Evolution of GPU-Optimized Computing
SuperBlade®
GPU Solutions
Storage
Universal I/O Twin Architecture Double-Sided
Datacenter Optimized
2011 GPU Tech Taiwan Presented by Peter Yang Agenda Confidential
Overview: Super Micro Computer, Inc. Evolution of GPU Servers GPU/CPU Optimized System Advantages 1st Generation Solutions New Generation Solutions Successful Case Studies
2 Supermicro Overview Confidential
SMC Inc., HQ SMC BV, San Jose, CA The Netherlands
SMC TW, Taiwan
Founded in 1993, HQ– San Jose, CA / NASDAQ: SMCI
Revenues: FY09 $500M, FY10 $721M $1Bn run rate Global Footprint: >70 Countries Production: US, EU and Asia Production facilities Engineering: 70% of workforce in engineering (30% growth through recession) Market Share: #1 Server Channel (SMCI enables ~10% of global server market) Brand Equity: Growing public profile since 2007 IPO Corporate Focus: Energy Efficiency, Erath-friendly, Technology Innovation Supermicro’s Evolutionary Technology Confidential
Serverboards Chassis & Power Supplies The most configurable rack IPMI & System Management solution in the Industry – Blades & Storages GPUs & Networking
Supermicro’s organic growth path to complete building block solutions® Technology Progression & Partnership Confidential
The Possibility is The fastest 1U server Unlimited!!! in the world
Telsa S1070
New generation CPU & GPU supports 1U 4-GPU GPU Blades Standalone The first optimized CPU/GPU integrated 2U GPU / with QDR IB onboard PCI-E x16 PCI-E x16 Hybrid System 2U Twin What’s Next? ?
6016TT-TF SC827HD-R1400B 1U Twin™ Working with the right technology front view partner is the key! SW7046GT-TRF
The most powerful PSC Rear View
2008 2009 2010 2011 The Twin™ Architecture Confidential
Up to 2X Compute Density Two DP/UP server nodes in 1U form-factor Doubles Performance per unit volume Higher Efficiency - Shares Common Resources Without sacrificing performance Lower TCO Save Space (Room, Cables, Cabinet Hardware, etc.) Save Power (Efficiency, Cooling) Optimized for HPC Scaling
1U Twin™ GPU/CPU Optimized 1U 6 Optimized from Twin™: the GPU / CPU Servers Confidential
True PCI-E x16 Gen 2 bandwidth for both GPUs Tesla S1070
PCI-E x16
Optimized DP Serverboard Non-Blocking PCI-E x16 Gen 2 links Enterprise Server Management: I2C thermal control 1U Twin™ SuperServer 6016TT-TF Fan speed adjustment optimum cooling efficiency
Performance improved 110% with only 60% additional power e.g. 31% efficiency improvement @ >2x performance
7 1st Generation GPU Server Solutions Confidential
Model 1026GT-TF-FM207 6016GT-TF-FM207 6026TT-GTF 7046GT-TRF-FC407
Image
CPU Dual Xeon 5600/5500 Dual Xeon 5600/5500 Dual Xeon 5600/5500 Dual Xeon 5600/5500
Memory Up to 192GB in 12 DIMM Up to 192GB in 12 DIMM Up to 192GB in 12 DIMM Up to 192GB in 12 DIMM
HDD 6 x 2.5” 3 x 3.5” 12 x 3.5” 8 x 3.5”
2 GPU 2 GPU 2 GPU 4 GPU Expansion slot 1 LP PCI-e x4 1 LP PCI-e x4 1 LP PCI-e x4 3 PCI-e x4
IPMI LAN Dedicate Dedicate Dedicate Dedicate
Power Supply 1400W Gold 1400W Gold 1400W gold redundant 1400W gold redundant
GPU HPC cluster Personal Supercomputer / Cluster w/ IB interconnect: AoC for 1U 2GPUs / Scaling up onboard for 2U The PCI-E x16 can be used for high >2000 cores in one system w/ non-blocking capacity flash connectivity Next Generation GPGPU Solutions Confidential
SYS-1026GT-TRF-FM307 SYS-2026GT-TRF-FM407
Key features Key features Support up to 4 double width GPU Support 4 double width GPU cards. cards Optional riser card for 2 additional GPU cards. Platinum level redundant power supply Platinum level redundant power supply 4 HDD bays 10 HDD bays Smart server management tool Smart server management tool GPU SuperBlade® Confidential
Announced in SuperComputing 2010 Up to 120x Nvidia M2050, M2070 or M2090* and 120x Intel Xeon 5600 Series per 42U Rack Up to 60TFLOPS Theoretical Performance per 42U Rack Up to 10X performance vs CPU-only solution Maximum Performance Density per Watt Up to 5X better efficiency vs CPU-only solution Integrated non-blocking QDR-IB or 10G Ethernet switches with Dual IOH Save up to half the IB cables and associated costs
Unmatched Density!
IBM - 7U/7 GPU Dell - 10U/8 GPU *validation pending Supermicro – 7U/20 GPUs! GPU Brings HPC to a New Level Confidential
3 in the top 5 fastest computers on Top500 are using GPUs to accelerate the performance 9 in the top 20 greenest supercomputers on Green500 are using GPUs to achieve high MFLOPS per watt Within 50 GPU properly equipped nodes, the cluster can get on today’s Top500 list (> 31.1TFLOPS)
Rmax = 2.56PFLOPS
635.15MFLOPS/w
GFLOPS
Top500: #22 / Green500: #8 – 740.78MFLOPS/w Top500: #144 / Green500: #12 – 555.5MFLOPS/w Supermicro 2U Twin Supermicro 1U GPU Summary Confidential
Computation Intensive Applications GPGPU is the key to efficient HPC computing Best Density Best Performance per Watt Best Performance per Dollar Optimize Non-Blocking Architecture Optimize System Efficiency Choose the right system for your application
New 2U 4/6 GPU solution
Tesla GPU Card
New 1U 3 GPU solution GPU Blade GPU Workstation 20 GPUs in 7U! (Most in the World) Personal Super Computer
Source: NVIDIA Confidential
Thank You!