Liqid - IT Press Tour #44 June 2022

Transcript

1. © 2022 Liqid, Inc. Proprietary & Confidential 1 Dynamic Infrastructure for Modern Workloads Liqid CDI Overview
2. © 2022 Liqid, Inc. Proprietary & Confidential 2 Company Snapshot SOLUTION Liqid Matrix Software & Fabric Technology for Composable Infrastructure VALUE PROP Increase Infrastructure Flexibility, Agility and Efficiency from Core to Edge OPPORTUNITY Transform the Data Center to Software Defined Infrastructure TARGET MARKETS HPC – $38B Cloud – $38B AI – $43B Edge – $33B 100’s of CUSTOMERS Key Awards OFFICES HQ: Broomfield, CO Offices: California, New York, and Texas and UK PATENTS ~85 Patents & IP Filed – ~55 Patents & IP Issued
3. © 2022 Liqid, Inc. Proprietary & Confidential 3 “Disaggregate the converged server …then compose and reconfigure …that’s a revolution.” Jensen Huang, NVIDIA CEO Three-Tier Hyperconverged Software Composable Software Defined Data Center Source: Gartner Datacenter Transformation
4. © 2022 Liqid, Inc. Proprietary & Confidential 4 Introducing the Dynamic Datacenter Storage NVMe Flash / Optane Hosts Intel / AMD / ARM Networking NIC Accelerators GPU / FPGA / DPU Dynamically Configured Servers Disaggregated Resource Pools Composable Software & Fabric Static Servers Legacy Infrastructure Modern Infrastructure Statically Configured Dynamically Configurable
5. © 2022 Liqid, Inc. Proprietary & Confidential 5 Dynamically Configured Servers Enable New Configurations Independent Scaling of Resources Improve SW License Efficiency Decouple Purchasing Decisions Pay As You Grow 2x-3x Extend Product Life Cycle Improved Resource Utilization Benefits of Composable
6. © 2022 Liqid, Inc. Proprietary & Confidential 6 Cloud Flexibility On-Prem at a Lower Cost
7. © 2022 Liqid, Inc. Proprietary & Confidential 7 Use Cases & Customers AI & Deep Learning HPC Dynamic Cloud 5G Edge
8. © 2022 Liqid, Inc. Proprietary & Confidential 8 Use Cases & Customers AI & Deep Learning HPC Dynamic Cloud 5G Edge Composable Architectures Benefit High Value Mission Critical Workloads
9. © 2022 Liqid, Inc. Proprietary & Confidential 9 Use Case: Composable AI Platform Data Pool Data Ingest Data Clean and Tag Training Inference Dynamic Resource Allocation for Each Stage of AI Workflow Liqid Mgmt. SW
10. © 2021 Liqid, Inc. All rights reserved. 10 Proprietary & Confidential Liqid Matrix Software and Fabric Servers Nodes Composable Resources (GPU, SSD, NIC, etc.) Liqid Matrix Software and Fabric Dynamic Resource Allocation for Each Stage of A.I. Workflow Use Case: Composable AI Platform
11. © 2022 Liqid, Inc. Proprietary & Confidential 11 Liqid Composable Elements Fabric Integrated PCIe Expansion CDI Software Liqid EX 4400 Series 10/20 Slot UI | API | CLI 4U GPU | FPGA | SSD | NIC Storage & Memory 32-Bay JBOF SSD | Storage Class Memory
12. © 2022 Liqid, Inc. Proprietary & Confidential 12 CDI Deployment Over Fabric • Plug and Play Performance • Add Up to 20 GPU/Server • Scale Throughput on Demand 16GB/s Throughput 32GB/s Throughput 64GB/s Throughput Liqid EX-4410 PCIe HBA 1U Server
13. © 2022 Liqid, Inc. Proprietary & Confidential 13 CDI Deployment Over Fabric • Scale from 1 to 16 Hosts • Reconfigure in Seconds • Flexible CPU<>GPU ratios Liqid EX-4410
14. © 2022 Liqid, Inc. Proprietary & Confidential 14 Deploy Large-Scale CDI Environments Storage Class Memory Expansion GPU | FPGA |NVMe | NIC | DPU Hosts NVMe Flash 48-Port PCIe Fabric Switch
