Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
The journey to Memory-Driven Computing
– Enrique Matorras SER CME Chief Tecnologist
– July 3rd, 2019
Agenda
2
HPE Memory driven approach
Gen-Z in the path to ultimate composability
HPE memory driven journey
The need
The needData Explosion
3
If you thought you were having trouble keeping up today, it’s only going to get worse
Exponentially-increasing data
But conventional compute improving only incrementally
Growing capability gap
DATACOMPUTE
2006
COMPUTE
DATA
2008 2010 2012 2014 2016 2018 2020
0.3ZB
0.8ZB
1.2ZB
1.8ZB
4.4ZB
7.9ZB
15.8ZB
31.6ZB
44ZB
Explodingdata sourcesX X Shrinking
time to action
Massive advancesin computing power
needed everywhere=
What’s driving the data explosion? Structured data
40 petabytesWalmart’s transaction
database (2017)
Human interaction data
4 petabytes a dayPer-day posting to Facebook
across 2 billion users (2017)
2MB per active user
Digitization of analog reality
40,000 petabytes a day*10m connected cars by 2020
Front camera20MB / sec Front ultrasonic sensors
10kB / secInfrared camera
20MB / sec
Side ultrasonic sensors
100kB / sec
Front, rear and top-view cameras
40MB / sec
Rear ultrasonic cameras
100kB / secRear radar sensors100kB / sec
Crash sensors100kB / sec
Front radar sensors
100kB / sec
* Driver assistance systems only
The New Normal: Compute is not keeping up
6
0,3 0,8 1,2 1,84,4
7,9
15,8
31,6
44
0
5
10
15
20
25
30
35
40
45
50
2006 2008 2010 2012 2014 2016 2018 2020
Data(Zettabytes)Data nearly doubles
every two years(2013-2020)
Data growthTransistors(thousands)
Single-threadPerformance(SpecINT)
Frequency(MHz)
Typical Power(Watts)
Number of Cores
1975197719791981198319851987198919911993199519971999200120032005200720092011201320152017
107
106
105
104
103
102
101
100
Microprocessors
Source: K. Rupp. 42 Years of Microprocessor Trend Data Source: Data Age 2025 study, sponsored by Seagate, April 2017
HPE Composable approachGen-Z the path to ultimate composability
7
Infrastructure is stretching to it’s limits
8
Circa 2008Circa 2019
Two socket server architecture is stretching to its limits
• Relentless drive for performance, greater core counts
• Processors are more complex than ever§ DDR data rates and DDR channel counts are increasing …
§ … but DIMMs per channel are decreasing
§ PCIe data rates and lane counts are increasing
§ Storage data rates and lane counts are increasing
• More complex system designs drive up costs
PCIe Gen 3
PCIe Gen 4PCIe
Gen 5
DDR3
DDR4
DDR5
SATA Gen 2SAS Gen 3
Inter-Processor Links(many variants)
SATA Gen 3
SATA Gen 4
SAS Gen 4NVMeCC
IX
NVLinkOpenCAPI
PCIe Gen 2
SAS Gen 2
Protocols on the horizon (in orange)
Infrastructure is stretching to it’s limits
9
Infrastructure networks & fabrics have challenges
• Still maintain a wide variety of networks/fabrics § Each has its special purposes
• None are designed to support native CPU load/store§ To support memory/storage convergence
PCIe Gen 4PCIe
Gen 5
DDR3
DDR4
DDR5
SATA 6GSAS 12G
Inter-Processor Links(many variants)
SATA 12G
SATA 24G
SAS 24GNVMeCC
IX
NVLinkOpenCAPI
Ethernet
The Tower of Protocol Babel
Omni-PathInfiniBandFibre Channel
PCIe Gen 3PCIe Gen 2
SAS 6G
The Solution:
10
Memory Semantic Fabric (language of compute)
Gen-Z!
