Latest news with #Phison
New Straits Times
06-07-2025
- Business
- New Straits Times
More than a trade fair: Taiwan Expo strengthens Malaysia's innovation agenda
THE buzz may have died down at the Kuala Lumpur Convention Centre, but the impact of the recently concluded Taiwan Expo 2025 continues to reverberate strongly across industries and conversations. Making its largest showing yet in Malaysia, the three-day event from June 23 to 25 wasn't just a showcase of Taiwan's technological prowess, it was a vibrant, forward-looking celebration of innovation, partnership and cultural exchange. At the heart of this 8th edition was one undeniable force: artificial intelligence. From AI-powered lung cancer screening and smart cybersecurity systems to responsive mechanical skin and autonomous guided tour robots, Taiwan brought cutting-edge breakthroughs to the Malaysian stage. A total of 32 leading AI companies took part, offering real-life applications that sparked the imagination of industry insiders and curious visitors alike. Highlights included Phison's next-gen AI storage solutions, Touché Solutions' tactile tech and DeepXray's medical imaging breakthroughs. Even education and finance sectors were not left out — ThinkCloud and NUWA Robotics presented AI-driven tools that could redefine learning and digital security. "This is more than just an expo. It is a gateway into the future. Taiwan's innovation ecosystem is thriving, and we are proud to share it with our Malaysian partners," said Simon Wang, president & chief executive officer of TAITRA, during the opening ceremony. A TASTE OF TAIWAN'S SOUL But it wasn't all circuits and algorithms. The newly introduced Taiwan Select Food Culture Pavilion offered a delectable journey into Taiwan's rich food heritage, pairing award-winning halal-certified snacks and clean-label products with the delicate ritual of a traditional tea ceremony. The sensory appeal of Taiwan's soft power was on full display — from smart lifestyle gadgets to culturally rich artisanal products. The diversity of exhibits reflected Taiwan's wide-ranging strengths: sustainability solutions like smart water towers and energy-saving tech stood alongside beautifully crafted home innovations, showing how thoughtful design and technology can transform everyday living. The Taiwan Expo wasn't just about displaying innovation; it was about building bridges. Supported by 26 Malaysian industry associations, the event underscored Taiwan's position as a valuable trading partner. With bilateral trade reaching USD 39.1 billion in 2024 and showing a 36.8 per cent growth in early 2025, the synergy between Malaysia's manufacturing might and Taiwan's tech expertise was undeniable. "Malaysia and Taiwan are natural partners. Our industries complement each other, and events like this foster the kind of collaborations that drive economic progress," said Ng Sze Han, Selangor State EXCO for Investment, Trade and Mobility. Federation of Malaysian Manufacturers (FMM) president Tan Sri Soh Thian Lai also praised the expo's timely relevance, saying: "What we are seeing here aligns with Malaysia's future direction — smart manufacturing, AI, green tech — it's all in sync with our industrial ambitions." HEARTWARMING HIGH-FIVE The crowd erupted in cheers on the final day as Taiwan Excellence's lovable mascot, FuBear, made history, setting a new Malaysia Book of Records with an astonishing 186 high-fives in just one minute, as proudly announced by Techtrans Taiwan. More than just a gimmick, the effort was a touching collaboration with First City University College, REACH Segamat (a centre for children with autism), and PSJKM, proving that even a tech-heavy event can carry emotional warmth and community spirit. THE LEGACY LIVES ON Though the booths have been packed away and the crowds dispersed, the Taiwan Expo 2025 has left an indelible mark. It wasn't just an exhibition — it was a statement: of shared futures, mutual respect, and a commitment to innovation that enhances life on every level. Malaysia may have said goodbye to the expo — for now — but the connections made, technologies introduced and friendships formed are already paving the way for what's next.
Forbes
24-06-2025
- Business
- Forbes
Kioxia, StorOne, Phison And Marvell Deliver Storage And Memory For AI
AI Data Growth Artificial Intelligence training and inference are a major factor in increasing digital storage and memory demand growth. According to a recent announcement from the Dell'Oro Group, the AI expansion cycle led to a 62% year over year growth in Q1 2025 in telecommunications, security, networks and data center industries, including memory and storage. Let's look at some recent storage announcements to support AI workloads from Kioxia, StorOne and Phison and Marvell Technologies. Kioxia recently announced its single port CD9P Series PCIe 5.0 NVMe 2.0 SSDs built with the company's 8thgeneration BiCS Flash TLC-based 3D flash memories (with 330 layers--?). These products also incorporate the company's CMOS directly Bonded to Array technology to boost power efficiency, performance, and storage density. Images of the product family are shown below. Kioxia CD9P SSD Product Family According to Kioxia, the CD9P Series is purpose-built for these next-generation environments, delivering the speed and responsiveness required by AI, machine learning, and high-performance computing (HPC) workloads to ensure GPUs stay fed with data and operating at maximum efficiency. The product is available in EDSFF E3.S form factor with 1.6-30.72 TB capacities and 2.5-inch U.2 form factor with 1.6-61.44TB capacities. The CD9P Series 15.36 TB model delivers approximately 60% and 45% improvements in sequential read and write speeds per watt, respectively, compared to the previous generation Kioxia SSD. It also achieves gains of approximately 55% and 100% in random read and write performance per watt, measured in thousands of IOPS (KIOPS), respectively. The product is being sampled to select customers and was on display at the on-going HPE Discover Las Vegas event. The table below gives more details on this model's performance with different workflows. CD9P Performance versus Workloads StorOne unveiled its ONEai, Automated AI Solution Optimized for Enterprise Data Storage in partnership with SSD company, Phison Electronics. StorONE integrated Phison's aiDAPTIV+ AI capabilities into the StorONE enterprise storage system to accelerate AI deployment and deliver domain specific responses on the stored data for end users. The Prompt: Get the week's biggest AI news on the buzziest companies and boldest breakthroughs, in your Subscribed! You're Subscribed! According to StorOne, ONEai leverages AI GPU and memory optimization, intelligent data placement and built-in support for LLM inferencing and fine-tuning directly within the storage framework, offering an efficient, AI-integrated system with minimal setup complexity. With ONEai, users benefit from reduced power, operational and hardware costs, enhanced GPU performance and on-premises LLM training and inferencing on proprietary organizational data. Marvell Technology, Inc. announced that it has expanded its custom technology platform with the launch of a 2nm custom Static Random Access Memory (SRAM), designed to boost the performance of custom XPUs and devices powering cloud data centers and AI clusters. It also introduced its Package Integrated Voltage Regulator (PIVR) power solutions, which it says enable hyperscalers to enhance the performance, efficiency and return on investment (ROI) of AI and cloud infrastructure. This custom SRAM delivers up to 6 gigabits of high-speed memory. Marvell said that it had previously introduced its CXL technology for integration into custom silicon to add terabytes of memory and supplemental compute capacity to cloud servers and unveiled custom HBM technology that increases memory capacity by up to 33% while reducing the space and power required for dense high-bandwidth memory (HBM) stacks inside XPUs. AI drives storage and memory demand. Kioxia, StorOne, Phison and Marvell Technologies announced digital storage and memory applications to support diverse AI workloads.
