 (EUV for Samsung). These are all similar between the companies except for Samsung’s use of EUV. In the Back End Of Line (BEOL) is where we see a significant differences. GF and TSMC both use SADP for critical metal layers, Intel uses SAQP for 2 metal layers and Samsung is expected to use EUV for critical metal layers.

We believe GF and TSMC are both ramping yield on schedule. It is possible that the yield issues Intel is seeing are related to SAQP in the BEOL. BEOL metal layers require multiple block layers and this is complex to implement. The first block layer would remove the layers needed for subsequent block layers, the way this is addressed is block layers are applied as reverse images and then once all the block layers are done, the whole pattern is reversed. Implementing multiple block layers in concert with SAQP versus fewer block layers at SADP is more difficult. This could explain why multi-patterning may be more difficult at Intel. Intel has a 36nm pitch in the BEOL versus GF and TSMC at 40nm and the smaller pitch is more difficult to achieve. We don’t know much about Samsung’s process yield ramp or exact specifications but certainly their early use of EUV may presents some challenges for them and we wouldn’t be surprised if Samsung encounters yield issues as well.

Cobalt Usage
I have also seen comments about cobalt usage suggesting Intel’s use of cobalt may be the issue. The first comment I want to make here is everyone is using cobalt, not just Intel although there are differences in usage.

  • Liners/caps – we believe all four companies are using cobalt for liners and caps on critical metal layers. Historically liners are Ta/TaN and switching to Co/TaN improves electromigration and the copper “wetting” during processing. Cobalt caps on top of the metal lines also improve electromigration.
  • Contacts– we believe all four companies will also use cobalt filled contacts although there may be differences in how it is deposited (more on this later).
  • Local interconnect – Intel uses cobalt filled metal lines for M0 and M1, GF does not and we don’t think TSMC does either. A key here is that as interconnect pitch shrinks, copper resistance goes up and eventually cobalt becomes a lower resistance solution. We believe Intel went to cobalt because it is beneficial for resistance at 36nm, with GF and TSMC at 40nm they likely didn’t see the need. We are curious to see what happens with Samsung, we believe they may also have a 36nm minimum metal pitch and it will be interesting to see if they use cobalt interconnect. They are co-authors on technical papers for 7nm with cobalt M0 so they have certainly looked at it.

We know that GF uses CVD to deposit cobalt for their cobalt filled contact and we have heard that Intel deposits cobalt with plating. We have also heard that Intel may have void issues. Perhaps plating cobalt is creating some cobalt issues, we do not think there are fundamental issues with cobalt.

Conclusion

I believe Intel’s comment on multi-patterning issues is probably the driver of their yield problems. They were more aggressive in their shrink than others and getting to 36nm minimum metal pitches with SAQP and multiple block layers is in my opinion the likely problem.

Cobalt may also be a contributor but I don’t think it is the main problem.

Also Read: Intel delays mass production of 10nm CPUs to 2019


Data Center Powers Intel but 10NM Still Slow

Data Center Powers Intel but 10NM Still Slow
by Robert Maire on 04-29-2018 at 12:00 pm

Intel (INTC) blew away expectations based on strong performance in the data center. Revenues of $16.1B versus street of $15.05B and EPS of $0.93 versus street of $0.72. While revenue was up 9% over prior year, earnings were 50% higher. Guidance is for Q2 revenue of $16.3B and EPS of $0.85 versus street of $15.55B and EPS of $0.81. IOT, NSG and PSG were also up a nice 18%.

We probably could have guessed that Intel would be a big beneficiary of the huge uptick in capital spending at Google who is obviously rolling out data centers and capacity as quickly as possible on the back of strong business.

Data center spending has long legs in our view and we think 2018 could be a super year for Intel’s data center group. Intel is well positioned to capitalize on this.

Intel’s financial performance and discipline has been very good and we think management will keep a tight rein on the model and profitability.

The only fly in the ointment was the long delayed 10NM roll out is still not rolling out. Though management talked about some products shipping, its pretty clear its not in volume. Intel and 10NM sounds a lot like ASML and EUV. Its coming, its just around the corner and the check is in the mail….

The delay hasn’t hurt Intel yet, and this quarters financial performance obscures the technology failings.

KLAC had a great quarter, breaking the $1B mark, beating both revenue and EPS expectations.
The most important fact that may be overlooked is that KLAC is projecting an increasing year with H2 higher than H1 , versus Lam that is projecting the opposite, a softening year. KLAC has three things going for it; Memory, China & EUV.

KLAC is seeing well over half its business from memory plus the additional driver of China which needs to buy metrology and yield management tools ahead of process tools made by Applied and Lam. Additionally KLAC gets a benefit from the transition to EUV as lots of new metrology and inspection tools are needed to deal with the new problems associated with EUV. This is compared to AMAT and Lam that see etch and dep steps reduced when process flow is replaced with shorter, but more difficult, EUV process flow.
KLAC remains in a very good position for 2018.

Tick, Tock, Tock, Tock, Tock, Tock…….
Moore’s Law in hibernation at Intel….It’s groundhog day all over again.

Intel, the company built on Moore’s Law and maintaining a technology leadership position has been stalled at 14NM going on 10NM since 2015. We were one of the first to point out Intels delay way back when. In the meantime both TSMC and Samsung seem to have caught up and may be about to pass Intel. Intel’s 10NM is about the same as TSMC’s 7NM and so far it looks neck and neck.

