+ All Categories
Home > Documents > Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC...

Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC...

Date post: 31-Dec-2020
Category:
Upload: others
View: 1 times
Download: 0 times
Share this document with a friend
24
1 | ©2014 Micron Technology, Inc. Todd Farrell Senior Member Technical Staff Computing and Networking BU [email protected]
Transcript
Page 1: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

1 | ©2014 Micron Technology, Inc.

Todd Farrell

Senior Member Technical Staff Computing and Networking BU

[email protected]

Page 2: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

2 | ©2014 Micron Technology, Inc.

Demand Drivers

Insatiable need for bandwidth

Impact of the cloud

Global demand for mobility

Big data analytics challenge

April 28, 2014

Annual Data Center IP Traffic 2012-2017 CAGR: 25%

Source: Cisco Global Cloud Index 2013

SOCs & Microservers

High-Performance Computing

Acceleration & Co-Processing

Enterprise Computing

0

2,500

5,000

7,500

10,000

2012 2013 2014 2015 2016 2017

Data Center to User

Data Center to Data Center

Within Data Center

EB/Yr

Page 3: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

3 | ©2014 Micron Technology, Inc.

0

15

30

45

60

75

90

2015 2014 2013 2012 2011 2010 2009 2008 2007 2006 2005 2004 2003 2002 2001 2000 1999 1998 1997

Traditional Memory Designs Do Not Scale and Drive Exponential Complexity

Historical System Bandwidth, Cost, and Complexity

April 28, 2014

1-2 Memory Channels Up to 6.4GB/s

85 page specification <3 Yrs to Standardize

DDR

2 Memory Channels Up to 10.7GB/s

3 Speed Bins

DDR2

3-4 Memory Channels 240 Pins Per DIMM

Up to 59.7GB/s 5 Speed Bins

DDR3

4 Memory Channels 284 Pins Per DIMM

Up to 85GB/s 12 Speed Bins

6+ yrs to Standardize

DDR4

Bandwidth Per Memory Channel

2012–2017 CAGR: 12.3%

GB/s

Page 4: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

4 | ©2014 Micron Technology, Inc.

Major Challenges to the Longevity of DRAM and NAND Technologies

Memory Process Nodes Over Time

Physical process limits?

Endurance

•Cycling capabilities

Shrink

•Physical limitations

Equipment capability

•Increasing equipment complexity

Cost

•Path of diminishing returns

-

0.020

0.040

0.060

0.080

0.100

0.120

2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018

DRAM Industry NAND Industry

New memory

April 28, 2014

Page 5: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

5 | ©2014 Micron Technology, Inc.

HMC - A Revolutionary Shift

Increased Bandwidth

Greater Power Efficiency

Lower TCO

Reduced System Latency

Smaller, Scalable and

Flexible

April 28, 2014

Page 6: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

6 | ©2014 Micron Technology, Inc.

High-Performance Memory Comparison

April 28, 2014

Requirements TCO Valuation

Bandwidth

Energy Efficiency

Board Footprint

Channel Complexity 90% simpler than DDR3L 88% simpler than DDR4

95% smaller than DDR3L 94% smaller than DDR4

66% greener than DDR3L 55% greener than DDR4

10.2X greater than DDR3L 8.5X greater than DDR4

Single-Link HMC vs. DDR3L-1600 and DDR4-2133 What does it take to support 60 GB/s?

0 250 500 750

HMC

DDR4

DDR3L

pins

0 3,000 6,000 9,000

HMC

DDR4

DDR3L

mm2

0 20 40 60

HMC

DDR4

DDR3L

pJ/b

0 300 600 900

HMC

DDR4

DDR3L

MB/pin

Page 7: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

7 | ©2014 Micron Technology, Inc.

Innovative Design & Process Flow

Incorporates thousands of TSV sites per die to reduce signal lengths and power

Enables memory scalability through parallelism

Sophisticated Package Assembly

Provides higher component density and significantly improves signal integrity

Enabling Technologies

Memory Vaults vs. DRAM Arrays

Significantly improves bandwidth, quality, and reliability vs. traditional DRAM technologies

Logic Base Controller Reduces memory complexity and

significantly increases performance

Allows memory to scale with CPU performance

Abstracted Memory Management

Through-Silicon Via (TSV) Assembly

April 28, 2014

Page 8: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

8 | ©2014 Micron Technology, Inc.

