eMemory Technology Inc. (3529) Earnings Call Transcript & Summary

Good afternoon, and welcome to eMemory's Second Quarter 2023 Webcast Investor Conference. Today, we have our Chairman, Dr. Charles Hsu; President, Mr. Michael Ho; Head of IR, Ms. Li-Jeng Chen; and the Director of the Finance Department, Ms. Teresa Kuo. The format of today's event will be as follows. First, eMemory's Chairman, Dr. Charles Hsu will give an opening remark. Afterwards, President, Mr. Michael Ho, will summarize our operations in the second quarter of 2023, followed by our business outlook. Next, Dr. Charles Hsu will give a talk about securities forefront, Confidential Computing, which is something you guys might be familiar with. Then we will go into the Q&A session where our management team will be ready to answer your questions. [Operator Instructions] As a reminder, this conference is being recorded, and a webcast replay will be available within 3 hours after the conference is finished. Please visit the website at www.ememory.com.tw under the Investor Relations section. As usual, before we begin, we would like to remind everyone that today's presentation may contain forward-looking statements subject to the risk factors associated with the semiconductor and IP business. Please refer to the cautionary statement on Page 3 of today's presentation. Now I would like to give the floor over to eMemory's Chairman, Dr. Charles Hsu.

Good afternoon, everyone. Thank you for attending our conference call today. We expect our revenue to grow sequentially in the second half of the year due to strong licensing and accumulated new tape-out entering into the production stage. There is a very strong demand for our 5-nanometer and 3-nanometer solutions, especially from mobile data center, AI and autonomous driving. We are very confident in our future growth for many years to come. Next, I'd like to invite our President, Michael Ho to share our second quarter performance and the future outlook. Michael?

Good afternoon, everyone. Now let's begin with our 2023 Second Quarter financial results. The second quarter revenue was TWD 697 million up 4.3% sequentially but down 12.5% year-over-year. Operating expenses were TWD 328 million, up 9% sequentially and mainly attributable to the increase in bonuses and rewards. Operating income was TWD 369 million with an increase of 0.5% sequentially but decreasing 19.9% year-over-year. The operating margin decreased by 2.1 percentage points sequentially and decreased by 4.9 percent points year-over-year to 52.9%. EPS for the quarter was TWD 4.71 and ROE was 53.5%. Now let's move on to the revenue contribution by licensing and royalty. Licensing in the second quarter accounted for 35.8% of the total revenue up 74.6% sequentially and 24.4% year-over-year. Royalties in the second quarter contributed 64.2% of the total revenue decreasing 14.8% sequentially and 25% year-over-year. Total revenue for the second quarter increased by 4.3% compared to the previous quarter but decreased by 12.5% compared to the previous year. With that, I will comment on our revenue contribution by super specific IPs. NeoBit accounted for 22.7% of total licensing revenue in the second quarter, increasing 101% sequentially and up 27.6% year-over-year. Its royalties accounted for 30.5% of total royalty, down 14.6% sequentially and down 42.8% year-over-year. NeoFuse accounted for 45.4% of the total licensing revenue in the second quarter, up 60% sequentially and up 20.5% year-over-year. In terms of total revenue, royalty revenue, NeoFuse royalties decreased by 15.3% sequentially and 10.2% year-over-year, accounting for 67.5% of total royalties. PUF-Based security IPs contributed to 11% of the licensing revenue, increasing 66.3% sequentially but decreasing 41.4% year-over-year. Its royalties accounted for less than 1% of total royalties, up 123.3% compared to the previous quarter but down 18.8% compared to the previous year. For MTP technology, accounted for 20.9% of total licensing revenue up 89.9% sequentially and up 239% year-over-year. Royalty from MTP decreased 1.6% sequentially and 58.4% year-over-year, accounting for 1.9% of total royalties. Now let's look at the royalties for 8-inch and 12-inch wafers. For 8-inch wafers accounted for 44.5% of royalties, down 19.2% sequentially and 34.5% year-over-year. 12-inch wafers contributed 55.5% of royalties, decreasing 10.9% sequentially and decreasing 15% year-over-year. In total, 147 product tape-outs were completed in the second quarter. We will provide more information in the management report. In the next section, I will address our future outlook. We expect our revenue to grow sequentially in the year's second half. For licensing revenues, we expect the licensing revenue to increase significantly compared to the first half of the year. For the royalty revenues, as new applications gradually entered the mass production stage. We expect royalties to grow quarter after quarter for the rest of the year. Moving on to the new IP technology and business development. One, the demand for 5-nanometer design licensing is very strong. We have customers adoption for autonomous driving data centers and AI-related this year; two, we are developing 3-nanometer in several foundries with many customer requests and continue cooperating with CPU partners for 3-nanometer confidential computing; three, 22-nanometer emerging memory, MRAM and ReRAM completed verification with customers' design ins. This concludes my comments. Next, I will pass the time to Charles. Thank you.