Also Read: Leading Edge Logic Landscape 2018

Broadwell, Skylake, Kaby Lake, Coffee Lake…we are getting tired of being under water at 14NM….maybe the next 14NM parts should be “Coffee Cake”…enough already, get on with it

Uptick in capex from $14B to $14.5B probably EUV high NA

We coincidentally saw an uptick of $800M in capex at TSMC for new mask capacity and a $300M pre payment to ASML for a place holder for a high NA system sometime in the future of ASML’s product line. We would be willing to bet that at least part, if not most ($300M) of Intel’s uptick is earmarked for a placeholder at ASML’s high NA waiting list.

KLAC has three strong drivers- Memory, EUV & China

KLAC has great positioning in that it has a more diverse set of business drivers as compared to others in the industry.

Both EUV and China are going to be multi year, long term secular growth stories that will go on despite what happens in memory of with Samsung. While a lot of new capacity in China is directed at memory, we think China will continue to spend even if the existing players such as Samsung, SK and Micron slow in the face of overcapacity. China needs and wants to get a foothold in the memory market and essentially has to build.

Also Read: SPIE Advanced Lithography 2018 – ASML Update on EUV

Transitioning to EUV is another “must have” for the industry. Sooner or later, Samsung, TSMC and Intel will all go to EUV and its clearly going to be very painful and expensive but there is simply no choice. The industry can delay the inevitable only so much, sooner or later (probably at 5NM) its going to happen and KLA will sell a lot of metrology and yield management tools to sort it all out.

An improving year is better than a slowing year
KLA is looking at high single digit to low double digit growth in 2018 with H2 bigger than H1. This is certainly better than a weakening H2 projected by Lam even though it was just a slow softening.
We thinks KLAC’s diversified drivers puts them in a much less risky position as compared to Lam at 84% memory and nowhere to grow.

We are also concerned that even though Applied has diversification in display tools, its highly likely that display revenue will be down sharply as Samsung cuts off display spending very quickly. This suggests that KLAC has better market positioning than either AMAT or LRCX for the near balance of the year.

The stocks
We saw both Intel and AMD blow away numbers as data center spend has been great as evidenced by Google. We think both stocks could still be buyable here as the data center spend is not going away any time soon.

Likewise, we think KLAC is one of the better positioned, less risky and more diversified plays of the semiconductor equipment tool makers. Business remains strong and future upside from Orbotech will add to the story in the fall.

And we still like Micron……

Also read: TSMC Adds Negative Semiconductor News


Samsung has another record quarter in chips

Samsung has another record quarter in chips
by Robert Maire on 04-29-2018 at 7:00 am

Samsung throws further gas on the fire of weak handset and CAPEX not set but will be down versus 2017. Samsung reported revenues of KRW 60.56 Trillion and KRW 15.64 Trillion operating profit ($56B and $15B). Chips accounted for whopping KRW 11.55 Trillion in operating profit on revenues of KRW 20.78 Trillion ( $11B and $19B)….a huge margin!

Capex was KRW 8.6 Trillion of which semiconductor was KRW 7.2 Trillion and display was KRW 0.8 Trillion ($8B, $6.7B and $740M) Although capex for the year has not been set yet the company forecasts a decline;

“Samsung has not yet finalized its capex plan for 2018, but the company expects it to decline YoY. Capex rose substantially in 2017 due to efforts to respond to market growth and emerging technologies, which included expanding the production capacity for flexible OLED panels.”

The company expects its memory business to continue strong into the second quarter but mobile handsets and display business will be down owing to weak demand (obviously pointing a finger both at their own handsets as well as their supply of OLED to Apple)

Samsung is projecting continued strength in both NAND and DRAM going into Q2 as demand maintains good pricing. It’s clear that memory products will be making up for the rest of the company for the next few quarters as handsets slow and displays follow. So far there seems to be no end in sight for memory but we still have concern about the weak handset disease spreading into DRAM

More bad news for smartphones and semi equipment
Samsungs report clearly blames weak handset sales as the culprit for holding back performance. Display sales to others, such as Apple just mean that Samsung gets a double whammy of weak handset problems. The saving grace is that memory is so strong that it easily overpowers even the bad handset news.

Moderating capex is a good thing
Samsung may be slowing capex to make sure it doesn’t get into an oversupply condition given all the capacity coming on line. The weak handset market which has only developed lately is likely the reason that they haven’t set the new, lower capex as they are likely still figuring out the impact of weak sales.

This may be a little hard to do as we are going through a seasonally weak period for handsets so it may be harder to figure out what’s seasonal and whats secular decline.

The slowing capex matches up to what we heard out of LAM
Lams projection of second half moderation sounds like it matches expectations of its biggest customer. The only problem is that even the customer doesn’t know how much it will slow so Lam’s projection is likely an educated guess at this point. The tone from Samsung sounds like the capex cut may be a bit steeper. We are sure that display capex will get an even sharper cut given profitability issues.

Obviously Samsung doesn’t want to hinder its memory business which is going gangbusters but it also doesn’t want to get to an oversupply condition and its probably better to come in on the low side of capex.

We’ll see what KLAC says
Given the negative flow of earnings news so far we have set low expectations for KLAC on Thursday. Metrology and yield management are less memory centric and are bought at different points in the life of a fab or a process so their view may differ some and probably differs more from Samsung than Lam which is in lock step with it biggest customer. KLAC also does not have the 84% memory exposure that Lam does. Samsung does sound like foundry is doing well just not as well as the blowout performance on the memory side.

Obviously when Applied reports we will likely get similar guidance as we heard out of Lam but that is weeks away.