Techday NZ
20-06-2025
- Business
- Techday NZ
StorONE launches ONEai for on-premises AI training & analysis
StorONE has introduced ONEai, an enterprise-focused artificial intelligence solution designed to enable large language model (LLM) training and inferencing directly within the storage layer. Developed in partnership with Phison Electronics, ONEai integrates Phison's aiDAPTIV+ technology directly into StorONE's storage platform. This arrangement means companies can carry out AI-related operations, including domain-specific model training and inferencing, on data stored locally—without the need for external AI infrastructure or cloud services. Technology integration ONEai offers an enterprise storage system with embedded AI processing capabilities. By leveraging GPU and memory optimisation, intelligent data placement, and direct support for LLM fine-tuning, the new solution aims to streamline AI deployment and improve access to proprietary data analytics. The system is positioned as a turnkey, plug-and-play deployment that does not require significant internal AI expertise or complex infrastructure. The solution's architecture is focused on reducing hardware costs, improving GPU performance, and offering on-premises LLM training and inferencing. This is intended to support organisations looking to gain deeper insights from their stored data while controlling costs and maintaining data sovereignty. Addressing enterprise challenges With more enterprises seeking to leverage AI on multi-terabyte and petabyte-scale data pools, the traditional requirement for separate, often complex AI infrastructure has been a significant barrier. Conventional methods often depend on external orchestration and cloud or hybrid AI workflows, which can increase both regulatory risks and total costs for data-driven organisations. StorONE and Phison have developed ONEai to deliver fully automated, AI-native LLM training and inference directly within the storage layer itself. The product supports real-time insights into file creation, modification, and deletion, and is optimised for fine-tuning, retrieval-augmented generation (RAG), and inferencing tasks. The system includes integrated GPU memory extensions, offering a user interface aimed at simplifying ongoing management. End-to-end automation "ONEai sets a new benchmark for an increasingly AI-integrated industry, where storage is the launchpad to take data from a static component to a dynamic application," said Gal Naor, CEO of StorONE. "Through this technology partnership with Phison, we are filling the gap between traditional storage and AI infrastructure by delivering a turnkey, automated solution that simplifies AI data insights for organizations with limited budgets or expertise. We're lowering the barrier to entry to enable enterprises of all sizes to tap into AI-driven intelligence without the requirement of building large-scale AI environments or sending data to the cloud." ONEai is presented as a system that automatically recognises and responds to changes in stored data, supporting immediate, ongoing AI analysis. Its plug-and-play approach is designed to remove the requirement for separate AI platforms and deliver full on-premises processing to ensure maximum data control. Phison's technical contribution "We're proud to partner with StorONE to enable a first-of-its-kind solution that addresses challenges in access to expanded GPU memory, high-performance inferencing and larger capacity LLM training without the need for external infrastructure," said Michael Wu, GM and President of Phison US. "Through the aiDAPTIV+ integration, ONEai connects the storage engine and the AI acceleration layer, ensuring optimal data flow, intelligent workload orchestration and highly efficient GPU utilization. The result is an alternative to the DIY approach for IT and infrastructure teams, who can now opt for a pre-integrated, seamless, secure and efficient AI deployment within the enterprise infrastructure." According to the companies, ONEai's plug-and-play deployment model can eliminate the requirement for in-house AI expertise while streamlining overall operations. The integrated GPU modules inside the storage layer aim to lower AI inference latency and deliver up to 95% hardware utilisation, while also minimising power consumption and operational costs by reducing the number of GPUs required. Use cases and availability ONEai is designed for immediate interaction with proprietary data, automatically tracking data changes and feeding updates into ongoing AI training and inferencing processes. This is intended to align with real-world enterprise needs for rapid, domain-specific data analysis. The solution will become generally available in the third quarter of 2025.
