 Okay, I think it's time. We can get started. Hello everybody, good afternoon. Thank you very much for joining this session. I'm going to talk about function-dedicated network deployment model in DPU and IPO-like device across the compulsory design in the infrastructure. You see the DPU slash IPU. See the device vendor is competing each other actually to innovate furthermore. My name is Haidu Sugiyama. I'm chief market director. I'm mainly contributing to two committees. One is open program infrastructure project, which is one of the Linux financial project, as you might know. And the other thing, I also are contributing Ion Global Forum. The quick question. How many guys know about the Ion Global Forum? Could you raise your hand if you know about the Ion Global Forum? Not yet, okay. Ion Global Forum was established in early 2020 by the NTT and the Sony and Intel. Unfortunately, at the time that you might know that the COVID situation, we couldn't have the onsite meeting many times, almost two years. But we have lots of remote conference to discuss the next generation AI and NTT infrastructure technology. And now the Ion Global Forum is growing into about 100 companies joining the Ion Global Forum. So through my talk, I'd like to use about Ion Global Forum, what technology we are going to develop. And I'd like you to understand why OPI, Open Program Infrastructure Project, needs for Ion Global Forum technology. We are in the middle of the hardware evolution. We are now developing the new desalate computing architecture. We call the Composer Desalate Infrastructure with adapting the CXL technology. You might know about the CXL aspect. After that, maybe PCIe Gen 5 and Gen 6 will adapt to the CXL protocol over the PCIe bus. Because due to the end of the moral, we need the solution to speed the computing processor. So we try to desalate the computing infrastructure. So Ion Global Forum, it's tried to develop the new architecture called data centric infrastructure, which adapting the Composer Desalate Computing Infrastructure. And primarily, we see the device innovation. Most of the device vendor is really keen to develop the DPU type of domain space hardware device. Because they also know about the current X86 centric architecture is one of the issues to speed the processor. So we address the domain space function into the DPU and IPU. And Ion also adopted in this technology, they call it the Function Dedicated Network. We're using the function card to build the Function Dedicated Network by like a DPU and IPU. Because we understand that the network function needs more intelligency, not just a simple smart nick to just off-road. There are lots of things that we need at the function on top of the network workload. So, yeah, 10 years ago I was busy to develop the NAB in the 4G era. And now we're shifting to the CNF with the Kubernetes for 5G era. But still not enough actually for the 6G, the CBP also feel that the network functionality needs more intelligency. That is also needed the DPU and IPU type of functionality to expand the, to increase the intelligence for network and also to support the end-to-end computing in addition to the edge computing. 5G are the most edge computing, but it's not enough actually. So actually the Ion Global Forum covered all of the challenge. And also that we are going to adapt the code package optics to the functional card like a DPU IPU that eliminated many electronic device actually. We can just directly receive the optical traffic over the optical network directly to the functional card and convert the photonic to electronics internally into computing architecture. And also maybe it's right word or not, but our next generation DPU IPU card we also see that the UCIE package. That is the new consortium launch in this summer. This is a die-to-die interconnect to increase the performance. I give you a little bit the scale of the computing interwork if available in the UCI package. And UCIE 1.0 adopted the CXL protocol to communicate the existing X86 computing architecture. So that's why Ion covered both innovation, one for the CXL-based design computing and the other is domain-specific hardware and innovation within DPU and DPU. So this time I like to focus on the DPU IPU deployment. Okay. The reason why that we need a Linux kernel supporting the CXL 3.0. Still not available enough, that's why. We are trying to do a parallel working. Once the CXL 3.0 is stable for the CXL 3.0, yeah, we can do the program together with DPU IPU. But at this stage, most of the DPU card is not available at the CXL because the new CXL is not available yet. So we are now separated working. And first of all, I'd like to explain about what is Ion. Ion stands for the Innovative Auticand Wireless Network. We try to eliminate the many electric network device with using the all-photonic network, with building all-photonic network. But out of that, still end users need communicating. End user cannot communicate over bit layer. They need pocket layer. So FDN, functional dedicated network would convert the photonic to electronics. So we still need internet protocol to communicate for service. So you see that this diagram, on top of the all-photonic network, you see the server, network service. On top of that, you see the overlay solution. But these are actually the logical view. Physically, these are written in each device. We try to compose the logical, we call the logical service node to compose with the available device like a GPU and DPU and the IPU to build the logical service node and to run the Kubernetes to deploy the network service work cloud as well as overlay solution like a HAI inference. So in the user side, they can see that there's a cloud environment. But for the internal infrastructure, we change lots of things for the more memory-centric architecture. Not necessary to send many network packets to each other. We just, once we receive the network packet, we try to use the memory technology to share the user data across the multiple node. Because if you think about AI development, AI needs a user's payload data. They don't need a network overhead. So once we receive that, we try to share the user's payload directly in the GPU. For example, the GPU directly is one of the technology. And also, we are targeting that