The Socks
The Samsung news is just piling on to what we already know and beating a dead horse does not usually push the stocks down further. Sometimes investors view this as getting past all the bad news flow as we finally let all the wind out of the sails.

What it does reinforce is that the recovery from this negative news will not be all that quick and will impact even Samsung over the next few quarters.

Teflon NAND
So far the NAND business and even DRAM has remained largely impervious to other issues in tech and the semi business. The pile on of negative news just increases our nervousness to even higher levels as things would get really bad really fast if memory fell off….but so far it hasn’t….we are still buyers of Micron because of that.

Also read: TSMC Adds Negative Semiconductor News


Webinar: ASICs Unlock Deep Learning Innovation

Webinar: ASICs Unlock Deep Learning Innovation
by Daniel Nenni on 04-27-2018 at 12:00 pm

In March, an AI event was held at the Computer History Museum entitled “ASICs Unlock Deep Learning Innovation.” Along with Samsung, Amkor Technology and Northwest Logic, eSilicon explored how these companies form an ecosystem to develop deep learning chips for the next generation of AI applications. There was also a keynote presentation on deep learning from Ty Garibay, CTO of Arteris IP.

Over 100 people showed up, including myself, for an afternoon and evening of deep learning exploration and some good food, wine and beer as well. The audience spanned chip companies, major OEMs, emerging deep learning startups and research folks from both a hardware and data science/algorithm point of view. The event covered a lot of ground.

For those who couldn’t make it and interested parties around the world, there will be a webinar version of this event broadcast on May 2 from 8-9AM and 6-7PM Pacific time. I’ll be introducing the webinar and moderating the event. More than 400 people have registered already which is a record number for webinars I have been involved with, absolutely! You can sign up here:

ASICs Unlock Deep Learning Innovation HBM2/2.5D Ecosystem for AI Webinar
Deep learning algorithms, powered by neural nets, hold promise to automate our world in ways previously reserved for science fiction. Computers and cell phones that recognize us and talk to us, along with cars that drive us are just a few of the revolutionary products on the near horizon.

Practical implementation of this technology demands extreme performance, low power and efficient access to massive amounts of data. Advanced ASICs play a critical role in the path to production for these innovations. In fact, many applications cannot be realized without the performance and security that a custom chip provides.

What is needed is an implementation platform supporting 14nm and 7nm FinFET process nodes to address the challenges of deep learning.

Please join Samsung Electronics, Amkor, eSilicon and Northwest Logic as we explore a complete implementation platform for deep learning ASICs. The webinar will be moderated by Dan Nenni, CEO and founder of SemiWiki.

May 2, 2018
8:00-9:00AM or 6:00-7:00PM

8AM Registration and 6PM Registration

And here is my opening statement thus far:

Our insatiable need for data is driving IP address traffic growth to increase by 3X from 2015 to 2020. In the five years following 2015 there will be (as we are seeing now) a dramatic increase in the number of connected devices and the improvement in broadband speeds by almost 100%. This coupled with the increase in internet users and the huge amount of video that we are posting (and viewing) is driving semiconductor companies to build highly complex chips to meet the underlying requirements for bandwidth. WAN redesign: moving applications to the cloud and improving Ethernet switching speed from 40Gbps to 400Gbps by 2020.

Deep learning is a specific machine learning technique that uses neural network architectures, requiring large amount of data & compute power. There is a need for hardware acceleration for deep learning computation.

Deep learning is deployed today in all major data centers (cloud) and in devices (edge)

Deep Learning chips have 2 main functions: Training & Inference

Deep Learning applications can be divided into 3 main categories covering most industries:

  • Image/Video: (object/image/facial recognition) Main industry: Automotive, Social Media, IoT, Advertising, Surveillance, Medical Imaging
  • Voice: (speech recognition, language translation) Main industry: Social Media, Smart Homes, IoT
  • Text/Data: (data mining, big data analytics, decision making) Main industry: Finance, cloud services, research.

Deep learning chipsets focus on 2 main functions – training and inference.

Training:

Training the neural network requires massive amount of training data, storage and compute power. Training ICs are typically in the cloud and some high-end devices.

Inference:

Uses the trained network to provide an ‘intelligent’ analysis/result which requires less data, storage and compute power than training. Inference ICs are typically in the AI devices and some in the cloud for low latency applications.

I hope to see you there!


Achronix Momentum Building with Revenue Growth, Product/Staff Expansion, New HQ

Achronix Momentum Building with Revenue Growth, Product/Staff Expansion, New HQ
by Camille Kokozaki on 04-27-2018 at 7:00 am

5G Wireless, Network Acceleration, Data centers, Machine Learning, Compression, Encryption fueling the Growth

Building on its increasing momentum, Achronix Semiconductor Corporation held a ribbon-cutting ceremony on Tuesday, April 25, with the presence of Santa Clara’s Mayor Lisa Gillmor, customers, and partners, employees and company executives. President & CEO Robert Blake addressed the attendees with details on the company’s roots, from its founding in 2004 to its move from New York to the Valley in 2006. He thanked the investors and the employees for their continuing support and dedication in getting them to the current state of fast growth, exceeding $100 Million in revenue, reaching 700% growth year-over-year, achieving profitability and prompting an increase in talent search to develop and support their growing customer base.