Yahoo
15-06-2025
- Business
- Yahoo
SMI CEO says no PCIe 6.0 SSDs for PC 'until 2030', Nvidia demands SSDs with 100 million IOPS — Wallace C. Kou on the future of SSDs
When you buy through links on our articles, Future and its syndication partners may earn a commission. The consolidation of the storage industry in the 2010s has injected every major SSD supplier with an in-house team designing SSD controllers, leaving the market with only a few independent SSD controller makers. These three makers — Maxio, Phison, and Silicon Motion (SMI) — are thriving, and their businesses are growing rather rapidly. At Computex 2025, we sat down with Wallace C. Kou, the founder and chief executive officer of Silicon Motion. Silicon Motion is one of the leading independent developers of SSD controllers, and we discussed the company's strategy as well as industry trends. During the conversation, we discussed numerous topics, including PCIe 5.0 and PCIe 6.0 SSDs, how the market of SSDs has changed over the last 10 years, why vertically integrated 3D NAND manufacturers are poised to continue using controllers from third parties, the cost of chips produced on advanced process technologies, SMI's plans to use Samsung Foundry in addition to TSMC, AI storage, as well as the viability of PLC 3D NAND and other exotic types of non-volatile memory. Anton Shilov: Large makers of NAND and SSDs have acquired multiple SSD controller startups from 2015 to 2020. However, they not only continue to work with you, but they even use your controllers for many of their branded products. What is your secret? Or their weakness? Wallace C. Kou: What is our strategy in order to win? We make sure our roadmap direction is aligned with current and future market trends and NAND makers' plans [and capabilities]. Compared to 5–10 years ago, [the] NAND flash market is very different. Back then, investments for 2D NAND were smaller, but with 3D NAND, the capital expenditure (CapEx) has doubled. So, NAND makers must focus on areas where they can be profitable and grow. So, each of these have a different play. Looking at Samsung and SK hynix, they have quite different businesses than SanDisk because the latter does not have a DRAM business. We leverage what we are good at. Fundamentally, we support NAND makers who need to sell wafers (This involves selling uncut wafers to SSD makers, which then dice, test, and and package them themselves for use in their products), or enable their channel business. We also help module makers manage their inventory or utilize excess supply, building out a broader ecosystem. This gives us a much better opportunity to grow. Now, we engage not only with NAND technology for each generation, but also with internal projects. Whenever they have a need, they come to us for all sorts of support. I think even 10 years ago, we already saw that NAND makers couldn't do everything themselves. For certain products, they have to outsource to SMI, right? They keep the most important parts in-house, but for others, where they need more control, they work with us. Some NAND is great for mobile, some for client, some for enterprise, some for automotive — in certain areas they need help, and that is where we step in. That is how we continue and make sure we can capture opportunities whenever they arise. Third-party controllers are also a key part of the current solution. You can see it with all the NAND makers today, including YMTC. When they have a new-generation pre-sample, they come to SMI. We quickly test it and give them feedback. That is how we build relationships — from technology to personnel, from management to R&D, from planning to sales — across all departments. We are probably the only controller maker today that works with all six or seven NAND makers, depending on how you count. And we do it independently — collaborating not just with R&D, but also with executive teams, planning, marketing, sales, and procurement. Anton Shilov: Indeed, SMI supplies all types of controllers for all types of applications, starting from humble PCs and then going up to enterprise-grade controllers and even automotive-grade controllers. Surprisingly, you can address all the segments of the market. Which of them will grow faster, what kind of market share should we expect? Wallace C. Kou: I mean, it is a very natural trend for a company like ours to grow. We have to grow bigger because client SSD controller revenue alone — even a hundred million dollars in some areas — is not that much, right? So, moving into enterprise and automotive is a natural direction for us. And we see the muscle we have. I can give you a couple of examples. Last week, while I was in Boston attending the JP Morgan Investor Conference — the 53rd conference — it was overwhelming with investors. I had not been to that event for 15 years, but I sent a very clear message. Silicon Motion has four major categories: client consumer SSD controllers, mobile eMMC/UFS controllers, enterprise controllers, and automotive controller solutions. These are the four major product lines where we will continue to grow. Especially for enterprise and automotive, we will grow much faster because the base is still small. For client SSDs, we already have about 30% market share. We are confident we can grow to 40% within three years — maybe even two years, depending on how fast PCIe 5.0 drives become mainstream. With eight NAND channel PCIe 5.0 controller [the SM2508], this is the first time we have had a real opportunity to serve the top tier. In the past, we never had the chance. Now, among 3D NAND makers, we have won four of them, and nearly all of the module makers. So, when [the] high-end PCIe Gen5 [platform] starts to ramp [at module makers] in July, we will have more than 60% market share (This is a targeted projection). There are only two makers I cannot name, but they have their in-house solutions. Eventually, I believe they will also come to us — I cannot say 100%, but that is my belief. So, 60% is our goal. Even for four-channel [PCIe 5.0 SM2504XT controller] mainstream CPO (Client Platform OEM) designs, which will start ramping in the middle of next year, we expect to win minimum 50% market share. So if you do the calculation, as PCIe Gen5 adoption continues to grow, our blended market share will move up to 40% or even higher. Anton Shilov: So, is it safe to say that once you have the core IP technology, which can manage certain types NAND, you can actually leverage that core technology to different types of applications. Wallace C. Kou: So, our, our core [IP] is in our ASIC architecture, LDPC error correction, and efficiency. I have to make sure the ASIC area is sufficient but not too large. We have to carefully manage margins — keeping around 50% gross margin — to stay competitive against internal [in-house] controller solutions. Otherwise, customers would not choose our controllers. Anton Shilov: This is exactly what I am talking about. The moment you have the right implementation of the NANDExtend technology, you can use it for different types of applications. Wallace C. Kou: In certain areas, we can reuse the IP, but for some [new 3D NAND] generations, you still need innovation — that's key. Let me give you an example: Kingston (A shareholder of Phison, SMI's major rival). Kingston definitely uses our controllers. I think their CEO once said — people asked Kingston, 'Why don't you use SMI solution? [They] are