this, maybe you can see that this number, we try to reduce the power consumption and increase the high transmission capacity and the low end latency. These numbers are really aggressive now. But based on the device research engineer, actually the number is based on the 2030. If we know where the device, we can read this number. If we change the architecture, if we keep the current architecture, it's not easy to meet this goal. But yeah, so I want to try to change a new architecture with using the current device innovation. So on top of that, what use case I want to try to realize, because when we launched the Ion Global Forum, first question is what use case we can try to realize for the 2030? It's really not easy to summarize, but there are many cross-industrial expertise joining the Ion Global Forum. So many cross-industrial expertise must summarize the 10 use case, yeah. And maybe we can classify the two type of use case. One is the cyber-physical system. The other one, the AI-integrated communication use case. And also the key requirement, we summarize the five key requirements, but these requirements are really similar, actually, based on the implementation. Many key requirements is that they need to increase, they need to make a low-processor latency. One of the big issues, because for the network latency, as long as we deploy the optical fiber infrastructure, it's net latency equal to the distance latency based on the photonic layer. But even we launched the photonic network, still problem is the computing latency. Most of the case, when we send the traffic in the current network, there's not much wide bandwidth to send the low-capture data. Most of the case, user try to compress the data packet to send the cloud. And then the cloud side to decompress the packet and then start analyze. So this compression, the decompression makes the CPU consumption, that's mean the more power consumption they spend. So I, in case we try to send, we can try to send the low data without compression. So that AI, GPU can directly receive the user's low data to start analyze with AI engine. So we can reduce the processor latency by using the ION. And by eliminating the electric network device and eliminate to reduce the CPU consumption, we can have that activity for net emission. That's why the many industries are keen to know about the activity around ION global forum. And the same thing for the entertainment side, AI-intuitive communication. One of the big issue, actually one of the key player entertainment, you guys asked to support 10 millisecond motion to photon latency for high performance XR. 10 millisecond motion photon latency is really hard to realize because the encoding and decoding takes time. I like maybe less than 10 millisecond, but that is a problem. So with using the low data, we can send the low data to directly the GPU memory. The GPU can start rendering. So we can try to make a 10 millisecond motion photon latency with that mechanism. So yeah, within ION technology, we try to make that a performance. So what technology we are now developing is that these are slides summarized at ION technology. For the foundational infrastructure side, yeah, open photonic network functional activities is one main part of the ION global forum technology. We build all photon network nationwide, not just local access. We try to establish all photon about 1,000 kilometers. That is one of the requirements. And on top of that, we can build a data-centered infrastructure. That's, this infrastructure is aggregating the packet to share the user data directory memory space for the guest users. And on top of that data-centered infrastructure functional architecture, we have a multiple workload for the network service provider. Data hub is one of the things, and also mobile network, 5G, radio access network and the 5G core is one of the things. We can run the container workload on top of the data-centered infrastructure. And fiber sensing is another unique use case. Actually, still under developing, but it's a very interesting use case ION Global Forum made. Fiber itself becomes a sensor. So we can probably in the future, we can reduce many electric sensor device by replacing the fiber anyway. But at this stage it's not ready actually. Maybe we can explore how we can collect data. Maybe I think the open telemetry project probably might be a good project to collaborate with this area. Still they have a problem to collect, how to collect the data and how to summarize data anyway. And at this stage, we are working on the reference information model for the existing sensor and security device. And to build the reference information model, one of the current challenges is scalability. Yeah, we have many sensor data collecting from each monitored area to the edge, but the problem with the performance, how to analyze and in the current situation, if we do the compression data, we collect the compression data. There are lots of CPU and processor usage we need use at the edge. So challenge, we try to collect low data from each monitor area to the edge and using that edge can analyze or perform the air inference directly with GPU memory. That is one of the reference model we already made and published. And now we are working on another two use case for the one is industry management. This is a remote robotics inspection reference information model we are now working. And another is interactive live music use case we are now doing. This is how to increase the rendering performance we are now making the reference information model still working on. And you can get the QR code. This QR code showed that you are there to download each document. And this slide is one of the main slide what I'd like to discuss. I'd like to introduce what is data centric infrastructure in the IOM Global Forum. So basically we are adopting the CXL-based desalated computing architecture, but CXL service is not available yet. So I cannot say that the CXL, so maybe we can say that PCI extension bus was something anyway in the document anyway. But once a Linux kernel available to the sheikhs we try to adopt the sheikhs different. And there are the two processor, one for the main CPU. The other is a processor in the DPIPU. Many maybe ARM processor for most of the case. Yeah, so we can