15. © 2022 Liqid, Inc. Proprietary & Confidential 15 Eth/IB Switching Storage Class Memory (48TB) Accelerators (90x) GPUs, FPGAs, NICs NVMe SSD (1.1PB) Compute Nodes (96x) PCIe Switch Fabric Rack Scale CDI Reference Design
16. © 2022 Liqid, Inc. Proprietary & Confidential 16 Enable High Performance GPU Scale-Out Solutions Resnet  Octane  V-Ray Benchmark Record Holder 20x 1x GPU CPU 960 1,908 3,747 7,369 13,386 15,814 1 2 4 8 16 20 NUMBER OF NVIDIA RTX-8000 GPU Resnet50 – Images/Sec
17. © 2022 Liqid, Inc. Proprietary & Confidential 17 20x RTX8000 GPUs 1x CPU Enable Impossible GPU Configurations Resnet  Octane  V-Ray Benchmark Record Holder 20x 1x GPU CPU
18. © 2022 Liqid, Inc. Proprietary & Confidential 18 Performance Composed vs. Static
19. © 2022 Liqid, Inc. Proprietary & Confidential 19 Highest ResNet Published Performance for Nvidia Based Solution World Record AI Benchmark Performance 8-Way NVLink = ~22K Images/Sec Source: Orange https://www.liqid.com/blog/orange-silicon-valley-composed-fastest-single-node-gpu-supercomputers-with-liqid-matrix
20. © 2022 Liqid, Inc. All rights reserved. 20 Proprietary & Confidential Liqid ioDirect Peer-to-Peer Technology GPU to GPU w/P2P Disabled • Bandwidth: 9 GB/s • Latency: 25 µs GPU to GPU w/P2P Enabled • Bandwidth: 49 GB/s 474% Higher BW • Latency: 3 µs 89% Lower Latency Host CPU Memory PCIe Fabric Host CPU Memory PCIe Fabric GPU Enclosure GPU Enclosure GPU Enclosure GPU Enclosure
21. © 2022 Liqid, Inc. All rights reserved. 21 Proprietary & Confidential Introducing Liqid GPU Direct Storage (GDS) Technology GPU to SSD w/P2P Disabled • IOPS: 179K • Latency: 583 µs • BW: 11.7 GB/s GPU to SSD w/P2P Enabled • IOPS: 2,900K 15X More IOPS • Latency: 80 µs 86% Lower Latency • BW: 16.5 GB/s 41% More Bandwidth Host CPU Memory Expansion Chassis PCIe Fabric Host CPU Memory PCIe Fabric Expansion Chassis
22. © 2022 Liqid, Inc. All rights reserved. 22 Proprietary & Confidential Data Flow Improvements of Peer-2-Peer CPU Memory PCIe Fabric GPU SSD Host CPU Memory PCIe Fabric GPU SSD NIC Legacy P2P
23. © 2022 Liqid, Inc. Proprietary & Confidential 23 Accelerate Deployment with Orchestration Storage Host NIC GPU Rest API Liqid Mgmt. SW Resource Pools Bare Metal Hosts
24. © 2021 Liqid, Inc. All rights reserved. 24 SLURM Northbound API Integration
25. © 2021 Liqid, Inc. All rights reserved. 25 VMware vCenter Integration https://www.vmware.com/resources/compatibility/detail.p hp?deviceCategory=wcp&productid=53418&vcl=true VMware Certified: Compose via vCenter • Single pane management • Dynamically configure servers • Scale beyond server sheet metal
26. © 2022 Liqid, Inc. Proprietary & Confidential 26 Use Case: Machine Containers Software Container Deployed with Predefined Hardware Containers
27. © 2021 Liqid, Inc. All rights reserved. 27 Proprietary & Confidential Policy Based Management
28. © 2022 Liqid, Inc. All rights reserved. 28 Proprietary & Confidential Full Stack Composability PCI Express (Gen 4) Fabric Layer (Commodity HW) Device Layer (Commodity HW) OS, Application & Virtualization Layer Ethernet InfiniBand PCI Express (Gen 5) CXL 5G Accelerators Liqid Matrix SW Layer Networking Pool Storage Pool CPU Pool FPGA Pool Future (i.e. DPUs etc) Memory GPU Pool