Okay. So in this meeting, I'd like to again, [indiscernible] about the confidential computing, particularly on the GPU, since last summer, I have presented the confidential computing in CPU, and because GPU now is so hard for the AI applications. And we do see that NVIDIA has promoted their GPU as confidential computing, so I would like to talk about that. So let me turn to the page 13, okay. So if you remember my previous talk, I mentioned confidential computing, which has become even more relevant than two years ago. Today, I will share with you a more detailed version of how it works, what NVIDIA is using confidential computing for. And most importantly, how exactly a PUF-based confidential computing method difference -- is different from others and why it is more secure. In many applications, computers deal with sensitive data or face certain data regulations, such as health and financial services. And the integrity and the confidentiality of this data can be compromised by an authority or unauthorized for this, this is why we need to establish confidential computing. And otherwise, known that -- known as protection of data in use by performing the computation in a hardware-based trusted execution environment, okay? So you can see from the Page 13, I have illustrated three protection of the data which is including the data and the rest and the data in transit and the data in use. And the protection of data in use is mainly by -- used by the confidential computing. And in the confidential computing a trusted execution environment is deployed to provide an isolated environment separated from the main operating system, where data can be processed without exposure to the rest of the system. Okay. So please turn to Page 14. So to understand how confidential computing works please look at the left graph. For example, for CPU has the figure of the CPU has secured and non-secured zone application and OS on regular CPU cannot assess the secure zone. However, with the confidential computing sensitive data from the APPs and OS can be stored in the secured zone under individual -- individually isolated TEE on the CPU. Please turn to Page 15. And another example of confidential computing is in a graphic processing unit, GPU. GPU has its own hardware rule of trust to enable an isolated TEE for protecting the integrity and the confidentiality of data and the code during the transfer from CPU to GPU and also processing in the GPU. Based on the NVIDIA's H100 [indiscernible] it shows that NVIDIA use confidential computing for their AI training, as you can see in this viewgraph during the code and the data transferred from CPU to GPU in multiple individual TEEs are created so the GPU's hardware rule of trust, so that unauthorized entities are prevented from accessing the AI application data and the code. Therefore, the TEE, trusted execution environment, okay, needs to have a robust and hardware-based security which we call it the hardware rule of trust to ensure the following actions. One is to do the device authentication. For example, in this viewgraph, there are two parts. One is on the top is CPU and many users use CPU and on the bottom is GPU okay? So when the communication begin all the processing begin then the CPU has to start to communicate with GPU. So first of all, is the device authentication need to be done between CPU and the GPU? Okay, so this -- that is when the user and the device are communicating with the GPU, they have to confirm if the firmware and the software of the GPU are secured and the GPU also needs to authenticate the firmware and the software from OS or APPs to check that they have not been tampered with. And the second function is to guarantee data integrity and the confidential entity between the CPU and GPU okay? A better way to understand this is that the data being transferred will be encrypted and decrypt with encryption key exclusively to CPU and the GPU and the data transmitted from the CPU and the GPU will be protected, okay? So this viewgraph demonstrate the security confidential computing between the CPU and the GPU. And please turn to the next page. Okay. So we know that in order to set up the -- in order to set up the confidential computing environment, then we have to know that how to use -- identify since we have to set up the extent trusted execution environment, we have to identify that the trusted execution environment is basically needs to be accomplished by the hardware rule of trust. So from the Page 16, the hardware rule of trust in TEE can help protect the integrity and the confidentiality of the data and the code. It consists of key storage, key generation and the key protections and the cryptographic engines, all combined for the competent comprehensive securities. So let's look at this viewgraph. So our PUF-based crypto processor, we call it PUFcc can fulfill the role of all functions mentioned above. Our secure OTP consisting of both our Anti-Fuse technologies and PUF-based technologies creates secure key storage that hides the key, and there we hope we are above from the potential attackers. Furthermore, our PUF-based TRNG can efficiently generate truly random keys to protect the data and the code. So from