Vector Databases – Delivering Value through AI
Created with DALL·E 3

Vector Databases – Delivering Value through AI

Part 3 – High Tech Manufacturing Use Cases

As we saw in Parts 1 & 2, Vector Databases enable extremely fast, efficient searching and analysis of multi-dimensional data, making them essential for developing scalable and efficient AI systems, in particular uses cases for? image recognition?and anomaly detection and decisioning. Now let’s look how AI can be applied in HTM and chip manufacturing use cases.

With so much data being produced daily using industrial IoT and real time connected systems, semiconductor manufacturers have already turned to machine learning and deep learning neural networks to better analyse the enormous amounts of data and make decisions in real time in the fabrication plants, to stay ahead of their competition. To give you an idea of the data volumes of a semiconductor fab plant, at one Asia fab plant they have over 2000 real time connected line tools that create in excess 2 Terabytes of data per day. And that is only 1 of 13 fabrication plants this company operates.

?

Firstly, some background on semiconductor creation. ?Chip companies create large silicon wafers which contain multiple chips at once, and the wafer is then sliced into many individual processor dies. This is a slow and meticulous process with ultra fine tolerances. It's normal, especially early in the life of a new manufacturing process, for many of those dies to end up with defects — either they do not work, or they don't perform to the required specifications. ?The percentage of usable chips that come off of each wafer is known as the yield. ?Defects that impact the yield can be caused by various factors but key ones are the cleanliness of the clean rooms, the accuracy of the manufacturing equipment ?(deposition, photoresist coating, lithography, etching, ionization and packaging ) and the conditions of each step process.


A Semiconductor Wafer

Now lets’ look at yield management which will be one of the most important applications of AI in Semiconductor Manufacturing going forward.

TSMC (Taiwan Semiconductor Manufacturing Company) is the world’s leading manufacturer of advanced computer chips, producing over 50% of all advanced semiconductors sold globally. ?At the leading edge of chip manufacturing, TSMC has now delivered 3nm chips (A human hair is 80,000nm-1000,000nm wide as comparison) ?, which are powering the latest Apple iPhone 15’s and iPhone 15 Pro’s, the new Apple Laptop M3 cpus ?and also the upcoming Nvidia RTX5000 GPU’s. ?Currently about 90% of TSMCs 3nm production is dedicated just to Apple.

Creating a 3nm wafer for Apple from start to finish takes about four months and for the A17 chip which has over 19 billion transistors and is only about 100mm2 in size, roughly 620 chips are produced per wafer?or about 450 M3 chips per wafer. Presently the foundry price of advanced 3nm wafers is about $17,000-$20,000 each and they produce about 70,000 wafers per month (Reference 1).

TSMC is currently delivering a yield rate of around 55% good chips for its new 3nm manufacturing process.

What’s the importance of yields to Wafer manufacturing you ask?

Simply put, increased yields equals increased profit.

Just a 1% increase in yield can mean as much as $300 million dollars of extra revenue in the above scenario. A very compelling value engineering business case to increase yields.

As TSMC works to increase their yield percentages, it has had to resort to other commercial incentives to keep Apple satisfied. ?Most chip buyers have a ‘wafer-buy’ deal, ?where the clients usually has to pay for the wafer and all of the dies it contains, including any defective ones but TSMC is only charging Apple for good finished chips. That means TSMC has to absorb the cost of any defect dies in the wafer but that’s because Apple is such a valuable customer who contributed over $17.5 billion in revenue to TSMC in 2022 alone.

So ,we can see it is a financial imperative to increase yield. History has shown that though advances in technology and better manufacturing processes and tolerances , yields have been able to be increased about 5% per quarter.

As a comparison, Samsung, one of TSMC’s competitors is already producing 3nm chips at a reported yield rate of greater than 60%.

?

So how can AI and Vector Databases Increase yields?

During silicon wafer processing, impurities and particles are deposited on wafer surfaces or are left over from previous process steps. These particles can cause defects in the final semiconductor product, so the created wafers undergo analytical analysis where the wafer is scanned for particles ,the analytical tools will publish a wafer map that contains the location of the particles and sampled for failure analysis . Humans manually review these wafer maps to check for possible issues but due to the sheer amount that are generated, many issues go unnoticed until it is too late. ??There are advanced yield optimization techniques that leverage image recognition techniques with a certain amount of machine learning for semiconductor production quality and anomaly detection already in use, but going forward new AI applications are being proposed and developed that can automatically detect and classify issues as soon as the wafer maps are generated. As the volumes of wafer maps are very large , it is a perfect use case to store these wafer maps in a Vector Database ?and leverage the power of AI driven similarity search and relationships to historical manufacturing data to efficiently recognize and understand the images and anomalies in real time. Corrective action than then be taken and thereby increasing yield and reducing manufacturing waste and cost. ?This will have a direct impact on revenue. Vector database vendors such as Milvus, Pinecone, KX, and Weaviate are well positioned in this space to provide core capabilities.

?

Other Generative AI Semiconductor use cases.

