Exclusive Interview With Uri Beitler, CEO, Pliops

Uri Beitler is the CEO and co-founder of Pliops, a company he launched in 2017 with Aryeh Mergi. Before that, he worked at Samsung and NovelStat. He is based in Israel where the company HQ is located.

StorageNewsletter: Pliops was founded in 2017 and you just unveiled the XPD for Extreme Data Processor as your first product, could you give us a company snapshot with some financial figures, employees, offices… with founders background as well?
Beitler: Pliops is on a mission to massively accelerate performance and dramatically lower infrastructure costs for applications that run on flash. These include all types of databases, analytics, AI, and ingest heavy workloads like 5G and IoT.

I’m very proud of what the team has accomplished in bringing our first product to market. We have created a new category of products that brings tremendous value to our customers. We recently hired our 100th employee and are rapidly building the team out in all functional areas. Most of our people are on the R&D team located in Israel, and the business team is headquartered in San Jose, CA. Our sales teams are in the US and APAC today, and we are rapidly expanding.

Before co-founding Pliops, I led Samsung’s SSD development center in Israel, and with CTO Moshe Twitto, designed advanced NAND storage technologies. Both of us also served together in the Unit 8200 intelligence division of the Israel Defense Corps. Our chairman was co-founder of  M-Systems, which invented the SSD, and he co-founded all-flash array innovator, XtremIO. Together, we built an elite engineering team from the deep Israel pool of technology experts in database, flash storage, and semiconductors.

What was the genesis of the project? How did it evolve as a company? What does Pliops mean?
Data is one of the most powerful tools we have to make the world a better place for ourselves and our families. After 10 years of analyzing and architecting data storage, it became clear that we have hit a wall. The data needs of today are not compatible with the data center architecture of yesterday.

Moshe and I founded Pliops when we saw how massive data growth is colliding with legacy compute and storage shortcomings. The industry is struggling with slowdowns in computing, storage, and networking efficiency. These are impacting our collective ability to innovate and our ability to use an increasing volume of data to make the world better. The challenge keeps getting harder, and I knew we needed a revolutionary data processor, but it had to be easy to deploy and scale.

Our name Pliops is simply a combination of the Greek word Plio, which means many or greater in size, and IO/s. Our mission is to multiply storage performance and efficiency, so it’s a good snapshot of what we do.

Did you suffer from the health situation or being in a product development phase?
I am very proud of the our team’s resilience through a very challenging year and a half. Our team continued to execute without missing a beat, instilling a high level of trust in customers. This trust is critical, as we had to overcome the difficulties of engaging with customers when meetings could not be in person. Also, getting our hands on servers was tricky (and still is), as development teams focused on dealing with the pandemic. In the end, we are thrilled that customers found the benefits of doing business with Pliops were worth the effort.

So far you raised $115 million, and we see various type of VC profile participating in your capital, why this is so wide? And do you plan to accelerate in this domain to boost your growth and geo presence?
The company has raised $115 million to date, which puts us in great shape to drive our plans aggressively. With the majority still in the bank from previous rounds, this was about growing as fast as possible. 

There is a tremendous amount of interest in what we are doing. The last round was oversubscribed, with Nvidia increasing its share. We are scaling our technology into new use cases, expanding product lines, and we expect to double the number of employees this year and again next year.

While storage and the computational requirements for processing the data are growing exponentially, status quo approaches are insufficient. Our solution takes a new direction, one that enables exceptional scalability and efficiency for data centers.

Having investors and partners like Intel, NVIDIA, Western Digital, Xilinx (soon to be AMD), Softbank, and others is a one-of-a-kind combination in the industry. We take that as a decisive vote of confidence that we struck a nerve and are solving critical problems. We are growing as fast as possible to meet high demand and building long-term partnerships with our investors and strategic customers.

We count several initiatives in the industry, lots of new players, how do you describe this market and what are the various directions and iterations promoted by these vendors? There is Data Processing Unit aka DPU, Computational Storage or CS, could you help to understand similarities and differences between them?
First, it’s important to note that we are complementary to GPUs, DPUs, and CPUs. We like to compare our Extreme Data Processor (XDP) to these devices in terms of applicability (breadth) and performance/cost.