this viewgraph, you can see that for the hardware rule of trust basically consists of two parts. One is the key management and the one is cryptography. So in the key management, key management also consists of key storage and key generations. And for key storage we have OTP and we have also PUF plus OTP to create secure OTP and also the actual secure key storages. And for the key generations, and we use the PUF-based TRNG to create high speed and high entropy key generations. And I will focus on talking about the key generations and the key storage since our fundamental technology have achieved a very high scale, key storage and the key generations by comparing to the conventional methods. So our PUF-based OTP helps the industry migrate to a very high-level security regime by improving the security of key storage and key generations. In the conventional key storage by using eFuse which is basically has not very secure because of the following two reasons: First of all, the burning of the -- because eFuse is by burning of the fuse. So eFuse require the fuse burning to determine the key whether it is 1 or 0, okay? So please refer to the Page 17. And you can see the low left side, you can see that eFuse, conventional eFuse lead to when the data is original zero and they need to be -- the fields need to be burned and become one. And so when the fuse is burned [ it was broken ]. So -- but this method will cause -- the eFuse can be easy to reverse engineer by looking at which fuse was burned. And there's also the risk of the fuse going back because the thin piece of metal can sometimes reattach itself, which causes the key to disappear. And the second problem by using eFuse is because eFuse -- the size of eFuse is very large. So it cannot overcome the density over [indiscernible] and due to -- because if the density becomes larger and the [indiscernible] will become lower. So the -- as a technology migrate to more advanced nodes, eFuse is inefficient -- insufficient to provide the key storage and also other security functions. So with our Anti-Fuse OTP, we call it NeoFuse. We will not face any issues with reverse engineering and also we will not face the size limitation and even key experience. So we use an exact term because we use [ electron tunneling ] method at the transistor level, which can shrink with the process node and they cannot be seen under [ Microscope ] okay? So combining our OTP with our PUF-based technologies makes key storage even more secure when used with PUFs unique value different from chip-to-chip, the key address can be scrambled and the store in a unique randomized location on the secure OTP to make the key storage ultra-secure. Okay. So let's turn to the Page 18. After I talk about the key storage, the next, I'm going to talk about the PUF-based key generation. As for key generation, the conventional True Random Number Generator used for hardware rule of trust is not PUF-based. My comparing PUF-based True Random Number Generator versus conventional True Random Number Generation in the 3 categories, PUF-based is superior for the following 3 reasons. First of all, the PUF-base has higher entropy okay? PUF-based TRNG has a much better and a much higher entropy because it is derived from the PUFcc, which is a perfect random sources, make it truly random. Conventional TRNG uses digital circuit to generate random number, which has had a low entropy. For example, our entropy is actually about 100 times than the conventional TRNG. And the second advantage of PUF-based TRNG is high speed. Another critical feature of TRNG is speed. For PUF-based TRNG, the speed can generate a high volume of keys, which is about 100x faster than conventional TRNG. So because the conventional TRNG require a lot more post processing, which will [ slow ] the speed. And the third advantage is about the power consumption in our PUF-based TRNG, we have less power consumption. So the power consumption with PUF-based TRNG is also around 100 times lower. Again, this has to do with the post processing that conventional TRNG requests. So that we're leading to more cost and energy use. Okay. So after -- compared to the advantage of our PUF-based key generation and the PUF-based TRNG next. Now, I'd like to make the summaries, so for the confidential computing in the future, it is a must for GPU and the CPU application, because it need to -- the confidential computing needs to provide the integrity and the confidentiality of the data and the code. And as compared to conventional rule of trust in memory is a rule of -- PUF-based rule -- how well rule of trust provide the best quality of unique identity and secure key storage and a much higher speed key generation for CPU, GPU and the DPU to facilitate their confidential computing. And all this, the difference is mentioned above why confidential computing architectures are moving away from conventional solutions toward PUF-based. For example, is putting our solution as a reference design for its V9 confidential computing with all developments reaching the most advanced process nodes. We expect increase in adoption of our security solution in the future. And this concludes my remarks. Next, we will enter the Q&A section.