29. © 2022 Liqid, Inc. All rights reserved. 29 Proprietary & Confidential Compute Pool Storage Pool PCIe Fabric Storage Pool Networking Pool Accelerators Ethernet Fabric Data Path Control Path Fabric-A PCIe Switch Fabric-B Eth Switch Fabric-A Connection Fabric-B Connection Liqid Mgmt. SW Multi-Fabric Composability
30. © 2022 Liqid, Inc. All rights reserved. 30 Proprietary & Confidential Vision: Enable Every Server With Composable Software Infrastructure Has Become Software Composable Every server should have dynamically configurable resources to best meet the application layer requirements …. Liqid software is the solution Software Defined Everything
31. © 2022 Liqid, Inc. Proprietary & Confidential 31 Source: https://bit.ly/3uYLoUs Non-Production Oversized Idle Time Overspent Low Infrastructure Utilization rate THE SPEND: $266.4 Billion THE WASTE: 44% 40% 76% 50% THE SOLUTION: Achieve Up to 90% Datacenter Utilization
32. © 2021 Liqid, Inc. All rights reserved. 32 Proprietary & Confidential 20% of the world’s electricity by 2025 Massive Amounts of Water Consumption 1,600T Gallons Consumption Per Year 1MW = 8.5M Gallons of Water Per Year 190GW Global Datacenter Power Consumption 8MM+ Datacenters Today 500k Datacenters in 2012 Liqid Enables Up To 2-3x Datacenter Utilization Liqid Enables Up To 75%+ Reduction in Power The Problem: Massive Inefficiency in the Data Center Increasing Power Needs Datacenter Footprint Growth
33. © 2021 Liqid, Inc. All rights reserved. 33 Proprietary & Confidential Comment: Exascale Compute Requires Massive amounts of water and energy Annual Energy and Water Consumption Statistic: • 40 MW @ $0.10 kWh = $35M Per Year • 6,000 GPM @ 90% Efficiency = ~320M Gallons Per Year Exascale Computing * 6,000 gallons per minute = ~3.15B gallons of water per year * 90% of water is recycled = ~320M gallons lost to evaporation https://www.eia.gov/electricity/state/
34. © 2021 Liqid, Inc. All rights reserved. 34 Proprietary & Confidential Increasing Energy Prices Power and cooling cost now make up majority of the costs to deploy large scale data centers
35. © 2021 Liqid, Inc. All rights reserved. 35 Proprietary & Confidential ESG White Paper ESG White Paper on the benefits of composable architectures on sustainability https://cdi.liqid.com/esg- showcase-liqid-sustainability
36. © 2022 Liqid, Inc. Proprietary & Confidential 36 Liqid Matrix Benefits • Reduce Deployment Times • Scale Resources Independently • Run any OS or Hypervisor • Dynamically Configure Servers • Deploy Advanced Configurations • Enable as-a-Service Simplicity • Improve Resource Utilization • Address Sustainability Initiatives • Decouple Purchasing Decisions Accelerate Time to Value Improve IT Agility Increase Efficiency
37. © 2021 Liqid, Inc. Proprietary & Confidential 37 NVMe Storage CDI Management Liqid Matrix Software on Liqid Director 1U Liqid NVMe PCIe Storage SSD or Storage Class Memory FHFL AIC HHHL AIC Liqid Product Portfolio
38. © 2022 Liqid, Inc. Proprietary & Confidential 38 High Performance Storage Fabric Single Device • 8x SSD Controller • 64x NAND Channels • Aggregated Performance • High Capacity • Enterprise Features Performance, Capacity, and Features Enabled Through PCIe Fabric • Patented Pending Design • Enterprise Optimized Firmware • Managed PCIe switch • On-board Pfail Protection • Active ARM Mgmt. Processor • Enhanced Monitoring • High Margin SI Design • Efficient Thermal Dissipation • Active Thermal Throttling • Flexible Drive Topology • Dual and Quad Port Supported PCIe Fabric SSD Pfail ARM SSD SSD SSD SSD SSD SSD SSD