He highlighted the evolution from the 80’s glue logic integration to ASIC prototyping to the current state-of-the-art implementation efficiencies replacing in certain cases CPUs, GPUs, and ASICs in Data Acceleration, Machine Learning, Artificial Intelligence applications. Blake stressed the importance and the contribution to their success by their partners in the Semiconductor, EDA, Packaging, IP, Test, Manufacturing, in addition to their customers and employees. He was followed by Santa Clara’s Mayor Lisa Gillmor who expressed her delight to see Achronix grow amid Santa Clara, a city rapidly becoming Silicon Valley’s technology center of the world. She reminded the audience that a testament to Achronix success was the 2017 Americas ACE Awards selection of Achronix as Company of the Year.

Santa Clara Mayor Lisa Gillmor and Robert Blake, Achronix CEO inaugurating the new Achronix headquarters

The new HQ facilities will occupy 25,000 square feet and house from 75-80 employees out of about 100 worldwide and is located at 2903 Bunker Hill Lane, Santa Clara right by the Santa Clara Convention Center and Levi’s Stadium. Achronix has a Research and Development and Design Center presence in Bangalore, India and other sales offices and representation in the US, Europe, and China.

Achronix offers both FPGA and embedded FPGA leading technologies with its Speedster® FPGA family and its Speedcore™ eFPGA IP solutions, respectively. These offerings are in high demand in the high-growth market of hardware accelerators designed to offload CPU data processing and supercharge system performance for applications such as 5G wireless, AI/ML, high-performance computing (HPC), data centers, and automotive stressed Steve Mensor, Achronix VP of Marketing. Speedster® 22i FPGAs support up to 1 million effective Look-Up-Tables (LUTs) and 86 Mbit of embedded RAM built o Intel’s 22nm FinFET process. Speedster22i HD devices include embedded hard IP for communication applications including 10G/40G/100G Ethernet, 100G Interlaken, PCIe Gen3 x8 and DDR3 ×72. Additionally, Speedster22i FPGAs have up to sixty-four (64) lanes of 10.375 Gbps SerDes and up to 996 high-speed general purpose I/O.

Achronix also provides accelerator boards and design tools. The Achronix PCIe Accelerator-6D board is the industry’s highest memory bandwidth, FPGA-based PCIe add-in card for high-speed acceleration applications. On the card is the Speedster22i HD1000 FPGA, which connects six independent DDR3 memory controllers allowing for up to 192 GB of memory and 690 Gbps of memory bandwidth. The ACE tools work with industry-standard synthesis tools, easing the mapping of user designs into Achronix.

Achronix is currently developing its next-generation FPGAs based on TSMC’s 7nm process node, capable of up to 5 Million LUTs and a core frequency of 750 MHz. The ACE tools will be extended to address Machine Learning (ML), AI, search, compression and encryption solutions and use cases.

About Achronix Semiconductor Corporation
Achronix is a privately held, fabless semiconductor corporation based in Santa Clara, California. The Company developed its FPGA technology which is the basis of the Speedster22i FPGAs and Speedcore eFPGA technology. All Achronix FPGA products are supported by its ACE design tools that include integrated support for Synopsys (NASDAQ:SNPS) Synplify Pro. The company has sales offices and representatives in the United States, Europe, and China, and has a research and design office in Bangalore, India.


Monitoring Process, Voltage and Temperature in SoCs, webinar recap

Monitoring Process, Voltage and Temperature in SoCs, webinar recap
by Daniel Payne on 04-26-2018 at 4:00 pm

Have you ever wondered how process variation, thermal self-heating and Vdd levels affect the timing and yield of your SoC design? If you’re clock specification calls for 3GHz, while your silicon is only yielding at 2.4GHz, then you have a big problem on your hands. Such are the concerns of many modern day chip designers. To learn more about this topic I attended the 45 minute webinar from Moortec, titled “The Importance of Monitoring Process & Voltage in Advanced Node SoCs“. Ramsay Allan provided the introduction and overview, then Stephen Crosher, CEO presented about 20 slides on the challenges for IC design at 40nm down to 7nm, along with their semiconductor IP used for in-chip PVT monitoring.

Here’s a summary of physical effects that adversely effect chip timing and reliability:

  • Thermal hot-spots
  • Device and process variability
  • Increased resistance of interconnect from 40nm to 7nm
  • Power Delivery Network
  • Lower Vdd trends from 40nm to 7nm, lower design margins
  • Ageing that changes Vt
  • Self-heating accelerates BTI and HCI
  • Delays increasing from interconnect resistance
  • Delays increasing from Vdd variations

The webinar was chocked full of diagrams and charts to point out all of the effects that make you worried at night, for example the increase in interconnect resistance as you progress from 40nm to 7nm nodes:

Total delays within a chip are the combination of gate delays and wire delays, so as we used smaller process nodes the percentage of total delay caused by interconnect is now approaching 50%:

Process variation now shows it’s ugly side when on the same chip you can see different process corners, making it especially challenging to do timing analysis and reach timing closure.

To mitigate these issues clever IC designers have come up with several approaches that use in-chip monitors:

  • Voltage scaling optimization per chip by finding the lowest center functional voltages to meet frequency
  • Adaptive Voltage Scaling (AVS) as a closed-loop system using on-chip monitors
  • Self-adaptive tuning
  • Embedded chip monitoring to minimize power consumption at the enterprise datacenter level
  • Using AVS to do speed binning of parts

The British chaps from Moortec founded the company back in 2005 are are experts at applying in-chip monitoring to a wide range of commercial ICs. Their semiconductor IP has both hard blocks and soft blocks combined to create a subsystem for in-chip monitoring.

Placing multiple PVT sensors on a single chip makes sense, but how many of them should you add, and exactly where should they be placed for optimum impact? Good questions, so rely upon the technical support that comes along with their service. Placement of IP is really application specific.