Thank you, Charles. This concludes our prepared statement. We will now begin the Q&A session. [Operator Instructions] I will hand the mic over to Li-Jeng.

[Foreign Language]

[Foreign Language]

[Interpreted] The question that was asked was what's your view on the operating margin for the second half of the year. Michael?

It will be a few percentage higher than the first half of the year. Thank you.

[Foreign language]

[Foreign Language]

[Interpreted] The question was, how will the post-quantum encryption algorithm affect the company? Charles?

Encryption uses crypto algorithms to protect sensitive information, including a secure website or even e-mails. Now why do they use the public key encryption system rely on algorithm typical -- rely on algorithms difficult for even the fastest conventional computers to solve in limited time. However, quantum computers will strengthen existing public key encryption within a decade. The National Institute offer standards and technology and NIST is expected to announce new algorithms for encryption in 2024. The main problem with the potential post-quantum encryption algorithm is that the key size is very low, which may be thousands of times longer than current RSA and the ECC keys. The length of eMemory's PUF-based rule of trust lies in its flexible and high-speed key generation which is currently the world's easiest, fastest and the safest method it will be necessary for customers to adopt PUF-based hardware rule of trust to face post-quantum crypto algorithm. The last large scale, replacement of encryption algorithm was around year 2000, at the that time, the United States decided to adopt the AES algorithm. It took 10 years for major related companies to switch to the new encryption methods. So we expect that changing to a new encryption algorithm will also take around 10 years. This is very positive for speeding up customers' adoption of our solutions.

[Foreign Language]

[Foreign Language]

[Interpreted] For power management, I see the market is quite sluggish. Coupled with [ TI's ] price competition, is a [ TI ] customer of the company? And if not, will [ TI's ] price competition grab orders from existing PMIC company customers and affect the company's future royalty income. Michael?

TI is not currently our customers. We have many PMIC customers. Still, the main royalty contribution comes from like a dialogue and nontraditional PMIC companies, mainly from processors or system companies. This type of customers annual wafer production volume is around 700,000 to 800,000 pieces. They have replaced some of the chips of the traditional [ IDN ] like TI, TI's pricing competition will not affect such companies because they are all top-tier customers of foundries with the low list foundry cost. Even when foundry raised the wafer price in the past years, it didn't affect theirs. Their PMIC products are mainly used in-house or bundled sales with processors. Their royalty was affected primarily by the end market demand and the inventory adjustments and such customer situation differs. The first customers to adjust inventory have already started placing wafer orders, the rest will gradually return to normal production level next year. We believe that PMICs royalties will grow strongly next year. In addition to the low year over year base comparison. New customer's gains like Korean smartphone penetrating into EV cars, data center wearable through the key customers as well as the ASP enhancements through the customer migrating into more leading process node. Our cost U.S. customer PMIC migrating from 8-inch to 12-inch will own the PMIC royalty growth strongly next year.

[Foreign Language]

[Foreign Language]

[Interpreted] The question was, do you have any AI processor customers, Charles?

The definition of AI is very broad. I think that most people care about now is the large-scale language training models made by NVIDIA and AMD, both of which has made contact with us in the past. NVIDIA started reaching out in 2020, and we have been discussing how to implement our OTP and the PUF into their security architecture. The reason for not been adaptive was that our 4- and the 5-nanometer at the that time, did not complete the qualification yet. For the key storage and the well rule of trust they currently use eFuse. As I mentioned earlier, eFuse is not safe for key storage OTP and the PUF and the OTP or PUF plus OTP must be used to provide the secure key storage. For key generation of hardware rule of trust, they currently use the traditional or conventional to run the number generator with our using PUF. So that the entropy and the speed of key generation is integral to what we provide in PUF-based TRNG. In fact, our randomness is about 100x then theirs. They have also have consumed the 100x of power for the key generations especially in the future, the length of the key required by post-quantum algorithm will be greatly increased, which will also increase the calculation time having high-speed key generation such as ours is very advantageous. We are very confident that with the progress of our 3-nanometer and the more production records, we have great opportunities in the future. In addition to these two major chip supplier, we also have cloud customers doing in-house AI processor in China and the U.S. this year. Furthermore, if AI in a broad sense also covers the edge computing, then we already have many customers. Our solution is only a small area and can generate a basic security function in a simple way which is very competitive in edge computing.