cheaper, faster, better.' Then Kingston's David Sun [Kingston Technology's COO] said, 'We should really use Silicon Motion, because when I use SMIs controller, I can use Samsung NAND, I can use SK hynix NAND, I can use others — even YMTC.' So, they say we are the controller that supports multiple NAND sources, and that is very valuable for Kingston. That is why, even for companies like SK hynix or Samsung, sometimes it makes more sense not to develop their own controller [for an application]. Use mine, and sell [your] wafers faster and better by partnering with us. Anton Shilov: Now that you mentioned both Kingston, an SSD maker, and SK hynix, a NAND maker, what is the primary focus of the SSD business? OEMs, turnkey solutions? Wallace C. Kou: You asked a good question. I will tell you the trend. For module makers, the retail portion is getting smaller and smaller. OEM business is becoming bigger. So, we have to focus on OEMs now. For consumer SSDs, the OEM side is important — but that does not necessarily mean NAND makers are winning (with their own branded products). You should also watch the PC OEM business. It is starting to move away from NAND makers, especially in the value and mainstream segments. High-end still tends to stick with the NAND makers, but in the value line, more and more PC OEMs are accepting module makers. We already saw this trend starting last year. For example, Seagate now uses more module makers because their pricing is more competitive. Even Lenovo in China is using Union Memory for their domestic PCs — and 100% of Union Memory's controllers are from us. We see this trend continuing. NAND makers are trying to sell complete solutions to win back some market share, but the volume is still small. Meanwhile, we see Adata, Kingston, Longsys, and Biwin starting to engage with Acer, Asus, HP, and even with Dell. This is happening right now. That is why module makers are gaining more strength and position. The retail portion for module makers is going down, but their share in the PC OEM segment is getting big, an important part for everyone. Anton Shilov: But module makers must make those models more competitive in terms of pricing, right? Wallace C. Kou: Yes. NAND makers might be able to take 30% – 35% margins, but module makers can operate on much lower margins — even 5% or 10%, especially Chinese makers. That is why they see an opportunity. In the past, our main concern with module makers was always about quality. Price, they could always meet, but quality was not good enough. Now, with SMI controllers and firmware, over the past 10 years, they have built up their own test capabilities and improved their factories. Now, their quality can meet the standards required by Lenovo, Dell, and HP, which is very good. So, we see even more opportunities now. Module makers are more flexible and faster in meeting urgent demands from PC OEMs. Anton Shilov: How important is the OEM SSD market is to you? Wallace C. Kou: It is very, very important. We are very happy. Beyond just working with NAND makers and module makers, we now work directly with PC OEMs. For example, we have quarterly reviews with Lenovo — both in Europe and Shanghai — with their procurement teams. We also have quarterly reviews with Dell in Singapore to talk about roadmaps, firmware updates, security requirements, performance, power budgets, everything. We discuss their needs and how we can provide value. This makes us more and more suitable as a partner. Sometimes, we even know better than the NAND makers how to help provide a stronger solution. With HP, the meetings are a bit less frequent compared to Lenovo and Dell, but we still maintain regular contact. Anton Shilov: Your focus on OEMs is probably one of the reasons why Silicon Motion was late with enthusiast-grade PCIe 5.0 controllers compared to Phison. Wallace C. Kou: But you need to understand why. PC OEMs do not plan to launch [PCIe 5.0 SSDs] until the middle of this year, right? Phison, on the other hand, is focused on retail, so they wanted to create buzz early. But if you look at how many units they have actually sold so far — it is not that much. Anton Shilov: Alright, but what do you think about enthusiast-grade SSD market? Are drives for gaming an important product category for Silicon Motion? Wallace C. Kou: Gaming was not our major focus. But right now, for PCIe 5.0 eight NAND channel SSDs, we are the only ones offering a low-power, 6nm solution on the market. So, it has become very popular — the 'darling' — for gaming PCs, workstations, high-end notebooks, and even what is called white-box servers. By default, they are using our SM2508 controller — I do not even know why exactly — but that is the situation. That is also why we faced shortages in Q1. Customers were buying off the shelf. So, we are not particularly trying to penetrate the gaming market, but it just happened because we are the only available eight-channel solution with DRAM, high performance, low power, and total power consumption under 7W. Anton Shilov: Do you expect nearline storage to become a big focus for you over the next few years? Wallace C. Kou: Yes, because if you follow the announcements, Meta has already said they are moving to QLC 3D NAND, and you probably know who Meta's suppliers are. Also, Pure Storage announced a major hyperscaler design win. I do not want to comment specifically, but if you look at Meta, Microsoft, Google, Amazon — in the past, they only used 16 TB drives. There was no need for bigger density because they managed storage on the cloud side with their own software. But now, they all want to move to higher-density QLC — they want to go to 128 TB drives as soon as possible. It is about slot density — they need higher density for AI training and inference because that requires massive amounts of storage. That's why not only us but also Chinese companies want to replace part of the nearline HDD market. Nearline SSDs will eventually replace tape drives, too. So, I think both the HDD and SSD markets are going to grow rapidly over the next five years. Anton Shilov: Now that you have mentioned 6nm as a means to reduce power consumption, it is well known that SSDs and SSD controllers are getting hotter. Can you deal with that by adopting more advanced process technologies earlier? Wallace C. Kou: Yes. Our PCIe 6.0 enterprise controller with 16 NAND channels — which is in development, uses 4nm. You have been in the industry for a while, so you know the history and are familiar with the competitive landscape. Starting from PCIe 5.0, we have seen competition decrease because the requirements for signalling and design make it very difficult for startups to enter. For example, if you do not count R&D expenses and just focus on the IP and mask costs for PCIe Gen5 — including memory, controllers, and Arm cores — you probably need at least $16 to $20 million just for one tape-out at 6nm. If you make mistakes, the costs add up even more. For 4nm, the costs are even higher — about $30 to $40 million for one tape-out, again without including R&D expenses. It is a huge investment. That is why it is very hard for new entrants, and it makes us even more unique. In the past, we saw many startups, especially in China and the U.S.