For me the most powerful bit of information was saved for near the end when they unveiled a list of customers using their in-chip monitoring.

I plan to visit Moortec at DAC in San Francisco, however they’re also having multiple events across the globe to support their unique IP.

Summary
Having a plan to meet critical timing by having in-chip PVT monitors as part of a subsystem makes more sense as you reach the 40nm process node and going to ever smaller geometries. Yes, you could cobble together something proprietary in-house if you have lots of spare engineering resources and the time to design, verify, fabricate and test a one-off system. My hunch is that your product schedule and budget would be better served by looking at something off the shelf from Moortec, because that is their sole focus and the proof is in their ever-expanding list of adopters. They’ve setup distributors in all of the high-tech centers around the world and that would be a good starting point to learn more about their technology, approach and benefits.

Related Articles


Open-Silicon, Credo and IQ-Analog Provide Complete End-to-End Networking ASIC Solutions

Open-Silicon, Credo and IQ-Analog Provide Complete End-to-End Networking ASIC Solutions
by Camille Kokozaki on 04-26-2018 at 12:00 pm

The end-to-end principle as defined by Wikipedia is a design framework in computer networking. In networks designed according to this principle, application-specific features reside in the communicating end nodes of the network, rather than in intermediary nodes, such as gateways and routers, that exist to establish the network. [1]There are usually tradeoffs between reliability, latency, and throughput. High-reliability networks usually negatively impact the other components of the parameters of this data transmission triad, namely latency and throughput. This is particularly important for applications that value predictable throughput and low latency over reliability – the classic example being interactive real-time voice applications.

Another example of end-to-end application is the network that handles on-demand content delivery from preparing, packaging the audio/video assets adding metadata, transcoding them and then sending them to distributors. It goes without saying that end-to-end networking solutions need to satisfactorily address these requirements for fast reliable transfer and delivery and it is thus not surprising to see that ASIC standardized solutions are best suited for the task. Those solutions are needed by leading-edge networking applications, such as long-haul, metro and core, broadband access, optical, carrier IP and data center interconnect use cases.

Open-Silicon, Credo, and IQ-Analog have put together a complete end-to-end ASIC solution for leading-edge networking applications, such as long-haul, metro, and core, broadband access, optical, carrier IP and data center interconnect which they have showcased at OFC 2018 last month.[2]Open-Silicona comprehensive Networking IP Subsystem Solution, which includes high-speed chip-to-chip interface Interlaken IP, Ethernet Physical Coding Sublayer (PCS) IP, FlexE IP compliant to OIF Flex Ethernet standard v1.0 and will be compliant to the upcoming v2.0, and Multi-Channel Multi-Rate Forward Error Correction (MCMR FEC) IP. Open-Silicon complements this with its High Bandwidth Memory (HBM2) IP Subsystem Solution.

Credo has its high-speed 56Gbps PAM4 LR Multi-Rate SerDes solution and 112Gbps PAM4 SR/LR SerDes targeted for next-generation networking ASICs. IQ-Analog rounds up the solution with its high-performance, patented TPWQ hyper-speed 90Gsps ADC/DAC IPs analog-to-digital converters (ADCs) and digital-to-analog converters (DACs). This teaming up gives the three companies the opportunity to demonstrate the power of complete solutions for the next generation of high-performance networking applications.

Open-Silicon has a comprehensive Networking IP Subsystem Solution portfolio and includes:

1.High Speed Chip-to-Chip Interface Interlaken IP – Open Silicon’s 8th generation Interlaken IP core supports up to 1.2 Tbps high-bandwidth performance and up to 56 Gbps SerDes rates with Forward Error Correction (FEC). This high-speed chip-to-chip interface IP features an architecture that is fully flexible, configurable and scalable, making it ideal for high-bandwidth networking applications, such as routers, switches, Framer/MAC, OTN switch, packet processors, traffic managers, look aside processors/memories, data center applications, and several other high-end networking and data processing applications.
http://www.open-silicon.com/open-silicon-ips/interlaken-controller-ip/

2.
Ethernet Physical Coding Sublayer (PCS) IP – Open-Silicon’s Ethernet PCS core is compatible with different MII interfaces for connecting to the MAC and is uniquely built to work with off-the-shelf MAC and SerDes from leading technology vendors. It supports 64b/66b encoding/decoding for transmit and receive, and various data rates, ranging from 10G to 400G. The Ethernet PCS IP complies with the IEEE 802.3 standard and supports Ethernet and Flex Ethernet interfaces, making it ideal for high-bandwidth Ethernet endpoint and Ethernet transport applications. https://www.open-silicon.com/networking-ip-subsystem/

3.
Flex Ethernet (FlexE) IP – Open-Silicon’s FlexE IP core features a generic mechanism that supports various Ethernet MAC rates, and is uniquely built to work with Open-Silicon’s packet interface and OTN client interface or off-the-shelf MACs. The FlexE IP supports the Optical Internetworking Forum (OIF) Flex Ethernet standard 1.0 and will be compliant with the upcoming v2.0. The IP supports FlexE aware, FlexE unaware, and FlexE terminate modes of mapping over the transport network, making it ideal for high-bandwidth Ethernet transport applications. https://www.open-silicon.com/networking-ip-subsystem/