39. © 2022 Liqid, Inc. Proprietary & Confidential 39 LQD4500: Gen4 NVMe SSD Vendor A Liqid Seq Read 128KB 8,000 24,000 Seq Write 128KB 3,600 10,000 Rnd Read 4KB 1,450,000 3,150,000 Rnd Write 4KB 260,000 950,000 LQD4500 Product Details Product: Gen4 NVMe SSD AIC Host Interface: Gen4x16 Form Factor: FHFL AIC Power Loss Protection: Yes Max Capacity: 30.2TB Higher Performance High Capacity Advanced Features
40. © 2022 Liqid, Inc. Proprietary & Confidential 40 LQD4900: Gen4 Intel Optane Storage LQD4900 Product Details Product: Gen4 NVMe SSD AIC Host Interface: Gen4x16 Form Factor: FHFL AIC Power Loss Protection: Yes Max Capacity: 3.0TB Intel Optane Technology Liqid Powered Intel Optane Technology Capacity 1.5TB 3.0TB Seq Read 128KB 2,500 20,000 Seq Write 128KB 2,200 20,000 Rnd Read 4KB 550,000 4,000,000 Rnd Write 4KB 550,000 4,000,000 Interface Gen3 x4 Gen4x16
41. © 2022 Liqid, Inc. Proprietary & Confidential 41 CDI Demo
42. © 2022 Liqid, Inc. Proprietary & Confidential 42 GTM, Customers & Roadmap
43. © 2022 Liqid, Inc. Proprietary & Confidential 43 Target Market Segments CDI Market • Higher Ed (research) • Fed/Govt • BIO/Healthcare • Fintech/Crypto • Manufacturing • Media & Entertainment Storage Fabrics Market • Industries • Media & Entertainment • Workloads • GPUs Consumers • Database Applications • Edge • AI/ML • HPC
44. © 2022 Liqid, Inc. Proprietary & Confidential 44 Durham University ““Durham University is using cutting-edge CDI to accelerate research, improve resource utilisation, and reduce the university’s carbon footprint” Alastair Basden, Technical Manager DiRAC Memory Intensive Service at Durham University What: • The COSmology MAchine (COSMA8), an HPC system for studying the origin and development of the universe Challenge: • Needed to increase speed and efficiency of COSMA8. • Improve utilisation of GPUs Why Liqid • Maximize utilisation of small number of GPUs • Populate servers with GPUs with the click of a button
45. © 2022 Liqid, Inc. Proprietary & Confidential 45 Deployment: ERDC - Liqid Awarded as Prime Contractor - Consumes over 85K CPU cores & ~600x GPUs - Liqid was Selected as the Prime Contractor - Listed on Top 200 List for Supercomputers Deployment: ARL - Liqid Awarded as Prime Contractor - Consumes over 100K CPU cores & ~300x GPUs - Liqid was Selected as the Prime Contractor - Listed on Top 200 List for Supercomputers https://blocksandfiles.com/2020/09/24/compose-this-liqid-bags-another-supercomputer-deal/
46. © 2022 Liqid, Inc. Proprietary & Confidential 46 Orange • Top TelCo provider in EMEA … over 300,000 employees worldwide • Goal is to deliver faster but also smarter and more efficient Edge Service • First Goal: Replace cloud-based AI workloads with On-Prem deployments • Second Goal: Offer GPU as a service for Edge deployments … premium product offering • Dynamic edge pod will enable the flexibility needed to meet the most demanding SLAs
47. © 2022 Liqid, Inc. Proprietary & Confidential 47 Competitive Landscape Traditional Infrastructure GPU Optimized Servers Other Composable Vendors
48. © 2022 Liqid, Inc. Proprietary & Confidential 48 Roadmap