— maybe 20 to 30 players — trying to enter the enterprise SSD market. But now it is shrinking. I do not see any of them announcing PCIe Gen6 designs. I am not even sure if they can survive until the end of the year if they do not have enough cash. Anton Shilov: Well, PCIe 6.0 is hard both in terms of controller and in terms of Physical Layers (PHY). Good to hear that SMI is developing its own PCIe Gen6 solution for enterprise applications. When do you expect it to come to market? Wallace C. Kou: Late 2026 to 2027 — probably late 2026. I think our customers want to align with Nvidia's Rubin platform. But it is a tough schedule because PCIe Gen6 is much more complicated, and the industry — especially the media side — is trying to push storage to the next level, particularly compute storage, to align with Rubin. For traditional data storage, I do not think PCIe 6.0 is really needed until maybe late 2027 or 2028. But for compute storage, they want to try to launch it together with Rubin, since it connects GPU and SSDs for direct access. Anton Shilov: How are these extra costs expected to affect the pricing of PCIe 6.0 controllers for enterprise applications? Wallace C. Kou: I expect the price of enterprise PCIe Gen6 SSD controllers to be at least 25% to 30% higher compared to enterprise PCIe Gen5 SSD controllers. Anton Shilov: When would you expect consumer-grade PCIe 6.0 drives to arrive? Wallace C. Kou: For consumer? You will not see any PCIe Gen6 [solutions] until 2030. PC OEMs have very little interest in PCIe 6.0 right now — they do not even want to talk about it. AMD and Intel do not want to talk about it. That is why we feel very good about the situation — because we dominate PCIe 5.0, both 8-channel and 4-channel controllers. For the next four years, we will be in a comfortable position to continue growing in the client market. Anton Shilov: I have seen a demonstration of a PCIe 6.0 SSD at Computex. Wallace C. Kou: Micron tries to be the first with PCIe 6.0 SSDs, but I think Nvidia has to figure out how they want to use compute storage first. It takes some time for them to figure out how to support peer-to-peer communications from SSD to GPU right now. But the solution is too expensive. Anton Shilov: That solution is expensive because it uses a very specific switch, advanced retimers, and specific software. But if you need it today — for peer-to-peer SSD to GPU communications, you are going to pay for this. Wallace C. Kou: It is true, but compared to the Rubin platform, maybe it is a relatively not that big [in terms of volumes] anyway. Anton Shilov: You mentioned that with every new node, tape outs get more expensive, so you have to choose nodes that provide the right balance between price, yields, and power. Wallace C. Kou: Yes, for returns. You need to have the customer lined up, and you need to have a certain market demand. Otherwise, the investment will not deliver a return. Anton Shilov: Would you expect SSD controllers to get more expensive over time because of more expensive IP, photomasks, and production? Wallace C. Kou: From PCIe 5.0 to PCIe 6.0 controllers, — yes. But for PCIe 4.0 and PCIe 3.0, prices will still gradually decline because there is still competition, especially from some Chinese customers like Maxio and others. PCIe 4.0 controllers will stay in the market environment for at least three to five more years. So, we also have plans to reduce our costs — we are working on cost reduction strategies and will leverage Samsung Foundry to help with that. Anton Shilov: When it comes to Samsung Foundry, which nodes are we talking about? Wallace C. Kou: If I tell you everything, then everyone will know our plans! [laughing] But I can say we have certain technologies that fit perfectly for PCIe 4.0 and 5.0 SSDs. We also want to make sure TSMC does not have any concerns, because currently about 95% of our products are manufactured at TSMC. We use Samsung Foundry for specific reasons — either for Samsung-related projects or for certain cost reductions. If you are asking about process technologies. For TSMC, we are use 12nm, so when we want cost reduction with Samsung, we would choose 8nm for that level (possibly Samsung's 8LPP, which is closer to 10nm than to 7nm-class nodes). If we are on TSMC 6nm and want further cost reduction, we would move to Samsung's 5nm. That gives you a reference. But the big challenge is IP — how to get the IP quickly. Our internal resources are limited, and our teams are already fully focused on new-generation technologies like PCIe Gen6 and UFS 5.0. So, how we manage IP — that is an important secret between Silicon Motion and Samsung Foundry, and it is key to how we grow together. If there is a need to reduce the price, we will have a Samsung version ready. Anton Shilov: Okay, so you are talking about double-sourcing the same controllers? Wallace C. Kou: Not really, just for certain cost reductions. If I see my margins are under threat, I have to lower the costs. But you know, for NAND makers, once they have a design win [with a PC OEM], it is hard for them to change. They do not want to change unless there is a major generational shift, otherwise, it is just a name change. That is why we feel very comfortable. Module makers may need more competitive pricing, but Phison has a different business model than us. They try to sell their full solutions (a controller with NAND). [Inaudible], for example, is [their] only for mobile SSDs — and they can't really expand beyond that. Anton Shilov: Alright, so just to clarify — when you talk about Samsung, does it mean double-sourcing the same controller or developing a new controller to be produced exclusively at Samsung Foundry? Wallace C. Kou: New controller — it will be a new controller, because I have to redesign everything. I have to think about how to maintain the same performance while reducing costs — not just from wafer pricing, but also from design improvements. I need to make sure my product stays competitive, not only for module makers but also for some NAND makers like Micron or SanDisk. They do not focus on retail; they want lower pricing, so I have to offer them something. Anton Shilov: So, in the future, you are going to have more controllers — some produced by TSMC, others by Samsung? Wallace C. Kou: Well, PCIe 5.0 is just now ramping and not yet at full scale. You probably will not see that until maybe two years from now. Anton Shilov: Speaking of IPs, do you actually invent them in-house, or do you license them from third parties? Wallace C. Kou: Key IPs — we always want to develop in-house. But for certain IPs, like DDR controllers or I/O controllers, where we do not have deep expertise, I don't think it's necessary to insist on owning everything. For those, we use third-party IP. As for Samsung — there are certain secrets I cannot share with you. Anton Shilov: But what about PCIe IPs? Wallace C. Kou: All PCIe IPs are in-house today. All UFS IPs are also in-house. Anton Shilov: Apart from SSDs, which segments of the storage market do you expect to grow explosively over the next, say, three to five years? Wallace C. Kou: Enterprise is definitely our focus, not just our focus, but also the focus of NAND makers. Look at what Jensen Huang says — everyone tells you AI relies on three major technologies: compute, storage, and networking. We are already playing in storage, and in the future, we also want to play in connectivity. To really reach certain goals, storage and connectivity must come together to reduce latency, of course. If you ask me where future investments will go, I would say, we definitely need