4.
Forward Error Correction (FEC) IP – Open-Silicon’s FEC IP core is capable of multi-channel multi-rate forward error correction in applications where the bit error rate is very high, such as high-speed SerDes 30G and above, and significantly improves bandwidth by enabling 56G PAM4 SerDes integration. This single-instance IP core is compatible with off-the-shelf SerDes from leading technology vendors and supports bandwidths up to 400G with the ability to connect 32 SerDes lanes. It can easily achieve a Bit Error Rate (BER) of 10-6, which is required by most electrical interface standards using PAM4 SerDes. The FEC IP core supports the Interlaken and Ethernet standards and significantly improves bandwidth by enabling high speed, multi-channel SerDes integration, making it ideal for high-bandwidth networking applications. https://www.open-silicon.com/networking-ip-subsystem/

5.
High Bandwidth Memory (HBM2) IP Subsystem Solution – Comprehensive High Bandwidth Memory (HBM2) IP Subsystem Solution for 2.5D ASICs in FinFET Technologies – This solution is now available for 2.5D ASIC design starts and also as licensable Intellectual Property (IP). The IP includes the controller, PHY and custom die-to-die I/O needed to drive the interface between the logic-die and the memory die-stack on the 2.5D interposer. Open-Silicon’s HBM2 IP subsystem is silicon proven on a 2.5D HBM2 ASIC SiP (System-in-Package) platform. The platform is used to demonstrate the high bandwidth data transfer rates of >2Gbps and interoperability between Open-Silicon’s HBM2 IP subsystem and HBM2 memory die-stack. http://www.open-silicon.com/high-bandwidth-memory-ip/

To summarize:

Open-SiliconNetworking IP Subsystem Solution

  • High-speed chip-to-chip interface Interlaken IP
  • Ethernet Physical Coding Sublayer (PCS) IP
  • FlexE IP compliant to OIF Flex Ethernet standard v1.0 and will be compliant with the upcoming v2.0
  • Multi-Channel Multi-Rate Forward Error Correction (MCMR FEC) IP
  • High Bandwidth Memory (HBM2) IP Subsystem Solution

Credo

  • High-speed 56Gbps PAM4 LR Multi-Rate SerDes solution
  • 112Gbps PAM4 SR/LR SerDes targeted for next-generation networking ASICs

IQ-Analog

  • High-performance analog-to-digital converters (ADCs)
  • High-performance digital-to-analog converters (DACs).

[1]End-to-end principle- Wikipedia

[2]Open-Silicon, Credo and IQ-Analog Showcase Complete End-to-End Networking ASIC Solutions at OFC 2018


Safety in the Interconnect

Safety in the Interconnect
by Bernard Murphy on 04-26-2018 at 7:00 am

Safety is a big deal these days, not only in automotive applications, but also in critical infrastructure and industrial applications (the power grid, nuclear reactors and spacecraft, to name just a few compelling examples). We generally understand that functional blocks like CPUs and GPUs have to be safe, but what about the interconnect? To some among us, interconnect just means wires; what’s the big deal, outside of manufacturing defects and electromigration? No joke – Kurt Shuler (VP Marketing at Arteris IP) tells me he still struggles with this perception among some people he talks to (maybe physical design teams and small block designers?).


If you’re familiar with SoC architecture, you know why interconnect is just as important as endpoint IP in functional verification, safety, security and other domains. Just as the Internet is only possible because connections are virtualized by routing through networks of traffic managers, so in modern SoCs, at least at the top-level, traffic is managed through one or more networks-on-chip (NoCs, though the implementation is quite different than the Internet). That means there’s a lot of logic in this interconnect IP, managing interfaces between endpoint IPs (such as ARM cores, hardware accelerators, memory controllers and peripherals) and managing those routing networks. Moreover, since the interconnect mediates safety-critical operations between these IPs, it is inextricably linked with system safety assurance.


Completing failure mode effects analysis (FMEA) and failure mode effects and diagnostic analysis (FMEDA) is particularly difficult for interconnect IP, first, because they are safety elements out of context(SEooC, as are all IP and generally chips), which therefore have to be validated against agreed assumptions of use. Secondly, interconnect IP are highly configurable, which makes reaching agreement on the assumptions of use, defining potential failure modes and effects, and validating safety mechanisms even more challenging than for other IP, as safety assurance must be determined on the configuration built by the integrator.

To Arteris IP, the right way to handle this complexity is to start with a qualitative, configurable FMEA analysis, which can then guide the creation of a configuration-specific FMEDA. Naturally, this requires hierarchy around modular components with safety mechanisms tied to those components, so that as you configure the IP, those safety mechanisms are automatically implemented in a manner that ensures safety for the function. For example, in the Arteris IP NoCs, the network interface unit (NIU) responsible for packetization can be duplicated. Then in operation, results from these units are continually compared, looking for variances which would signal a failure. You can also have ECC or parity checks in the transport. Following the FMEA/FMEDA philosophy, these safety mechanisms are designed to counter the various potential failure modes identified by the IP vendor. And because of the modularity, functional safety diagnostic coverage of the entire NoC can be calculated based on individual module coverage metrics. Kurt also stressed that it is important for the IP provider to work with functional safety experts (in Arteris IP’s case, ResilTech) to ensure maximum objectivity in this analysis, and naturally, to suggest opportunities to further enhance safety solutions.

Who guards the guardians? Arteris IP provides BIST functions to check their compare logic for safety mechanisms at reset and during run-time. All of this safety status rolls up into a safety controller (part of the Arteris IP deliverable), which can be monitored at the system level.