49. © 2022 Liqid, Inc. All rights reserved. 49 Proprietary & Confidential • Multi-Fabric Support PCIe Gen-4, Gen-5, CXL, Ethernet, IB • Multi-Device Support CPU, GPU, FPGA, Storage, Networking • Advanced Protocol Support NVMe-oF and GPU-oF • Software-only Composability Compose devices directly from existing servers • Secure Multitenancy Support Role based access control and advanced security policy • Northbound Integration Compose via Slurm, VMware, K8s • P2P Direct Performance SSD, NIC, GPU, & CPU • Multi-cloud Integration Multi-site Multi-pod composability • Fabric Management & Telemetry Single Pane of Glass UI, CLI, and API • Bare-metal and Virtualized Compose for multiple workload types Increase Flexibility Improve IT Agility Unlock New Levels of Datacenter Efficiency Liqid Matrix Software Features
50. © 2022 Liqid, Inc. All rights reserved. 50 Proprietary & Confidential Composing Resources for Workloads Item Bare-metal Virtual Machines Containers Management / Control Plane Resource Composability GPU, FPGA, NVMe-oF, Networking, Memory GPU, FPGA, NVMe-oF, Networking, Memory GPU, FPGA, NVMe-oF, Networking, Memory Workload / Resource Unit Bare-metal Servers Virtual Machines Containers Composed Unit Bare-metal Servers Hypervisor Servers K8s Cluster Nodes Liqid Integration Point Liqid Slurm Plug-in* Liqid vCenter Plugin* Liqid Openstack Integration** Liqid K8s Integration** Integration Functions Modify server resources at server cluster creation. Modify resources in response to new workload needs. Modify server resources at hypervisor cluster creation. Modify resources in response to new VM needs. Modify server resources at K8s cluster creation. Modify resources in response to container needs. *available today; **roadmap item
51. © 2022 Liqid, Inc. All rights reserved. 51 Proprietary & Confidential PCIe Interconnect Evolution *Bandwidth is for x16 lanes, bi-directional **Transfer Rate is GT/s x1 lane Compute Express Link (CXL) • Provides coherency and memory semantics on top of PCIe Gen 5.0 • High bandwidth/low latency between host processors and devices • Accelerators • Memory buffers • Smart I/O devices PCIe Gen 3.0 • Bandwidth: 32GB/s • Transfer Rate: 8GT/s PCIe Gen 4.0 • Bandwidth: 64GB/s • Transfer Rate: 16GT/s PCIe Gen 5.0 • Bandwidth: 128GB/s • Transfer Rate: 32GT/s Intel Sapphire Rapids & AMD Genoa Servers likely in 1H-2023
52. © 2022 Liqid, Inc. Proprietary & Confidential 52 CXL Technology Preview Demo • What are we demonstrating? • Liqid Matrix software updated to support CXL fabric type and CXL connected memory endpoints • CXL connected hosts • CXL fabric interconnects • CXL connected memory devices • Results of the demonstration: • Integrate CXL fabric-type into the Liqid Matrix software for composable infrastructure • Utilize a CXL fabric for memory composition to a host • Demonstrate the disaggregation of DRAM from the CPU complex • Proof of concept for end-to-end composable memory from host to memory end-point device
53. © 2022 Liqid, Inc. Proprietary & Confidential 53 Liqid Enables Composable Memory Storage NVMe Flash / Optane CPU Intel / AMD / ARM Memory DRAM Accelerators GPU / FPGA / DPU Dynamically Configured Servers Disaggregated Resource Pools CXL Fabric Liqid Mgmt. SW
54. © 2022 Liqid, Inc. Proprietary & Confidential 54 Composable Memory Management New Device-type for CXL memory New Fabric-type for CXL memory
55. © 2022 Liqid, Inc. All rights reserved. 55 Proprietary & Confidential Composable Memory for AI & HPC
56. © 2022 Liqid, Inc. Proprietary & Confidential 56 Backup