to develop certain networking and connectivity technologies. We are looking at integrating these into a single package — a single chip — to reduce latency and improve efficiency. This can solve some of the infrastructure problems we see today in AI systems. It is something similar to what Intel tried to do with Optane — but they did not really get it done properly. I also see Nvidia trying to build an ecosystem with BlueField. I do not know if you know, but we had a design win with Nvidia's BlueField-3, providing Boost storage. From BlueField, they connect through PCIe switches (likely coming from Broadcom) to 30+ SSDs in a storage bay or storage system. We are also trying to win storage bay and storage system projects with our MonTitan platform. For Boost storage, we already have a design win with Nvidia. Nvidia typically chooses two solutions — one is ours, and the other I cannot disclose, but it is another player that also uses SMI controllers and firmware, just under a different name. Anton Shilov: Do you plan to offer retimers? Wallace C. Kou: No, not retimers. Retimers are not in the picture. Boost Storage sits next to BlueField-3? BlueField is a networking accelerator — a data processing unit (DPU). It serves the storage network as an accelerator, helping offload tasks from the Grace CPU. This is really about data storage. Earlier on, they did not pay much attention, but starting from around Q4 2023 they began defining it seriously and kicked off the project. That is why we had already been working with Nvidia for two years, but it was not really aggressive until Q2 of last year. Since then, they have become very aggressive in defining the product, and they are planning to launch it in late Q3 or Q4 this year. Anton Shilov: You also mentioned compute storage. Does that mean you are working in that direction as well? Wallace C. Kou: No, we do not have the resources to develop compute storage ourselves, but our customers are moving in that direction. They are developing their own hardware and working with Nvidia. We provide the controller and the SDK, but we are not directly developing compute storage. Anton Shilov: Does that storage solution use CXL? Wallace C. Kou: It uses a PCIe interface. But I think we will have more internal meetings to review what they want to define, because right now they are aiming for 100 million IOPS — which is huge. The solutions you see today, like HBM3-based accelerators, are too expensive and not really workable. I believe they are looking for a media change. Optane was supposed to be the ideal solution, but it is gone now. Kioxia is trying to bring XL-NAND and improve its performance. SanDisk is trying to introduce High Bandwidth Flash (HBF), but honestly, I don't really believe in it. Right now, everyone is promoting their own technology, but the industry really needs something fundamentally new. Otherwise, it will be very hard to achieve 100 million IOPS and still be cost-effective. Anton Shilov: Now that you mentioned High Bandwidth Flash (HBF) — are you exploring various innovative types of non-volatile memory beyond NAND? Wallace C. Kou: Many startups come to us, but I do not see any that are really mature or executable at scale with acceptable costs. So far, it is very difficult. I do not see anything today that can replace NAND — it is very, very difficult. Kioxia's XL-NAND, which is 3D SLC NAND, could be one option, but it needs to become four times faster. And maybe it also needs a wider interface — I am not sure yet. We will have to see what the NAND makers can come up with, and how Nvidia will define what they really need — how to truly offload tasks. DDR5 and HBM cannot do that. What they want is to move data directly from GPU to SSD without involving the CPU — that would save a lot of transfer time and reduce latency. Anton Shilov: So basically, for now, you do not see anything replacing NAND in the next few years? Wallace C. Kou: I do not see anything happening in the next two or three years. But this is a long journey, and I am happy that at least now they are paying attention, starting to address the issue — because there is a real need. Anton Shilov: But you do not expect storage-class memory, like Intel's Optane, to become a major thing over the next, say, five years? Wallace C. Kou: I cannot say for sure about five years, but within the next three years, I would say no. Others are trying very hard — Micron has some development work, and others too, but nothing concrete yet. Except for XL-NAND from Kioxia — they are already selling it to hyperscalers today. Anton Shilov: Companies like Micron and SanDisk (previously Western Digital) have had SCM in their roadmap for years. So, they have shown SCM in the roadmap for years. Yet, these technologies change from time to time: from ReRAM a decade ago to something else today. There is no clarity at all. Wallace C. Kou: Today, Nvidia is stepping in to define things. Before, everyone talked about SCM — hyperscalers like Google, Microsoft, Amazon, and Meta — but in the end, they just showed interest; they did not actually commit. Now it is different — they are willing to pay, that changes everything. If they are willing to pay the price, it will change the whole mentality. If Jensen Huang really wants to drive this forward, it could transform the future of data centers — making AI training much faster and more efficient. This would create a new media ecosystem, and I think more players will invest and try harder to move in that direction. Anton Shilov: Then again, specialized storage solutions could be based on the NAND technology. Wallace C. Kou: It could be XL-NAND, just with a wider interface for extra bandwidth, right? Anton Shilov: You have mentioned QLC 3D NAND, XL-NAND. Would you expect PLC to become a viable thing over the next few years? Wallace C. Kou: Whatever Kioxia said before, I cannot comment on that. But I see it as very difficult for PLC to become commercial within the next five years. It's very challenging because the signal management would be extremely complex. I am already starting to see beyond 400-layer 3D NAND stacks, and the controller has to work very hard to overcome the inherent weaknesses of more bits per cell — like lower retention and endurance. We have already started using 16K LDPC (Low-Density Parity-Check) per codeword to provide better protection for enterprise products. There are many challenges — and nobody really wants soft decoding because it introduces the risk of reduced performance. At the same time, you must extend the lifecycle of the entire drive. It is all about balancing the technology and keeping up with the dynamic developments from NAND makers like Samsung, SK hynix, Kioxia, Solidigm, and YMTC. That is exactly what we are focused on. Anton Shilov: So, TLC and QLC will stay with us for years to come? Wallace C. Kou: Yes. Kioxia's and SanDisk's QLC is very good, and Micron's N69 QLC also looks strong. For Samsung and SK hynix, we are waiting to see their V9 generation next year — with 2TB QLC dies. They are promising very competitive products, but we'll have to test them before making any final comments. Anton Shilov: What is your short-term and mid-term business strategy and plans? Wallace C. Kou: Storage looks like it is going to grow rapidly — maybe not from unit volume, but from capacity, because storage sizes are getting much bigger. For us, our strategy is to keep gaining