GPUs are fantastic for what they do, but their applicability is mainly for AI. CPUs can do just about anything, but performance/cost is getting increasingly worse with Moore’s Law slowing. DPUs are emerging and provide important security improvements in IaaS environments for hosting VMs.

Computational storage is a good description of our concept, but we take a different standards-based, turn-key approach, using an easy-to-deploy, low-profile PCIe card that works with any SSD. Installation is plug-and-play using a standard block interface requiring no application changes or a native key-value interface that uses the defacto RocksDB APIs for the highest possible performance and compatibility. In many cases, there are no software changes required to use our native RocksDB interface as well.

Pliops develops the XDP as a DPU, beyond general and common characteristics you share with other players in that area, what are the key differentiators in favor of the XDP?
We don’t use the term DPU to describe our product because that implies integrated network functions. We are focused on a universal problem – at the last stage of any data path, there’s a processor reading/writing to a disk – we want to make that as efficient as possible. We are also accelerating host-level functions for database performance, analytics, and other data-intensive applications.

XDP manages data flow from application to storage and shapes the data for optimal placement on the lowest-cost SSDs. It also delivers breakthrough performance with fully offloaded data protection and inline transparent compression. We work with all standard servers and with all existing SSD suppliers. XDP is unique because it is the only solution that maximizes nearly all SSD-based application performance, efficiency, and resilience using the lowest cost TLC or QLC SSDs.

In the current data deluge climate and the few years coming, how this kind of architecture can help and solve the data management challenge in front of us?
As technology penetrates every aspect of our world, storage requirements are doubling every 3 years. That’s remarkable growth, and there is no end in sight. Not so many years ago, SSDs arrived to solve performance problems, and soon 75% of enterprise data will be stored in flash. But today’s data volumes have exposed the inefficiency of compute in legacy architecture, putting data-hungry, storage-intensive applications at risk.

To accelerate these applications, you need more from your infrastructure: more application speed, more storage capacity, more data protection, and more scalability. Until now, the only option was to throw money at the problem by adding servers, SSDs, and data center footprint. But that’s not solving the real problem, and it’s not helping the environment.

The success of GPUs has proven, a purpose-built processor that accelerates specific workloads is the right approach. Our game-changing XDP overcomes storage inefficiencies to massively accelerate performance and dramatically lower overall infrastructure costs. We enable companies to do more with less, and that’s more important than ever, given today’s ongoing chip shortage.

Does it mean you also implement some AI/ML algorithms? Could you illustrate with some examples?
With XDP, we have implemented breakthrough data structures and algorithms that deliver performance equivalent to ~500 Xeon Gold cores on a single <45W device. While we don’t need AI/ML techniques to accomplish this today, XDP provides tremendous value for accelerating AI/ML workloads.

Today the XDP is a FPGA-based PCI card, will it be the strategy for the next few years or do you think about a new approach? It is particularly interesting as Xilinx is one of your investors and a well known market leader in FPGAs.
We focus on the complete product and how it integrates into our customer’s solutions. To our customers, what device is “under the hood” doesn’t matter. With Xilinx as an investor and supplier, we have access to their most advanced FPGA technology, which gives us the flexibility to innovate quickly and deliver more value. We are actively developing an ASIC version that will provide even higher performance for high volume because XDP accelerates so many applications and makes sense anywhere there is storage – every server, storage server, or storage system.

Addressing the market and high demanding applications, what are your priorities in terms of use cases? And who are your target customers, we imagine hyperscalers are in the radar, who else?
Yes. Hyperscalers have been an initial priority, but we are also engaging with large and medium-sized CSPs, Fortune 500 companies, HPC labs, and more. We have been focused on flash-based workloads starting with databases, MySQL, NoSQL, in-memory, and analytics, but XDP accelerates any flash-based application. 

Our current customer base consists of many of the largest cloud data center providers and users. We have multiple public customers including Western Digital’s validation and manufacturing engineering and Sakura Internet, the largest hosting provider in Japan. Stay tuned for more details on other strategic customers including a >$100 billion pure SaaS provider, one of the top 3 China hyperscalers, and a large US Government HPC lab.