allow the multiple operator to manage the resource. One is for support based on the support CPU. The other is DPIPU. So for example, the mobile network operator focus on the DPIPU management. So without interacting the first OS. That is one of the goal what the open program infrastructure project is doing. Open program infrastructure project try to create a common API to manage the life cycle manage for the dedicated DPIPU. Anyway, so now I'm working focus on the OPI side. Yeah, if we are focused on this here there are many, many solutions actually. We have talked about the Kubernetes or OpenSciAC or there are many ways to virtualize the resource anyway. So what we need to do now is the DPIPU. DPIPU, how we can manage the resource inside DPIPU and IPU. How we can interact with DPIPU and other device like a personal memory device or GPU anyway. So that area we are now collaborating the open program infrastructure project. So let me summarize what use case we are targeting Ion Global Forum. Ion Global Forum called Function Dedicated Network. This is actually just one of the use case in the DPIPU and IPU what we can integrate. So for example that RDMA is one of the things. There's many use case with RDMA. If we build that data hub, stream hub we can use it as a personal memory with RDMA. And also for the addition node and the AI inference or HAI inference over all photonic network we can also use it RDMA type technology. Try to collect data directly from the remote monitor side to the edge without additional network overhead. We can get the user's data directly to the memory space. For example, the GPU direct is one of the technology. We can translate the user data directly to GPU memory. So GPU directly analyze the data without interacting the main CPU. So we can reduce the CPU power consumption in the case. And also another thing that in XR performance is also a use case. When we make it a 10 millisecond motion photo latency to increase the rendering performance we can eliminate the compression and decompression task. We can just receive the raw data directly to the GPU memory for the rendering. And also another use case is high performance blockchain of RDMA. This is also another use case. And in terms of the other workload like a 5G radio access network and much access edge computing. Yes, also we can integrate the whole workload into the GPU and the IPU. That is one of the unique point is that mobile network operator manage the GPU and IPU directly without the main CPU that needs to share the resource for the guest user. So we can use the same box for the two types of the administrator. One for the network infrastructure operator. The other for the guest user operator. I like to share that maybe three use case what we are now working on. First use case is the area management use case based on the Ion Poc difference document. We are now doing the security use case to collect the many video camera data directed in gestion node with RDMA. So usually if we collected the video data in, I think for example, the 5,000 video camera. Traffic between the ingest and the localization need a 6G. So with using the optical fiber, we can do that. We can collect the 6G traffic and this ingestion node directly send the capture data directly to the GPU and the GPU can perform the AIH inference and store the result of data to the data hub. The data, between the data, H data hub and the central cloud data hub, we can also use the RDMA to share the data over the parsed memory. So the intelligent application in the cloud side just use the memory to get the data. Not necessary to the network protocol at the time. And there's one challenge is that because currently in the monitor area, we have to collect the 60 Gbps traffic if we aggregate, but it's not easy for the mobile network. Currently we cannot make it to transit the 60 Gbps traffic over the wireless access. So in shift to that, we can try to get each video data over RTP directed to the mobile edge. Then we can send, we can decaps RTP packet to get the user data directly to the GPU memory. This is similar way to the GPU direction to get the user data directly to the memory space of the GPU. It's not necessary to interact the CPU processor data. So we can keep reduce the CPU power consumption in the scenario. And another thing is that we established the front four network and the mid four network by DPU and IPU. So these DPU and IPU are managed by the mobile network operator separate from the other administrator. So with using DPU and IPU, we can isolate the administrator. So even though everybody use the same system, but each guy can isolate the resource each other. So we can keep secure the resource for each. So more easy way for the infrastructure sharing service. My one mobile network, maybe several mobile network operator use the same system, but actually each of the mobile network use all DPU and IPU. Maybe in near future, softback DPU and NTTDocomo DPU and Rakuten DPU. So everybody can use a focus on the DPU, IPU card level rather than just a code server. The other use case is that we are still working on products, but I like to show that we are also working on the industry management, remote inspection, robotic inspection. This is a smart plant project. Actually we use a UAB, which is a drone in the 3GPP name, okay. And we provide the reliable access to the mobile edge in the customer side. And we collect the user's data and 4K, 8K video encoding to here and we can decouple the user data and directly send the user data over the RDMA to the remote inspection side. So because one of the issues in the industry management, there's not so many expertise who can manage, monitor the plant. So then most of the case that expertise in the remote side needs more operate. So that's why we try to send low data to the user remote side for the rendering. At the source at the same time that also we try to make an enterprise communication for the UAB controller to the navigation application. So one of the challenges, if we use just TCPIP, TCPIP itself makes a distance issue because TCPIP, they need check up. So we try to terminate the multiple TCPIP locally. So no need distance concern and to send control packet to the UAB controller. Then this is just running the processor to communicate the remote