[Foreign Language]

[Foreign Language]

[Interpreted] So this is also a question related to AI, which AI server-related applications have adopted the company's IPs and does this include CPU and GPU, Michael?

In the application inside the AI server adopting our IPs with step-outs or in production our SSD controllers, [ CXL ] memory controllers, retimers, PMIC and SPD Hub for DDR5-themed [ car ]. The CPU and the GPU has not yet adopted our IP, but we have a good chance when we complete the 3-nanometer OTP verification, especially for those which adopt ARM solution.

[Foreign language]

[Foreign Language]

[Interpreted] This question is also related to AI. So you mentioned that AI-related applications are amongst the few early adopters of PUF, which is the case if you attended our meetings. And why are they the first to adopt PUF? Charles?

Okay. In advanced AI model are typically stored in conventional commodity fresh memory and this standard storage device has no security function. So it can be stolen or modified or PUF-based will security rule of trust IP and crypto coprocessor can encrypt and authenticate these AI models to protect them from being stolen or modified, which is why customers need the PUF security and eMemory's solutions. In addition, all processors have SRAM. One type of SRAM is one type of volatile memory just like [ DRAM ]. So when SRAM is manufactured, it will have a [indiscernible] that need OTP for memory repair. In the past, eFuse was user to do SRAM repair functions. However, the SRAM density in AI chips is much bigger, which eFuse cannot fulfill up due to the density limitation. Our OTP offers 500 to 1,000x bigger density than eFuse, therefore, for AI with large density SRAM our IP become a necessity based on SRAM repair and the security requirements, most AI chips will convert from eFuse to our OTP.

[Foreign Language]

[Foreign Language]

[Interpreted] So many mature processes are located in China. And in the past, even ARM faced problems of collecting royalties from Chinese customers. If customers all move their mature products to Chinese foundries will eMemory face trouble receiving royalties, Michael?

Foundries, customers and over 900 chip customers in China, accounting for about 10% to 15% of our total revenue. There have been no problems with collecting royalties and other payments. This is related to the nature of IPs. ARM offers [ soft ] IP, providing customers RTL is similar to software for chip design. They collect the license and royalties from chip companies. They must audit their customers if customers use the same RTL to other chips without being licensed. We licensed the fundamental transistor to the foundries. The foundry paid royalties based on the royalty rate and wafer volume while the chip companies pay design license fee only, since there are only 9 foundries by more than 900 chip customers in China is much easier to collect and audit the accounts of foundries than the chip companies. In addition, we have continuously licensed our technology to foundries for new process nodes and new applications and the continuing tape-outs through the end customers' adoption. The contract will be breached if the foundry fails to pay the royalties as a result. All customers tape-out and the pipeline cannot go for production. The loss to the foundry will be much bigger than not paying royalties, not to mention the infringement of lawsuit penalty. The risk of embedded nonvolatile memory is that there is a problem with the memory reading after production. The entire chip or even the system will shut down. And the resulting penalty for damage will be much bigger. So customers don't want to risk illegally obtaining the IPs. Thank you.

[Foreign Language]

[Foreign Language]

[Interpreted] Why has MTP's license fee increased in quarter 2? And what applications have adopted our MTP technology, Michael?

The utilization rate of foundries is very low. There are incentives for foundries to develop the specialty technologies, especially embedded non-volatile memories to increase wafer value with the requirement of high-speed, low-power and [ smartphone ] factors, replacing external memory with embedded memory is accelerating. Our MTP technology is compatible with the larger process. So the foundry can easily expand the [indiscernible] large process for embedded nonvolatile memory without investing in additional equipment with [ fabs ] low utilization rate, foundries are allocating more resources to speed up the technology development. We believe our MTP related technology have started to enter a multiyear growth cycle and will have a significant royalty growth next year.

[Foreign Language]

[Foreign Language]

[Interpreted] So this person asked, I heard that Synopsys, your competitor, bundles OTP, PUF and EDA together and even offered it as a free IP. Will this affect eMemory, Michael?