Traditional Host Processor Memory Management

HOST

DRAM Layer

DRAM Layer

DRAM Layer

DRAM Layer

TSV

TSV

TSV

TSV

TSV

TSV

TSV

TSV

Re-drive Layer

Si Interposer

Manufacturing Test

Burn-in

At-speed functional

Manage field maintenance and

self test

Manage all present and future DRAM scaling and process variation issues

Manage 100+ different DRAM timing parameters

Non-Managed DRAM (DDR, WIO2, HBM, etc.)

April 28, 2014

Page 9: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

9 | ©2014 Micron Technology, Inc.

Simple HMC Memory Management

HOST

DRAM Layer

DRAM Layer

DRAM Layer

DRAM Layer

TSV

TSV

TSV

TSV

TSV

TSV

TSV

TSV

Re-drive Layer

Si Interposer

Manufacturing Test

Burn-in

At-speed functional

Manage field maintenance and

self test

Manage all present and future DRAM scaling and process variation issues

Manage 100+ different DRAM timing parameters

Simple memory requests and responses;

No DRAM timings or scaling issues to manage

Functions moved to HMC for management

Non-Managed DRAM (DDR, WIO2, HBM, etc.)

April 28, 2014

Page 10: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

10 | ©2014 Micron Technology, Inc.

HMC Architecture

Vaults are managed to maximize overall device availability -

Optimized management of energy and refresh

Self test, error detection, correction, and repair in the logic base layer

3DI & TSV Technology

DRAM0 DRAM1

DRAM2

DRAM3 DRAM4

DRAM5

DRAM6 DRAM7

Logic Chip

Logic Base

Multiple high-speed local buses for data movement

Advanced memory controller functions

DRAM control at the memory rather than at distant host controller

Reduced memory controller complexity and increased efficiency

DRAM

Logic Base

Vault

Memory Control

Vault Control

DRAM Sequencer

Crossbar Switch

Write Buffer

Read Buffer

Refresh Controller R

equest

Write

Data

Read D

ata

DRAM Repair

TSV Repair

Detail of Memory Interface

Vault Control Vault Control Vault Control

Memory Control

Crossbar Switch

Link Interface Controller

Link Interface Controller

Processor Links

Link Interface Controller

Link Interface Controller

Logic Base

April 28, 2014

Page 11: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

11 | ©2014 Micron Technology, Inc.

HMC Architecture Link Controller Interface

April 28, 2014

RX

TX

Host

TX

RX

HMC

16 Lanes

16 Lanes

8 or 16 Transmit Lanes

8 or 16 Receive Lanes

Example:

HMC-SR Options: 10 Gb/s, 12.5 Gb/s,

or 15 Gb/s Crossbar Switch

BIST

Processor Links

Link Interface Controller

Link Interface Controller

Logic Base

Vault Control Vault Control Vault Control Vault Control

Memory Control

Memory Control

Memory Control

Memory Control

Page 12: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

12 | ©2014 Micron Technology, Inc.

Packet-Based Communication

Packets comprised of 128-bit (16-byte) FLITs

Packets include 1 to 9 FLITs, depending on command

Host issues requests & HMC issues responses

Each packet contains 64-bit header and 64-bit tail (1 FLIT)

April 28, 2014

Protocol NOT affected by any DRAM-related timings, nor is it

DRAM-specific!

Multiple data transfer sizes supported (16B to 128B)

Commands include reads, writes, atomics, error responses

Simultaneous READs and WRITEs supported

Page 13: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

13 | ©2014 Micron Technology, Inc.

HMC Reliability Built-in RAS features

Logic / Interface

DRAM Array

Logic / Interface

DRAM Array

Logic / Interface

DRAM Array

Logic / Interface

DRAM Array

Host Vault data ECC-protected

Address/command parity for array

transactions

CRC protection on link interface

Link retry

Logic stability (DRAM controls in logic)

Reliable handshake (packet integrity verified before memory access)

April 28, 2014

Page 14: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

14 | ©2014 Micron Technology, Inc.

RAS Feature System Comparison

April 28, 2014

Page 15: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

15 | ©2014 Micron Technology, Inc. April 28, 2014

HMC Standard Packages

Standard BGA packaging solutions: Cost-effective packaging using existing ecosystems

MCM and In-Package options available

Up to 1.28 Tb/s memory bandwidth available TODAY!

Page 16: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

16 | ©2014 Micron Technology, Inc.