IT evolution towards composability
11
Appl
icat
ion
optim
ized
Operations optimized
Improve staff productivity– Masks some complexity
with people and software– Preconfigured physical IT– Hardware defined– Targeted workloads
Simplify deployments– Remove complexity by
eliminating SAN– Fluid virtual IT– Software defined storage– Virtual workloads
Simplify Infrastructure as a Service– Hardware & Software
architected as one– Fluid IT – Software defined everything– Physical, virtual and
containerized workloads
Siloed Infrastructure– Complex processes– Static, siloed IT– Silo defined– Physical, virtual, and
containerized workloads
Traditional
Converged
Hyper-converged
Composable
Composability
This document contains confidential information of Hewlett Packard Enterprise © Copyright 2018 Hewlett Packard Enterprise Development LP
HPE and Channel Partners - Internal Use Only
DEVDEVDEV
TESTTEST
ANALYTICS
HANAHANAHANAHANAHANAHANA
VDI
Oracle
App Dev/Test environmentis needed now that peak season is over
Modeling & Analytics Run at night
SAP HANA for running production mission-critical workloads
VDI / CAD applications run during the day
and Oracle database
HPE Synergy composable infrastructureToday’s photonic ready solution for your traditional and cloud-native workloads
13
HPE Memory driven approach
14
We are radically rethinking our approach to computingAdvancing computing without relying on Moore’s Law
Unconventional accelerators
Unconventional architectures
Unconventional programming
Unconventional architectures
16
GPU DSP
x86A
SIC
Phot
onic
Neuro
Quantum
POWER
SPARC
RIS
CV
ARM
FPGA
Memory+
Fabric
MemoryMemory
Mem
ory
Mem
oryM
emory
Memory
MemoryMemory
Mem
ory
Mem
ory
Mem
ory
Memory
SoC SoC
SoCSoC
SoC
SoCSoC
SoC
SoC
SoC
SoC
SoC
Future architectureMemory-Driven Computing
Today’s architecture From processor-centric computing
17
Traditional vs. Memory-Driven Computing architecture
18
Today’s architectureis constrained by the CPU
DDR
Ethernet
PCI
If you exceed the what can be connected to one CPU, you need another CPU
Memory-Driven Computing:Mix and match at the speed of memory
SATA
Memory-Driven Computing delivers ultimate composability
19
Processors Memory
Storage Class Memory/Persistent MemoryNon-volatile, fast
FlashNon-volatile, but slow
DRAMVolatile, but fast
Accelerators
GPUAccelerates images manipulation
FPGAArray of logic gates that can be hardware-
programmed to fulfill user-specific tasks
Dot Product Engine/Neuromorphic Computinghigh performance per-watt task-specific
analog accelerator computing using
Memristor Arrays
Chaos ComputingSupercharged version of the
Dot Product Engine, creating an
Analog Computation Engine (ACE)
to take advantage of chaos
principles for tackling a class of
non-deterministic compute problems
Optical ComputingComputing circuits operated solely
using light
Logical systems composed of physical components
Or subparts or sub regions of
components (e.g. memory/storage)
Logical systems match exact workload requirements
No stranded resources
overprovisioned to workloads
Facilitates data-centric computing via shared memory
Reduces data movement
Do more with less, reduces cost
x86
Arm
RISC-V
Fabric
Specialized processors designed to perform
specific functions more efficiently
Memory-Driven Computing is the future for every kind of computing
21
Edge deviceCloud
infrastructureBig memory
machineExascalecomputer
‒ Near-zero power‒ Persistent memory‒ AI task-specific
accelerator
‒ Composable infrastructure from every edge to any cloud
‒ Microservices in microseconds at massive scale
‒ High-performance data analytics
‒ Large shared memory
‒ Monte Carlo, graph analytics applications, etc.
‒ 100,000+ components‒ Ultra-fast message passing
and checkpointing
‒ 20x more energy-efficient than state-of-the-art
Fabric
Photonic interconnects “destroy distance” and makememory-speed communications practical over multi-kilometer distances
How big can Memory-Driven Computing go?