57. © 2022 Liqid, Inc. Proprietary & Confidential 57 Gen4 4U 10x FHFL Double-width Gen4 4U 20x FHFL Single-width Gen3 1U 32x U.2 Drives Model Number LQD400x10 LQD400x20 LQD300x32 PCIe Generation Gen 4 Gen 4 Gen 3 Connectivity PCIe PCIe PCIe or Ethernet Device Types Supported GPU | FPGA| NIC | AIC SSD FHFL GPU | FPGA| NIC | AIC SSD FHFL SSD Device Quantity 10x Double Wide 20x Single Wide 32x 2.5" Device Interface PCIe Gen 4.0 x16 PCIe Gen 4.0 x8 PCIe Gen 3.0 x16 Interface Ports 4x PCIe Gen 4.0 x16 4x PCIe Gen 4.0 x16 PCIe: Up to 4x Gen 3.0 x16 Eth: Up to 2x 200GbE NICs Bandwidth - Port Up to 64 GB/s – Full Duplex Up to 64 GB/s – Full Duplex Up to 32GB/s – Full Duplex Bandwidth - Total Up to 256 GB/s – Full Duplex Up to 256 GB/s – Full Duplex Up to 128GB/s – Full Duplex PCIe Fabric Latency 100 ns 100 ns 150 ns Liqid Management Port 2x PCIe Gen 4.0 x1 ports 2x PCIe Gen 4.0 x1 ports N/A Power - Chassis 4 x 2400W 4 x 2400W Dual 1000W w/PMBus PCIe Slot Power Mgmt. Yes Yes N/A Liqid Expansion Chassis Portfolio
58. © 2022 Liqid, Inc. Proprietary & Confidential 58 HHHL SSD FHFL SSD HHHL Intel Optane FHFL Intel Optane Model Number LQD3000 LQD4500 LQD3900 LQD4900 Type NVMe Flash TLC 3D NAND Intel Optane Intel Optane Raw Capacity Up to 16TB Up to 32.7TB Up to 1.5TB 3TB Read Bandwidth (256K) 7.0GB/s 24.0GB/s 7.0GB/s 21.7 GB/s Write Bandwidth (256K) 6.3GB/s 24.0/GBs 7.0GB/s 19.4 GB/s Ran. Read IOPS (4K) 1.25M 4M 1.6M 4.1M Ran. Write IOPS (4K) 900K 4M 1.6M 3.7M Read Latency (4K, QD=1) 80 μs 80 μs 10 μs 15 μs Write Latency (4K, QD=1) 20 μs 80 μs 10 μs 15 μs Protocol NVMe 1.2.1 NVMe 1.3 NVMe 1.0 NVMe 1.3 Bus Interface PCIe 3.0 x8 PCIe 4.0 x16 PCIe 3.0 x8 PCIe 4.0 x16 Endurance Up to 31 PBW Up to 62 PBW Up to 164 PBW 328 PBW Liqid NVMe SSD and Storage Class Memory Devices
59. © 2022 Liqid, Inc. Proprietary & Confidential 59 Model Number LQD9448 LQD9424 Ports 48 Ports 24 Ports Protocol PCIe Gen 4.0 Managed Switching Fabric PCIe Gen 4.0 Managed Switching Fabric Port Bandwidth 16GB/s (Full Duplex) per x4 Port 16GB/s (Full Duplex) per x4 Port Total Bandwidth 768GB/s (Full Duplex) 384GB/s (Full Duplex) Port Latency 100ns 100ns Port Configuration x4, x8, or x16 (SRIS Ready) x4, x8, or x16 (SRIS Ready) Cable Connections Photonics and/or Copper over MiniSAS HD Cabling Photonics and/or Copper over MiniSAS HD Cabling Management Secured, 1 GBASE-T Secured, 1 GBASE-T Failover N+1 Support via Multi Switch N+1 Support via Multi Switch Scalability Supported Supported Form Factor 1U 1U Liqid PCIe Fabric Switches
60. © 2022 Liqid, Inc. Proprietary & Confidential 60 PCIe Gen 3.0 HBA PCIe Gen 4.0 HBA U.2 to PCIe Gen 3.0 HBA Model Number LQD1316 LQD1416 LQD1325 Form Factor PCIe x16 HHHL PCIe Gen4x16 HHHL PCIe x4 U.2 Adapter Dimensions 6.14″ x 2.89″ (156mm x 73.4mm) 6.14″ x 2.89″ (156mm x 73.4mm) 4.42”x.63”x2.93” Power Maximum 15W Maximum 15W 10 Watts Max Connectors SFF-8644 or SFF-8643 SFF-8644 or SFF-8643 SFF-8808 Operating Temp 0C to 50C 0C to 50C Storage Temp -40C to 85C -40C to 85C -40 to 75C Operating Humidity 10% to 90% RH non-condensing 10% to 90% RH non-condensing 5% to 95% RH non-condensing Storage Humidity 5% to 95% RH non-condensing 5% to 95% RH non-condensing 5% to 95% RH non-condensing PCB PCI Express add-in card standard thickness .063″ +/- 0.008″ (1.6mm +/- 0.2mm) PCI Express add-in card standard thickness .063″ +/- 0.008″ (1.6mm +/- 0.2mm) Standard U.2 PCIe Switch PLX PEX9733 Capella 2, 8.0 GT/s 32-lane PCIe Gen 3 Switch, DMA Controller, SSC Isolation Support Broadcom PEX88032, 16.0 GT/s 34-Lane PCIe Gen4 Switch, DMA Controller, SSC Isolation Support Liqid PCIe Host Bus Adapters