HMC Maximum Bandwidth Configurations

April 28, 2014

All links between host CPU and HMC logic layer

▶Maximum bandwidth per GB capacity:

HPC/Server – CPU/GPU

Graphics

Networking systems

Test equipment

Page 17: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

17 | ©2014 Micron Technology, Inc.

HMC Scalable Memory Solutions

April 28, 2014

Far Memory:

Some HMC links connect to host – some to other cubes

Scalable to meet system requirements

Available in module form or soldered-down

Building blocks for multiple application needs

Future Products May Include:

Higher-speed electrical (SERDES) VSR-30

Optical interfaces (align to industry stds.)

Higher stack count for greater capacity

Non-DRAM memory technologies

Additional Atomic Operations inside Cube

Page 18: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

18 | ©2014 Micron Technology, Inc.

Hybrid Memory Cube

Micron Memory Innovation We’ve combined fast logic process technology and advanced DRAM designs to create an entirely new category of memory. Hybrid Memory Cube (HMC) technology provides a high-bandwidth, low-energy, high-density memory system that’s unlike anything on the market today. . Unprecedented Performance

HMC will provide a revolutionary performance shift that will enrich next-generation networking and enable exaflop-scale supercomputing:

Reduced Footprint 90% less space than today’s RDIMMs

Reduced Power Fraction of the energy per bit

* HMC SR-15G vs. DDR3-1333

Increased Bandwidth 15X the performance of DDR3*

April 28, 2014

Page 19: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

19 | ©2014 Micron Technology, Inc.

“…like adding a turbocharger to your computer”

- datacenteracceleration.com

“…unprecedented levels of memory performance”

- Electronic News

Industry Validation

“…get ready for some serious bandwidth to hit us in

the near future”

- tweaktown.com

EE Times 40th Anniversary: “one of the top ten technologies expected

to redefine the industry”

April 28, 2014

Page 20: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

20 | ©2014 Micron Technology, Inc.

Consortium Momentum

http://www.hybridmemorycube.org

April 28, 2014

Page 21: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

21 | ©2014 Micron Technology, Inc.

Milestones

April 28, 2014

2 0 1 1 2 0 1 2

2/11 - Micron announces the

development of HMC technology

10/11 - Micron and Samsung launch the HMC Consortium with fellow developers Altera, Open-

Silicon, and Xilinx

12/11 - HMC wins Linley Group’s Analyst Award

for Best New Technology of 2011

12/11 - IBM joins the HMCC as a developer

9/11 - Micron and Intel demonstrate Micron’s HMC Gen1

device and an Intel MIC CPU memory interface at IDF 2011

9/13 – Micron sampling 2GB HMC device

5/12 - ARM and SK Hynix join the HMCC as

developers

9/13 - Micron and Altera demo industry’s first FPGA and HMC

interoperability

4/13 – HMCC releases final HMC Spec 1.0 to

public 5/12 - HMC Spec 1.0 released to adopters for

review

2 0 1 3 2 0 1 4

4/13 – HMC named Memory Product of the Year

2/14 - HMC Spec 2.0 released to adopters for

review

11/13 – HMC part of Fujitsu next-gen supercomputer

12/13 – HMCC Developer Group membership final – ARM, Altera, IBM,

Micron, Open-Silicon, Samsung, SK Hynix and Xilinx

10/13 – HMCC adopter membership

tops 120

4/14 - HMC Stratix V Team named Design

Team of the year

Page 22: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

22 | ©2014 Micron Technology, Inc.

HMC - A Revolutionary Shift

Increased Bandwidth

Greater Power Efficiency

Lower TCO

Reduced System Latency

Smaller, Scalable and

Flexible

April 28, 2014

Page 23: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

©2014 Micron Technology, Inc. All rights reserved. Products are warranted only to meet Micron’s production data sheet specifications. Information, products, and/or specifications are subject to change without notice. All information is provided on an “AS IS”

basis without warranties of any kind. Dates are estimates only. Drawings are not to scale. Micron and the Micron logo are trademarks of Micron Technology, Inc. All other trademarks are the property of their respective owners.

23 | ©2014 Micron Technology, Inc.

Thank You!

Page 24: Todd Farrell - Los Alamos National Laboratory · Fujitsu next-gen supercomputer 12/13 – HMCC Developer Group membership final – ARM, Altera, IBM, Micron, Open-Silicon, Samsung,

Recommended