22
4,096 yottabytes
292bytes
Memory space:
16 million devices
224addresses
Fabric size:
1,600 exascale computers
270FLOPs
Processing power:
Unconventional programming
23
Transform performance with Memory-Driven programming
25
In-memory analytics
15xfaster
New algorithms Completely rethinkModify existing
frameworks
Similarity search
40xfaster
Financial models
10,000xfaster
Large-scalegraph inference
100xfaster
DZNE discovered HPE’s Memory-Driven Computing — and saw unprecedented computational speed improvements that hold new promise in the race against Alzheimer’s
60% power reduction cuts research costs
101x increase in analytics speed blasts research bottlenecks, leading to shorter processing time — from 22 minutes to
13 seconds
Memory-Driven Computing helps outpace the global time bomb of neurodegenerative disease
Unconventional accelerators
27
Beyond General Purpose. Beyond Moore’s LawUnconventional accelerators designed for Memory-Driven Computing
28
Analog neuromorphic computing Massive speedup for AI training and inference
Optical ComputingDesigned for “unsolvable” optimization problems
– Harnessing the properties of light at the microscale
– Aims to solves NP-hard problems in milliseconds– Airline optimization– Logistics networks– Vehicle routing– Graph analytics
– Prototype has world record1,000 optical components
– Easily scalable to100,000 components
– Complex matrix calculations in one step instead of thousands
– 10-100x faster
– 10-1000x more energy efficient(Compared to GPU)
– Simple to program
– Simple to manufacture
Memory-Driven Computing – built for accelerators
Combining memory and storage in a stable environment to increase processing speed and improve energy efficiency
Using photonics where necessary to eliminate distance and create otherwise impossible topologies
Optimizing processing from general to specific tasks
Radically simplifying programming and enabling new applications that we can’t even begin to build today
Fast, persistentmemory
Fast memoryfabric
Task-specificprocessing
New and adaptedsoftware
29
HPE memory driven journey
30
Accomplishments and highlightsDelivering proof points, capturing our customers’ attention and gaining traction across the industry
November 2016Memory-Driven Computing proven
2017November 2017DZNE gets 100x faster results– The Machine User Group 600+ users– Continued leadership in IEEE
Rebooting Computing initiative
January –March 2017The Machine prototype scales up
June 2017Exascale ComputingDoE PathForward award
2017Memory-Driven Computing principles across HPE portfolio– HPE ProLiant Gen 10
servers with Silicon Root of Trust
– HPE Persistent Memory– HPE Superdome Flex
January 2018Gen-Z major milestones– Spec 1.0 release– 52+ consortium
members
June 2018Memory-Driven
Computing Sandbox– 48TB of shared memory
Pointnext Memory-Driven Computing capabilities
– Enabling proof-of-concept use cases with customers
April 2018Catalyst UK program– HPE Apollo 70 HPC with ARM– Partnership with 3 UK
universities, ARM, Cavium, SuSE
Vanguard Astra- Apollo 70 HPC- World’s Largest ARM
Supercomputer
2016 2018
June 2017First collaboration announced with DZNE
May 2017World’s largest single-memory computer system announced– The Machine prototype with
160 TB of shared memory– 1280 ARM cores
Real-timeEnterprise
High-performance Data analytics
SAP HANAin-memory
Project “Kraken”Co-innovation with SAP
Superdome X
Memory-Driven ComputingThe ultimate composable infrastructure
Extreme scalabilityfor transactions & analytics
Superdome FlexDesigned with Memory-Driven Computing principles
HPE is leading the industryFrom in-memory to Memory-Driven Computing
Memory-Driven Computing prototype
160TBThe world’s biggest single memory computer
HPE: your right mix today, your right mix tomorrow
Memory-Driven Computing bridges the widening gap between our goals and our capabilities.
Address tomorrow’s AI-driven workloads much faster, using much less energy, than conventional systems.
Memory-Driven Computing is the ultimate composable infrastructure from every edge to any cloud
Memory-Driven Computing will power the data-driven future