But there are many other applications for data driven Generative AI in the modern high tech manufacturer, and it presents an enormous opportunity to AI ISV’s for their products and solutions. ?These include:- ?

1.???? Optimizing Chip Design

Engineers can optimize layouts design, performance, power consummation and area, by extracting intricate patterns from vast datasets and assemble new innovative designs compared to traditional EDA tools by using innovative new LLM’s and RAG. One example is Nvidia’s new Gen AI ChipNeMo LLM with 43 billion parameters that they are now using for chip design. (Reference 2)

2.???? Shorten Design Cycles

AI applications will shorten the design process to explore, test and refine complex design cycles and thereby reducing time to market.

3.???? Lights out Full Automation.

Lights-out manufacturing is the methodology of fully automating the production of goods at factories.

Already fabrication plants are semi-autonomous with the automated robotics, delivery vehicles but Samsung recently announced now plans to eliminate all humans from it’s Fab plants by 2030 to dramatically increase yields and efficiencies. ?(Reference 3)

This ambitious plan will require new levels of machine and data integration, completely autonomous tooling , Edge IIOT and AI driven oversight.

4.???? Connected Value chain

Generative models can simulate multiple scenarios considering variables like demand fluctuations, resource availability, and supply chain factors

A real time , data driven, AI enhanced strategy is driving the new “AI Factory of the future” vision. The fabrication plant data silos of the past with incompatible systems using legacy integration and analytics are rapidly being replaced with a completely connected fab plant and wider supplier ecosystem that utilizes vendor agnostic cloud based fabrics and AI design and management applications

The Open Innovation Platform (OIP) Ecosystem alliance with AWS, Microsoft, Cadence and Synopsys is just one example of new interoperability.

5.???? Predictive maintenance

AI driven predictive maintenance applications will allow for optimization of production parameters, such as temperatures, pressures, fine tuning tolerances of machines, and reduced down time resulting in enhanced yields and reduced defects.

6.???? Data Driven Decisions

Increasingly there is a movement away from legacy integrations and siloed databases to real time streaming multi cloud fabrics where AI and analytics can operate seamlessly on a combination of real time and historical data. ?For example, one semiconductor Fab plant I was at, had over 2000 separate MySql databases for production , control and reporting that they were trying to modernize.

Their desire for multi cloud architectures, is driven by the need for data and system interoperability and openness, risk mitigation and delivering the best commercial terms between cloud vendors to minimize OpEx.

Interleaved with these new data fabrics is the desire to run real time analytics ?but also new cross Fab Gen AI analysis and decision making that can be run internally and across their wider supply chain with the relevant security, governance and control.

AWS, Microsoft, Synopsys, Apache Spark, Snowflake and Dataiku are some of the leaders to watch here in this space.

?

So we can see the emergence of AI and Vector databases in high tech manufacturing will fundamentally alter the semiconductor landscape going forward. ?Seamless automation, connectivity, cloud and Generative AI will form the basis of the “AI Factories of the future” ?. Combine that with human creativity and ingenuity , and we will see amazing new levels of Innovation by the Semiconductor manufacturers ’s that will power the future of AI.

?

In Part 4, we will change direction and look at the application of AI and Vector Databases in the Healthcare industry and how AI can literally save your life.

?

#AI #Data #Innovtion #VectorDatabases #HTM #innovation #semiconductors? #milivus #KX #Pinecone #Weaviate #NVIDIA #AWS #Microsoft #Snowflake, #Dataiku #Synopsys

?

References

1.???? https://wccftech.com/tsmc-3nm-wafer-production-reaching-100000-units-by-end-of-2024/

2.???? https://www.eetimes.com/nvidia-trains-llm-on-chip-design/

3.???? https://www.tomshardware.com/tech-industry/samsung-plans-to-eliminate-humans-from-its-chip-fabs-by-2030-push-for-full-automation-continues-at-full-steam

?

?



An interesting update in APAC, ?TSMC had now progressed to 2nm technology nodes, including the innovative NanoFlex technology. It will begin mass production in 2025. It’s scheduled to boost performance of 10% to 15%?and a reduction in power consumption by 25% to 30%. Faster, cooler it will boost the data engines of the future, but all this comes at a much higher wafer price. It looks like TSMC's quote for a 300-mm wafer process using its N2 technology will exceed $30,000, where as a N4/5 wafer only cost ~$17,000 today. Of course this varies by volume and customer, but it will be much more profitable for TSMC, although they are building a couple of new fabs to make the N2 chip at a cost of 10’s of billions of dollars. Expect to see N2 chips in apple products in in early 2026. Mike Alperin #innovation #HTM #semiconductors

回复

The link below is a great overview of the semiconductor manufacturing process by ASML. They are one of the world’s leading manufacturers of chip-making equipment. Customers include Intel, TSMC, Samsung, SKHynix and Micron. https://www.asml.com/en/news/stories/2021/semiconductor-manufacturing-process-steps

Jan Bakker

Interim CEO at Shortcontentsolver

9 个月

Such an exciting intersection between technology and innovation! Can't wait to read more about it. ??

回复

要查看或添加评论,请登录

社区洞察

其他会员也浏览了