[ This ] acquired our previous main competitors, [indiscernible] and Kilopass for around 5 years. In the past, we often heard clients say that they bundle EDA tools to sell together. And the OTP price was very low or even free to get our customers to switch. However, no customers have ever switched because of this, and we haven't changed our pricing strategy either accurately more and more customers using Synopsys says has switched back to our IPs. The latest news is that Synopsys raised their price because previous price gearing strategy was ineffective, PUF is even more fundamental different. Synopsys used SRAM PUF and the performance is comparable to our new PUF IP. Synopsys security department is composed of many companies acquired through the M&A. Their strategy is focused on the security aspect of system integration, which differs from our focus on offering the best rule of trust in the world. Thank you.

[Foreign Language]

[Foreign Language]

[Interpreted] So I see someone asked this question. And the question is, when will you start seeing contributions from your collaboration with ARM, Michael?

With ARM has been ongoing. Some customers have already tape-out our security solution through ARM's preform and contribute a license fee already. Thank you.

[Foreign Language]

[Foreign Language]

[Interpreted] Does eMemory do 12-nanometer ISP? In addition, if the OTP of the ISP rises from 2,000 to 16,000, will the output value of royalty be multiplied?

Consumers have adopted our IP in 12-nanometer ISPs with small volume production. Since the price of 12-nanometer wafer is about 40% higher than that of 22- and 28- nanometer, if it's [ 16- ], 22-, 28-nanometer customers reached to 12-nanometer, chip size remains the same, then the royalty difference will be the increase of wafer pricing, which has nothing to do with how much density is used. Thank you.

[Foreign Language]

[Foreign Language]

[Interpreted] [ CXL ] framework, what role does eMemory IPs play?

CXL is the protocol to enhance the memory capacity sharing between CPU and the other devices such as GPU, and the DPU or CXL memory expander controller customers adopt our rule of trust to protect the security of the data confidentiality and the integrity during the data transfer.

[Foreign Language]

[Foreign Language]

[Interpreted] Other than PMIC, do other parts of DDR5 also need to use your IPs. For example, register clock driver, data buffer and SPD Hub, et cetera. Michael?

DDR are not currently used, but SPD Hub have adopted our NeoEE for the 100,000 endurance. So inside the DDR5-themed [ car ], there are two chips: One is the SPD Hub and the other for the PMIC will use our MTB solution. Thank you.

[Foreign Language]

[Foreign Language]

[Interpreted] eFuse cannot shrink below 28-nanometers. However, in terms of advanced manufacturing processes, if the customer's eFuse is actually accidentally blown out what kind of solution will be adopted if the eMemory solution has not yet been implemented, Michael?

Use the eFuse and [indiscernible]. It will usually increase the eFuse area. However, it will take up too much space and it would be very expensive for advanced process chips. In addition, eFuse cannot accommodate a circuit design that needs to burn in a large current. However, since the eFuse is provided by the foundry, if there is any problem, the foundry will be responsible. So eFuse is still dominating the advanced manufacturing process. Another type of technology is Anti-Fuse many, NeoFuse and the competitors' solution. However, the IP of our competitors have experienced a problem in the advanced manufacturing process. At present, no heard about any customer adopt their solution in leading process, but it also affects the customers' confidence in the slow down the adoption of this technology with more production records and successfully moving into the 5-nanometer and the 3-nanometer, we are confident to replacing eFuse. It's only a matter of time. Thank you.

Due to time constraints, this will be our last question. So Li-Jeng?

[Foreign Language]

[Foreign Language]

[Interpreted] The last question is, why is your royalty ratio of revenue much higher than other IP companies? And why has it been like this for such a long time, Michael?

IP companies mainly collect the license fees, but without a little royalties. Our strategy is to charge less on the licensing in the early stage but insist on collecting the royalties. Thank you.

Thank you. We will now begin the closing comments. Charles, can you please proceed?

Both conference today. And for more information about our PUF-based security IP and the technology. We encourage you to visit our PUF security website at www.pufsecurity.com and check out our articles and other materials. And thank you again for your patience and support for eMemory. We will continue to work hard on our technology and IP innovation and also PUF-based hardware security solutions for our customer and bring higher returns to our shareholders. Thank you.

Thank you, everyone, for joining us today. You may now disconnect. Goodbye, and have a good day. [Portions of this transcript that are marked [Interpreted] were spoken by an interpreter present on the live call.]