Supporting these needs means XDP can be configured in different modes. Do you support multiple cards per server? What about grouping servers via their DPUs and making a cluster of data processing units?
XDP is easy to integrate, supporting several interfaces. First is a standard block interface, which is universally adopted in the industry today. It looks like any drive, any storage device, and just shows up; no SW changes are needed. XDP also has a native Key-Value Library API. It’s RocksDB compatible, and we also support the emerging NVMe-KV standard. Native KV is the most efficient way to access the system for the highest possible performance.

Our customers deploy XDP in a variety of ways. Some use multiple XDP cards in a single server to manage SSDs. We also work with NVMe-oF to manage NVMe drives across multiple servers, storage servers, or JBOFs.

Using XDP in combination with a cluster of DPUs is an absolutely valid and exciting use case. This is the future of infrastructure and we are working closely with NVIDIA to make this the new normal in the datacenter. 

What are the gains of such a model, I mean a DPU and in particular your XDP card? Of course, performance but does it goes beyond that?
Our 4 value pillars are performance, rreliability, capacity, and efficiency. XDP exponentially increases these key capabilities by radically simplifying how data is processed and SSD storage is managed.

Our customers and partners see incredible numbers across a range of popular applications where they use SSDs. For example, with RocksDB, we see 10x, even 25x gain in performance compared to software. This improvement comes with rock-solid, very low QoS as measured by four nines latency even at full system and drive utilization.

RocksDB is a critical component of many popular applications, including Ceph, Nutanix, Spark, and Redis. We’re deploying Redis with up to a 7x increase in performance, matching the performance and latency of DRAM. Yet, XDP achieves this at a fraction of the cost of DRAM using low-cost SSDs.

With popular databases like MariaDB, MySQL Mongo DB, and others, XDP increases performance up to 3x over a software-only setup and requires zero application changes.

Beyond performance, we have Pliops Drive Fail Protection which enables users to maintain constant data availability and deliver on stringent SLAs while delivering this higher performance. Many customers tell us this is a game-changer for them, and we are unique in the industry. With the current chip shortage, even if you have an unlimited budget, you may not even get all the servers and SSDs you want. XDP lets you get more out of your existing infrastructure with our in-line compression and data reduction technology that provides up to 6x space savings without affecting performance using the lowest cost QLC SSDs.

Lastly, I would highlight that XDP is a single solution for every workload and helps optimize your data center infrastructure footprint. It can reduce data center footprint or increase instance density up to 10x.

What about your cloud strategy, both in terms of prospects/customers and processing model?
Our go-to-market focus so far has primarily been with large cloud/web service providers and you will soon see these companies offering Pliops-accelerated services, like Redis and MySQL, soon, as well as accelerated instances to allow users to take full advantage of Pliops benefits with a cloud deployment and consumption model.

How do you sell the product? Such technology seems to be perfect for an OEM model. If yes, are you ready to share with us some names? And what about tier-2 distribution?
We engage directly with the largest cloud and web service providers. They are big, growing fast, care about infrastructure efficiency, and control their supply chain. We are also expanding to Fortune 500 and HPC.  Several of the largest OEMs in the US and China are in process to integrate XDP in their next generation servers.  And we are scaling our channel ramp via our recently announced partnership with TD Synnex in North America, and previously announced relationships with Comtech and Allianz in China, as well as Servants International in Japan.

How do you price XDP? Is it per card, server, capacity under management, basic and optional features, standard and enterprise models…?
Today, pricing is per card and dependent on volume. Every card comes with the same feature set. 

TCO is where Pliops really shines with customers realizing 35% to 80% reductions in TCO. In radically simplifying how data is processed and stored, we multiply data-intensive application performance while slashing infrastructure costs.

What are your next steps? What can we expect for the rest of 2021?
We have a robust roadmap for the rest of 2021 and beyond and are executing aggressively on it. Working closely with many large customers, we’ve learned a lot about how they use XDP and what capabilities are essential to help them achieve their goals. We’ve observed 4 key areas (see above) our customers are asking for help deploying flash today, and we are leaning into them as the north star to inform our product vision. Each of these 4 pillars delivers tremendous value on its own, but together the value multiplies to something; we call that the Power of X.