A critical part of safety analysis is deciding how to meaningfully distribute failure modes for verification. This is an area where the IP provider (with guidance from an independent safety expert) can generate a template to drive FMEDA estimation based on the specific configuration and estimated area of sub-components. Having a modular and hierarchal interconnect IP architecture is key to calculating these safety metrics required for FMEDA. In other words, the user of the IP should be able to expect a largely self-contained solution for this part of their safety validation task.

Another important advantage for an automatically generated IP when it comes to safety is that it becomes possible to automate the generation of the fault-injection campaign for verification and the rolling up of results to the FMEDA table. Rather than taking a shotgun approach to faulting, the campaign can be more precisely targeted in this manner:

  • Define where a fault must be inserted to generate one of the failure modes
  • Define traffic patterns that ensure faults don’t appear as safe faults
  • Define which safety mechanism shall detect the fault (but ensure all possible safety mechanisms detect the failure mode)
  • Define observation points as close as possible to the sub-part being tested

Again, the fault campaign is largely automated for the integrator.

Overall, this seems like a reasonable strategy, meeting requirements for functional safety and the expectation that an IP should still be a self-contained solution, even when configured as extensively as is possible with a NoC. You can learn more about the Arteris IP safety solutions HERE.


Electrical Reliability Verification – Now At FullChip

Electrical Reliability Verification – Now At FullChip
by Alex Tan on 04-25-2018 at 12:00 pm

Advanced process technology offers both device and interconnect scaling for increased design density and higher performance while invoking also significant implementation complexities. Aside from the performance, power and area (PPA) aspects, designer is getting entrenched with the need of tackling more reliability issues such as Electrostatic Discharge (ESD), Latch-Up (LUP) and Time-Dependent Dielectric Breakdown (TDDB). Traditionally these issues are observed during the cell library and technology development stage under as-designed operation voltage ranges. However, foundries or integrated device manufacturers (IDMs) are encouraging full-chip reliability verification to prevent chip reliability failures either during burn-in stage or in silicon.

Process and Reliability Design Rules
In the traditional RTL2GDS2 design flow, the standard cell library is initially targeted as the test vehicle for new process role-out as they are easier to be implemented compared with the other macro or custom blocks. Both timing and power attributes are normally captured in the library and synced-up with the foundry provided SPICE model version. Subsequently, the library could also embed process variations parameters through the Liberty Variation Format (LVF) extension.

During the 0.18 micron technology day, we were accustomed to the notion DFM (Design For Manufacturability). Since then, frequent collaborations between foundries and designers were taken place to minimize surprises from the manufacturing field. Designers would attempt to incorporate all known critical process parameters prior to tape-out. Foundry would provide PDK (Process Design Kit) and techfile releases early and maintain close-collaboration with customers’ technology and library teams to ensure proper adoption. Customer subsequently will pipe-clean and apply changes using foundry pre-approved reference flow.

Similar to PDK, a Reliability Design Kit (RDK) may be provided by foundries or IDMs, but with some critical design rules being presented as guidelines. For example, wire width threshold requirement is posted as criteria for nets in design to survive an ESD event. This rule can not be directly applied on the traditional DRC tool as it is lacking of the capability to identify electrical current directionality among others. In other examples, applying DRC driven approach to identify both LUP and TDDB is impractical as it has heavy reliance on using physical markers to identify polygons under analysis. This approach either does not work well at full-layout or may present loophole that deems it to be ineffective. Recent foundries/IDM proposed methodology is moving towards non-physical marker approach, allowing a more versatile logic-driven-layout (LDL) based checking. The combination of topological data and static design rules enable better reliability check coverage.

Calibre PERC and Reliability Bottlenecks
Mentor Calibre product family has been the industry leader for IC physical verification. Calibre PERC reliability platform provides complex reliability verification using both foundry’s standard rules and project custom rules. It employs topological constraints to verify correct circuit structures are in place as specified by circuit design rules. It has access and ability to concurrently use netist and layout information to perform electrical checks utilizing parameters in both domains.

The combination of Calibre PERL LDL flow along with static simulation and static voltage propagation features has enabled foundries and IDMs defining robust reliability rules. These rules can be implemented and automatically verified at full-chip ensuring a full coverage. Analogous to static timing approach used for resolving timing the latency and the scalability of dynamic-based circuit simulation, using the above approach cuts down the complexity and allows both block and full-chip reliability verifications.

Let’s review on how Calibre PERC utilizes static simulation and voltage propagation to address ESD, LUP and TBBD compliances.

For ESD prevention, ESD or power clamping devices
(diode, transistor, resistor) with enough strength must be connected to IO, P/G and cross-power domain paths. As shown in figure 2, the types of checks Calibre performs on this type of condition include:
– Verify if the required circuits (ESD, power clamping, back-to-back diodes) exist.
– Check device parameters of the corresponding circuits have sufficient strength for ESD.
– Additional checks for for advanced nodes such as effective resistance and current density.

The second reliability issue is latch-up (LUP). IC’s LUP can be described as a short-circuit type event occuring in the parasitic-bipolar-pair equivalent structures and is triggered when disruptive excitation causes significant current or overcurrent looping in the positive feedback network (refer to figure 3a).

For a LUP prevention, the guard-ring or strap insertion is normally recommended. In addition, it requires the spacing among polygons involved with latchup (operate at different potentials) should be equal or larger than potential difference as shown in figure 3b. Calibre PERC handles LUP check by traversing through extracted layout netlist and propagate
external voltage values into internal nets based on user-defined constraints. It has mechanism to annotate to physical polygons attributes to signify aggressor or victim devices and voltage value for DRC checks.