navigator over shared memory communication with using RDMA. So between the site, a local plant and the remote inspection side, there's no TCPIP connection. Just run the IPC communication using the shared memory. So we can reduce the latency in this scenario. The other scenario use case I like to share is mobility management. One of the issue of the mobility management is if the vehicle move to the multiple cell group, users data need handover to another central unit. There's a not good solution, but within the IWON and using the RDMA over long distance APN, the CPU also can use the shared memory communication directly without sending network packet. So once this UE connecting the PDC packet here and when the UE moved to another cell side, this CPU connecting the secondary CPU over the memory, shared memory, so that even the UE connecting here still see you can keep the same PDC packet here. So that is the one of the solution we are now exploring and to keep the more minimum mobility latency. So we are still under discussion, but it's one of the use case I can introduce here. So there are many use case for the FDN deployment. And we can use the Kubernetes operator to deploy the EHR code, but one of the biggest challenges is you see that many vendor have the DPU product. Each vendor has each SDK. There's no common API. That's the real issue that when we deploy the software on top of each DPU, we need the different API. So that's why I realized that issue. And fortunately, Linux Foundation launched Open Planner Infrastructure project this summer. That's Open Planner Infrastructure project. It's trying to eliminate vendor lock-in, actually. It's trying to create device vendor analysis solution to create a common API. This is one of the main project Open Planner Infrastructure project. I like to spend the time to discuss this, but actually we have another session tomorrow to introduce Open Planner Infrastructure project. I like this topic tomorrow. Anyway, but what I like to explain why the Ion Global Forum needs the Open API project is here. Yeah, we have Intel and NVIDIA are the member of the Ion Global Forum. And actually this October, we have a first Ion member meeting held in New York. I invited an OPI member to introduce this and to discuss how to collaborate with each other. There are many outcomes we can make together, actually. So Ion Global Forum is not a software community. Ion Global Forum tried to adopt the latest software. So we need to collaborate with many software community to realize a new architecture. This is one of the use case, why that's why we are now working with OPI project, okay? So now Ion Global Forum published the seven Poc reference document. You can find that each Poc reference document, what technology we need implement, yeah. Actually, this already published, you can scan the QR code to get each document. And one of the unique thing in the Ion Global Forum is that we have a non-member program that collaborate the Poc project together. So we are very interested that if the software community like to contribute, yeah, it's very welcome to discuss each reference Poc document, okay? So let me summarize. Ion Global Forum is now developing data-centric infrastructure architecture, which is a composed of this type of infrastructure. To realize this architecture, we need a Linux kernel supporting CXS 3.0, actually. Parallel, we are also spending time to discuss infrastructure work load to integrate into the DPU and the IPU, okay? And at this stage, DPU card and IPU card are not available for the CXS type of protocol, but device vendor is now announcing the several device supporting CXS aspect. So I hope that the DPU and IPU will be available in near future support CXS. So far, we use the CXS.io because OPI uses SRV. SRV means the CXS.io in the CXS spec case. So we can keep using the existing interface, but what I like to explore is that how we can use it to share the data access memory using the CXS memory spec. So there are many activities we need to do anyway, but we have to do step by step. So we now try to do the PUC with PCI-GN4, which is not available to CXS at this stage, but we are trying to integrate a mobile network called in DPU and also RDMA. So type of the application also can be integrated in DPU and IPU, so we are moving to the PUC phase. And more, do we speak hardware like a DPU, IPU type of the use case will be available and it's needed. And the challenge is the common deployment process. So we are tracking the open program infrastructure project for DPU and IPU deployment because the I1 is not a software community. We are aligned to the software community to share our requirement to them. So we have a PUC project. If you are interested, please join that and also contact this URL and you can also scan this QR code to connect to that. And if any question about I1 PUC, please submit it here. Okay, and tomorrow we have an open program for such a project session in, I think at 11.10 a.m. I'm not sure that the room, I don't know maybe, but you can check, you can scan this page, yeah. Okay, that's all my side. So you have any question? Hello, I'm Aki from F5. I have a very basic question about I1. So does I1 develop any standards or specifications? Excuse me, could you repeat again? Does I1 develop standards or specifications? Standard, yeah, that's a good question. I1 try to make a standard for the next generation AI-NATV processor, but I also use a standard organization, right? We have a liaison agreement with many standard body and we share the idea and the information each other. So we don't do the same activity, but the organization already do, for example, the Linux Foundation already launched the open program reference project. We adapt that project. We try to support that project. So not necessary to make a new standard, okay? So it's more like a reference architecture utilizing existing standards and technologies. And show that the new use case, prior for the 2030, there are many cross-industry expertise here in I1. So they can appear at the new use case and to help the other committee to realize that. So the goal is to realize the new use case around the 2030 to reducing the power consumption and to increase the more high power performance, yeah. Thank you. Okay, thank you very much.