61. © 2022 Liqid, Inc. All rights reserved. 61 Proprietary & Confidential Unlock Trapped Performance With Liqid P2P NVIDIA A100 GB/s w/o P2P 9 GB/s w/P2P 49 0 10 20 30 40 50 60 GB/s Bandwidth (GB/s) NVIDIA A100 Latency w/o P2P 25 Latency w/P2P 3 0 5 10 15 20 25 30 Latency Latency (µs) 474%  89% 
62. © 2022 Liqid, Inc. All rights reserved. 62 Proprietary & Confidential Accelerate GPU to Storage Performance 0 500,000 1,000,000 1,500,000 2,000,000 2,500,000 3,000,000 3,500,000 4K RR 4K RW IOPS w/o GDS 179,000 203,000 IOPS w/GDS 2,900,000 2,700,000 IOPS I/O Per Second (IOPS) 0 100 200 300 400 500 600 700 800 4K RW 4K RR Lat w/o GDS 583 729 Lat w/GDS 80 112 Latency Latency (µs) 0 2 4 6 8 10 12 14 16 18 128K W 128K R GB/s w/o GDS 12 11.7 GB/s w/GDS 11.8 16.5 Bandwidth Bandwidth (GB/s) Run on LQD4500 in RAID0 with 128 workers ./gdsio -f /dev/md127 -d 0 -s 10G -i 128k -I 1 -T 60 -x 2 -w 128 ./gdsio -f /dev/md127 -d 0 -s 10G -i 4k -I 3 -T 60 -x 0 -w 128 1,520%  1,230%  41%  1.2%  85%  86% 
63. © 2022 Liqid, Inc. All rights reserved. 63 Proprietary & Confidential Data Traversing CPU/RAM Impacts Performance • Data flows through x86 Root Complex • Bottlenecks Throughput • Injects Latency CPU Memory PCIe Fabric GPU SSD
64. © 2022 Liqid, Inc. All rights reserved. 64 Proprietary & Confidential Eliminate Costly Data Bottlenecks with Liqid Host CPU Memory PCIe Fabric GPU SSD NIC • Information flows between devices on PCIe bus • Improves throughput • Minimizes latency
65. © 2022 Liqid, Inc. Proprietary & Confidential 65 Liqid Composable Elements NVMe Storage Fabric Integrated PCIe Expansion CDI Software Liqid EX 4400 Series 10/20 Slot UI | API | CLI 4U GPU | FPGA | SSD | NIC Liqid NVMe PCIe Storage SSD | Storage Class Memory FHFL AIC HHHL AIC
66. © 2022 Liqid, Inc. Proprietary & Confidential 66 LQD3000: Gen3 NVMe SSD Vendor A Liqid Seq Read 128KB 6,170 7,200 Seq Write 128KB 2,200 5,800 Rnd Read 4KB 1,200,000 1,500,000 Rnd Write 4KB 200,000 500,000 LQD3000 Product Details Product: Gen3 NVMe SSD AIC Host Interface: Gen3x18 Form Factor: HHHL AIC Power Loss Protection: Yes Max Capacity: 15.4TB
67. © Copyright 2020 Dell Inc. 67 of Y MX7000 & Liqid Enables PCIe Expansion (PCIe) Dell EMC PowerEdge MX7000 Modular Chassis Liqid PCIe Expansion Chassis Up to 20x FHFL GPU PCIe Gen3x4 U.2-to-PCIe Adapter Accelerator Expansion Over PCIe GPU Chassis JBOx PCIe Expansion Chassis (“Just-a-Bunch-of-GPU” or Other Devices) Interconnect PCIe Gen3x4 (Multiple G3x4 Links Possible) GPU Expansion 20x GPU (FHFL) Double Wide GPU Supported V100, A100, RTX, T4, and Others OS Supported Linux, Windows, VMware Devices Supported GPU, FPGA, and NVMe Form Factor 14RU (7RU MX7000 + 7RU PCIe JBOx) Command Center Software
68. © 2022 Liqid, Inc. Proprietary & Confidential 68 6/10/2022 68 Technology Leadership: ~7 years of software code and feature development – major first mover advantage Strategic Partnership: Strong strategic partnerships with OEMs, technology providers, and distribution channel Customer Traction: Significant customer adoption – strong product market fit for core technology Patent Portfolio: >85 Patents Filed and >55 Patents Issued Liqid Market Leadership Competitive Advantage
69. © 2022 Liqid, Inc. Proprietary & Confidential 69 Thank You