market share. In automotive, we are gaining share. In enterprise, we started from almost nothing and are gaining share. In client storage, we are definitely gaining share. Even in mobile, the market trend favors Silicon Motion. Today, we are about 20% of the 1.2 billion smartphone storage units, but within three years, we will grow to 30% for eMMC and UFS. I guarantee it, because the market trend favors us. You have to understand the whole mobile environment. For the past 15 years, Samsung, SK hynix, and Micron dominated mobile storage because they combined mobile DRAM and NAND in a single package — eMCP or uMCP — especially for mainstream and value-line smartphones. That made procurement much easier for smartphone makers. For us to win the smartphone business, we have to win over the NAND makers — Samsung, SK hynix, Micron. If they use fully integrated solutions, we are out. But we have worked very hard with module makers, starting more than 10 years ago with Longsys and Biwin. Back then, many mistakes were made, but three years ago, we started to see real results. Now, eMMC is declining in smartphones and UFS 2.2 is taking over. But eMMC is still growing in other areas — automotive, set-top boxes, smart TVs, IoT devices, even AI glasses. eMMC shipments are about 120 million units today and will grow to 130 million next year, and most of those devices use Silicon Motion controllers. Major DRAM players have exited eMMC because they do not see profits there anymore — that is where we stepped in. Similarly, UFS 2.2 is growing in value-line smartphones. We see a trend where Chinese smartphone makers — Oppo, Vivo, Xiaomi, Honor — are moving away from UMCP and instead using discrete NAND and LPDDR4/5 DRAM separately because it is cheaper and good enough. They use standard UFS solutions, and that is why module makers are jumping in. This also forces NAND makers to decide: should they continue offering UFS 2.2 integrated solutions, or should they work with Silicon Motion and outsource the controller and solution design? Anton Shilov: How is SMI's business split between SSDs, mobile storage, embedded storage, and expandable storage? Wallace C. Kou: So far, about two-thirds of our revenue comes from client SSDs, and about 30% comes from mobile storage. Mobile might even go a little higher this year. The rest — enterprise is still very small, and expandable storage is even smaller, just about 5%. But we are seeing growth in automotive — automotive already reached 5% of our revenue starting from Q3 last year. We expect automotive to grow and exceed 10% by 2026 – 2027. Enterprise should also grow — we expect it to reach around 5% to 10% of revenue by 2026 – 2027. Overall, non-client SSD and mobile segments will likely grow to around 20% by 2026 – 2027. So we want to become more balanced across segments. Also, we expect that by Q4, we will have a $1 billion revenue. Next year should be even stronger, and in the second half, you will start to see [new products] CR (customer ramp) revenue and new design wins ramping up. Even with tariffs and all the uncertainty, many investors and analysts tell us that most U.S. semiconductor companies are trimming down their guidance for the second half. But we look very promising because we are gaining market share, and our design-in pipeline is very strong. That's why we feel very comfortable with what we are doing today. Anton Shilov: What about your acquisition strategy? Over three decades you acquired Feiya Technology, Shannon Technology, and Future Communications IC (which was later sold). Only three companies. Why is that? Any future M&A plans? Wallace C. Kou: As I have mentioned before, we have done two major acquisitions in the past. The first was FCI (Future Communications IC). We were very interested in FCI because they were a successful RF company and a leader in mobile TV technology — particularly in Korea's mobile TV market. Japan was also launching mobile TV around that time — around 2005–2007 — and it looked very promising. We also saw some early adoption in Brazil, and the UK was trying to promote digital audio broadcasting. But in the end, mobile TV did not take off. iPhone did not support it, and video streaming — thanks to 4G — killed mobile TV before it could become mainstream. Even though FCI had success with Samsung for 4G and 5G RF, Samsung eventually developed their own RF solutions. Without strong demand and with no expertise on our side in RF, we decided to sell FCI to Dialog Semiconductor. The second acquisition was Shannon Systems, which we invested in starting around 2015. Shannon specialized in enterprise storage and had a unique architecture — atomic write and open-channel SSDs — and they worked closely with Alibaba and Baidu. Initially, it helped us gain a better understanding of the enterprise storage business in China. Revenue grew — Shannon almost reached $100 million a year — but it was difficult to make a profit because we had to procure NAND ourselves and build the whole product. Without strong, consistent customers and having to compete with NAND makers like Samsung, who could sell SSDs below cost, it was hard to sustain. Eventually, we changed our entire strategy. We decided to exit the full solution business because manufacturing and NAND procurement are not our strengths. Instead, we refocused on building high-quality PCIe controllers, because without strong PCIe IP and architecture, we could not grow in the mainstream market. About five years ago, we rebuilt the enterprise team. We shut down Shannon's solution operations and built a new controller architecture team — hiring people from Microchip, Marvell, and others. We expanded ASIC design, firmware development, C-modeling, qualification, and system architecture teams. That is how we developed MonTitan — our new enterprise controller platform — based on specifications from hyperscalers and leading enterprise server companies like Dell and HPE. MonTitan's architecture is unique, and it has helped us grow from PCIe 5.0 onward. Now we have tier-one customers in both the U.S. and China, and we're part of Nvidia's ecosystem. Looking ahead to PCIe 6.0, we have expanded our engine, and we already have formal engagements with top U.S. and Chinese customers. We have restructured R&D — China R&D focuses on Chinese customers, and U.S./Taiwan R&D focuses on U.S. customers. We maintain strict separation to avoid any IP crossing the Pacific Ocean and to stay fully compliant with U.S. regulations. At the same time, we continue to support our Chinese customers. We do not want to be labeled as anti-China — all major U.S. players still need their China business, and so do we. But we must balance that carefully and maintain a strong foundation for growing our enterprise business globally. Anton Shilov: Does that essentially mean that you strongly believe that internal organic development is a better strategy for growth than 'pure' M&A? Wallace C. Kou: For the enterprise portion, yes — I believe we need strong internal R&D teams. From an architecture perspective, we are capable, but I need more R&D resources — more ASIC and firmware engineers — to support many customers. I see QLC SSDs becoming very critical, and we are unique in our QLC management, especially with FDP (Flexible Data Placement), which makes us a complete solution provider and helps us build an ecosystem. For example, Kingston or Union can now be a supplier for SMI — many companies can use our controllers to make drives and support customers. But at the end of the day, I need to deliver real results. So far, I do not see M&A as a priority — at least not by acquiring large installed companies. If we were to consider acquisitions, I would prefer companies in Taiwan where I can manage them directly. U.S. or China M&A is very difficult — I have learned that lesson. U.S. acquisitions are too expensive, and with U.S.