For interconnect TDDB checks on block or full-chip, spacing checks among polygons of the same layer but different nets are executed against criteria dependent on delta-voltage range. Voltage propagation of the external nets into the internal nets are done until all targeted nets for potential TDDB have the appropriate voltage values to annotate to the corresponding polygons. Designer could apply constraint to control the static voltage propagation across multi voltage domains. For example, in static propagation shown in figure 4: the top port has a 3.3 v while nets A, B, C can be assigned with voltage 2.5v, 1.8v and 2.2v, respectively. A voltage shift is done by user-defined subcircuit pattern or through simulation. Subsequent static voltage propagation is done, followed by annotation of the voltage values to polygons in nets of concern for DRC checking.

Increased reliability issues stemming from the advanced process technologies are introducing additional complex sign-off requirements as they are not easily resolved and scaled through the use of dynamic simulation or traditional DRC checks. Such limitation is no longer the case with the availability of Mentor Calibre PERC reliability platform. It has both capacity and fast performance to allow an accurate full-chip verification for a full design-for-reliability (DFR) compliance. For more detailed discussions on Mentor PERC related to solving reliability issues, check this LINK.


SiFive’s Design Democratization Drive

SiFive’s Design Democratization Drive
by Camille Kokozaki on 04-25-2018 at 7:00 am

There is something endearing and refreshing in seeing a novel approach unfold in our Semi-IP-EDA ecosystem currently settled in its efficient yet, let us say it, unexciting ‘going through the motions’, constantly comparing, matching, competitively and selfishly sub-optimizing what the art of the possible can be.

Enter a new breed of technologists, industry veterans, academics and evangelists articulating and embarking on testing and applying a new business model, building on agility, collaboration and continuous delivery and improvement, emulating moves from the playbook of the widely successful parallel industry of Software, DevOps, IT and Social Media. I am talking about Open Source initiatives like the RISC-V movement, crowd-sourced and building on standardization of an instruction set architecture (ISA) while allowing differentiation for each company with a set of extensions built on top of those common constructs and testing new fresh and promising approaches to business as usual. What is also admirable is the well-intentioned strive to make a better world allowing contributions from parts of the world where opportunity and access to funds and technology are lacking. SiFive is now articulating and executing on this vision.

At the just held GSA Silicon Summit in San Jose, Naveed Sherwani, SiFive’s CEO outlined in his Closing Keynote the elements of this vision and practice. He challenged the audience to explain how Instagram, a 13-employee startup, ended up being a $1B acquisition. The answer was that it provided a minimum viable product (MVP) on top of an existing stack of tools, infrastructure, and technologies that did not need developing from scratch. He posited that in our industry MVPs cost too much ($1M-$7M+), Design takes too long (9-24 Months + Fab Time) and too many experts are needed (14+ disciplines from architects to package and test and all the expert steps in between). He challenged the industry to set the following goals:

  • Reduce Cost by 90%
  • Reduce Time to 1 Month + Fab cycle
  • Reduce needed expertise to System Level

Experts need not panic that they will be made obsolete, they just need to move to a higher level of abstraction (such as moving from writing assembly code to developing application code).

The options for the industry are, moving forward, the following:

  • Automation
  • Reduction of Options
  • Cloud leveraging
  • Deployment of new business models

SiFive’s approach to realizing this has 4 key components:

  • Freedom Designer (Core/Subsystem/Chip)
  • Cloud Platform offering
  • DesignShare
  • Operations (Fabrication/Package/Test/Logistics)

By allowing the definition and reuse of templates in the Freedom Designer, the individual core blocks can be specified, documented and incorporated into the design fabric. By providing a Cloud Platform Offering an IP can be verified while allowing the protection and security of the crown jewels and intellectual property of the offerors. With DesignShare, 3[SUP]rd[/SUP] Party providers supply their IP to SiFive’s Platform at zero cost, increasing design starts, and collecting NRE and royalties when production starts. This also allows interoperation of their IP with other vendors’ and customer internal IP at no initial cost for use and verification by the target customer, deferring the IP charge to a later stage and reducing development and verification time. The initial DesignShare participants were … shared with the audience.

While not explicitly addressed by Naveed, my take on this is that it requires a paradigm shift with some resilience to concerns about ‘Now my competitors can compare my metrics to theirs’ and other reservations about this business model. I view it instead as sharpening the edge and leveling the playing field while concurrently stepping up to match all (if I can mix my metaphors with abandon). Another question from IP providers might be ‘Is this a race to the bottom, selling wise?’. I think not, as selling more design starts is more revenue, albeit deferred, and there is nothing like emerging momentum to make all join with a fear of missing out on the wave.

SiFive is proposing to build a Core, Subsystem, and Chip Design Factory Software Platform with predesigned blocks and components, bringing down design cost 10x from a typical $7.5M down to $750K.

Naveed then walked the talk by showcasing HiFive Unleashed, the World’s first multi-core RISC-V development board booting Linux

Naveed summarized the dramatic benefits of this business model reducing prototyping cost, allowing more startups, design starts, and IP providers, reducing the needed expertise and allowing design contributions by excited young technologists from the abstracted software and hardware worlds. A pledge to offer his software free to all universities and to the fifteen poorest countries was admirable in terms of commitment and smart in setting the stage for a new generation of millennial contributors and entrepreneurs to innovate and prosper. Not specifying which millennial helps preserve inclusiveness in this ongoing revolution, count me in.

Disclosure: I am an active participant in the RISC-V ecosystem which includes SiFive

Related Article