-China tensions, it is not practical. In the U.S., Korea, Japan, and China, we already have teams and we can grow organically. The bigger challenge is competition — it is hard to compete for R&D talent when Nvidia, Google, and others are hiring aggressively worldwide. But if there is a candidate in Taiwan that fits well, I would consider an acquisition — especially in IP areas related to connectivity. We have already started to build internal teams to develop PCIe Gen5/Gen6 controller IP and connectivity solutions separately, but I would like to have more resources to speed up development. I need more than just storage products — I need connectivity products too. Anton Shilov: You have navigated Silicon Motion through various crises, including the 2008–2009 financial crisis and the 2020–2021 pandemic. Do you have any special strategy for navigating through major disruptions? Wallace C. Kou: Yes. Since COVID and even through the failed MaxLinear acquisition, we've made significant management changes. We've upgraded our technology and algorithm leadership — changed our VP of legal, VP of operations, and manufacturing operations. We also moved to a business unit-driven model and strengthened our relationships with top-tier customers globally. So now, we know better how to manage through potential disasters and crises. Frankly, the Trump tariffs are another crisis. But we have been through enough to understand how to prepare. For example, in the U.S., we do not ship a lot directly — we have very few direct customers there — but Cisco is one of them. Anton Shilov: Tariffs were actually my next question. Wallace C. Kou: I can assure you, every major U.S. customer knows how to reduce their risks. They work with global manufacturing partners in Malaysia, Thailand, Mexico, Brazil — you name it. They know how to adapt, just like Kingston, which has a fully booked factory in California now because of this risk management approach. When I was at the JP Morgan conference last week, I heard many CEOs and even ambassadors talk about it. I would not say people ignore the tariffs, but they seem prepared. Everyone faces the same challenges, but they believe they will move on. As for Trump — he is very unpredictable — but in the end, I do not believe he would go extreme on tariffs, because that would harm himself and the Republican Party. Anton Shilov: It does not seem it makes sense for your to outsource part of your production to TSMC Arizona, does it? Wallace C. Kou: Never — because we are too small, and we have much more control in Taiwan. Why would I move unless someone forces me? Apple, maybe they have to. Nvidia might have to. Broadcom, Qualcomm — they need to do it to keep Trump happy. But we are too small to even show up on the radar, so I will not do it. Anton Shilov: Sure, and you just mentioned you have only one big direct customer in the U.S., so it hardly makes sense to produce chips there just for one client. Wallace C. Kou: Exactly. Most of my major customers manufacture overseas. They will figure out how to avoid the tariffs. I do not have to ship directly to the U.S. Anton Shilov: Do you think tariffs will affect the storage market in general, especially considering that NAND memory production is almost entirely based in Asia? Wallace C. Kou: So far, I do not think tariffs will affect storage much. I think the impact of tariffs on storage will be much less than on other sectors. Because of the AI trend, everything needs more capacity, right? And I do not count by unit sales — I count by capacity growth. As capacity demand increases, it drives storage demand. Meanwhile, NAND makers are trying to control bit output carefully to keep the market balanced, because overall channel demand is still weak. For example, smartphone growth is now only around 3% to 5%, and PC growth is about 4% to 5% — not the 7% to 8% growth people predicted late last year. Tariffs may have some impact, but storage demand is still strong. For us, I cannot control the impact of tariffs, but I can control gaining market share. As long as we continue to gain market share, we will be fine and we will grow. Anton Shilov: Thank you for a great interview with deep insights into the NAND industry. Wallace C. Kou: Thank you, Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.
Yahoo
23-05-2025
- Yahoo
Phison's Apex RAID demo showed us blistering 113 GB/s speeds in Computex demo
When you buy through links on our articles, Future and its syndication partners may earn a commission. Phison has been showcasing its data storage innovations at Computex 2025. Of particular interest, we attended a partner demo where the firm utilized 32 of its latest and greatest PCIe Gen5 SSDs in a RAID setup. During the demo, we saw CrystalDiskMark report data transfer rates as high as 113.6 GB/s for reads, and 104.6 GB/s for writes. This demonstration of extraordinary speed was enabled by a powerful modern workstation PC and a mix of cutting-edge storage components. Specifically, Phison utilized an AMD Ryzen Threadripper Pro 7978WX CPU, installed on an Asus Pro WS WRX90E-Sage SE motherboard, as the foundation of the system. The storage subsystem consisted of 32 of Phison's potent new E28 Gen5 SSDs installed across a trio of Apex Storage X16 Gen5 add-in cards. All these components were installed in a sweet-looking be quiet! chassis, as you can see. Image 1 of 3 Image 2 of 3 Image 3 of 3 Phison's new E28 Gen5 SSD rubs shoulders with the best in consumer mid-2025 PC land. A single E28 Gen5 in a modern PC will be able to deliver data transfer speeds of up to 14.8 GB/s reads and 14.0 GB/s writes. The Phison E28 Gen5 is being marketed as 'the weapon of choice for serious gaming and productivity,' so it is definitely targeting consumers. Its controller features a quad-CPU architecture, is fabbed on TSMC's 6nm node, and supports up to 32 TB. Here we had 32 of Phison's new E28 Gen5 SSDs; however, each Apex Storage X16 Gen5 add-in card can fit 16 M.2 SSDs (there's a clue in the name), so two would have sufficed for a 32x SSD demo. We guess three cards were installed for optimal performance, load balancing, or another nuance of the Threadripper-powered Windows system. The storage performance in this setup seems remarkable. Witnessed data transfer rates as high as 113.6 GB/s for reads, and 104.6 GB/s for writes might seem amazing. However, it was confirmed by a Phison rep that the Windows kernel was actually holding back performance. Image 1 of 3 Image 2 of 3 Image 3 of 3 Apex Storage's new X16 Gen5 add-in card features the PM584 Microchip switch that has 84 lanes, claimed to support the full 16 M.2 at full bandwidth. We were told each X16 Gen5 card also has a limit of 128TB of addressable storage, at this time (that would require using 32x 4TB drives). If you are interested in the Apex Storage X16 Gen5 card, we were told that it ships in 30 days and will cost $3,995 without any storage onboard. Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.
