Terascale Memory Challenges and Solutions

By Dave Dunning, Randy Mooney, Pat Stolt, Bryan Casper and James E. Jaussi

December 6, 2010

Introduction

Modern computer architectures commonly include one or more CPUs, a cache or caches, a few DDR-based memory channels, rotational and/or solid state disks and one or more Ethernet ports.
 
Figure 1: System blog diagram
Figure 1: System block diagram

A high percentage of CPU-based systems use DDR-based DRAM for external memory. DDR-based DRAM currently provides very favorable cost/bit while providing enough bandwidth with low enough latency to meet the application demands. Although process engineers have continued to find ways to cost effectively scale feature size, the CPU power consumed has become prohibitive.

In contrast to the previous decade, CPU clock rates are scaling slower over time due to the power constraints. However, the number of transistors per silicon area continue to increase roughly at the rate of Moore’s Law. Therefore, CPUs are being designed and built with an increasing number of cores, with each core executing one or more threads of instructions.

This puts a new kind of pressure on the memory subsystem. Though the demand for instructions and data per thread is not increasing very quickly, the rapid growth in the number of available threads puts an increasing emphasis on memory bandwidth. This article summarizes the challenges that arise for the memory subsystem associated with these terascale CPUs.

Memory Key Metrics and Fundamentals

The key metrics for examining the memory sub-systems are bandwidth, capacity, latency, power, system volume, and cost.

Bandwidth (Bytes/second, B/s or bits/second, b/s). Bandwidth is the number of Bytes transferred in a given amount of time. Bandwidth is usually the most talked-about performance metric. The bandwidth required for a system is usually market segment and application (working set size, code arrangement, and structure) dependent. Interestingly, bandwidth alone is not a very useful metric for system design decisions. Other factors must be considered such as cost, power and form factor (size/space) constraints in conjunction with bandwidth.

Capacity (Bytes or B). Capacity is the total number of bytes that can be stored in the region of memory.

Latency (seconds, sec or simply s). This is the time it takes to read a word from the region of memory. The focus is usually on read latency. Write latency is often of less interest; the time required to write to a memory is often not a factor for the performance of the application.

Power (Watts or W). Power equals the energy consumed divided by the time in which that energy is consumed.

System volume, Form Factor. This is the volume required for different technologies into a system. This is usually driven by the physical size of components and/or cooling requirements.

Cost ($). Cost usually refers to the money required to use components in a system.

Often metrics are combined. Frequently used metrics include bandwidth/Cost or Watts/bandwidth (J/bit).

Memory Scaling

Double data rate (DDR) memory has become the dominant memory technology (in terms of number of units sold). DDR-based DRAM products are optimized for high capacity and low cost, not high bandwidth, low latency or low power.

As the CPUs continue to increase in capability toward the terascale level, many of the key metrics are not scaling well and are becoming system design challenges. The metrics being stressed most are bandwidth, power and latency. As potential solutions are investigated, the other metrics of capacity and form factor become challenging as well.

The expression “hitting the memory wall” is often used. Commonly the “memory wall” has the connotation that DDR cannot supply enough bandwidth for CPUs. A more accurate statement is that based on the DDR interface and channel specifications, the bandwidth per pin cannot scale up as quickly as the compute capabilities of CPUs. Simply adding more pins in parallel is not an appealing option due to system cost reasons. The problem becomes acute when CPUs reach the TeraScale performance level. Being more precise, the rate at which bits can be moved between CPUs and DDR devices is limited by the frequency dependent loss, impedance discontinuities, the power available and cost to implement. It will be extremely challenging to push and pull data at rates that exceed 2.4 – 3.2 Gb/s per data signal across DDR channels.

The need to reduce latency and the value of reducing latency is very difficult to assess. Most systems today have put a higher value on bandwidth and choose to use forms of pipelining such as pre-fetching to hide latency. As CPUs approach the terascale range via many threads running in parallel pipeline-based methods to hide memory latency will become less effective. To keep cost and power low, more emphasis will be placed on reducing the latency for the first level of the memory hierarchy that is external to the CPU chip.

Increasing the bandwidth by adding data pins as well as reducing the read latency of DDR devices could be done while maintaining the existing architectures of both the DRAM as well as the interface. However, addressing these bandwidth and latency metrics alone is not enough since one of the greatest challenges to achieving terascale bandwidths is maintaining low power consumption.

DRAM device power is composed of three main components: power consumed by the storage array, power consumed by the datapath and power consumed by the I/O pins. Roughly 50 percent of the power consumed is in the datapath, with the other 50 percent split between I/O circuits and the array. All three areas need to be addressed to create DRAM products suitable for terascale systems.

Evolutionary DRAM Summary

In summary, the key trends for evolutionary memory sub-system scaling are:

•Bandwidth scaling for traditional DDRx-based systems will end at about 2.4 – 3.2 Gb/s per pin (bump).
•To achieve the bit rates above, each channel will likely be limited to one DIMM without extra components, such as buffer on board (motherboard).
•GDDRx gives increased bandwidth but at the cost of capacity. Pin bandwidth will be limited to 5-6 Gb/s for GDDR channels being constructed today.
•Power in the memory sub-system varies from 40-200 mW per Gb/s, translating to hundreds of Watts for a TB/s of bandwidth.
•Adding capacity to evolutionary memory sub-systems is limited to adding channels, buffer on board or other forms of buffered DIMMs.
•Latency improvements for evolutionary systems will be minimal.

Terascale Memory Challenges and Future Memory Technologies

In the following section, we describe some of those challenges facing memory architects and designers and potential solutions.

Memory Technology

The first question we need to ask is which memory technology(s) will fill the needs of these systems. DRAM technology has long dominated the market for off-chip memory bandwidth solutions in computing systems. While non-volatile memory technologies such as NAND Flash and Phase Change Memory are vying for a share of this market, they are at a disadvantage with respect to bandwidth, latency, and power.

A holistic approach is needed to achieve the required results. The main factors that will need to be addressed to achieve the optimal solution for increased bandwidth and lower energy per bit of future terascale memory sub-systems are the channel materials, the I/O density, the memory density, and the memory device architecture. We examine the changes required in these areas.

Channel Materials

First we look at the materials that could be used to construct channels between CPUs and memory modules.

Figure 2: Data Rate versus Trace Length
Figure 2: Data Rate versus Trace Length for different materials

Adding complexity to the I/O circuits in the form of additional equalization, more complex clocking circuits, and possibly data coding can increase the data rate, but also increase the energy per bit moved. More complex interconnects, such as flex cabling, improved board materials, such as Rogers or high-density interconnect (HDI), and eventually, optical solutions, must be considered. The emphasis on higher bandwidth/pin, I/O density and lower energy per bit read/written will lead to selective use of new channel materials.

Memory Density

A DRAM technology that supports a high bandwidth per pin, high capacity and low energy per bit moved will be required. A promising solution to solve these issues is 3-D technology, based on through silicon vias (TSVs). 3-D stacked memory will provide an increase in memory density through stacking, and it will enable a wide datapath from the memory to the external pins, relaxing the per-pin bandwidth requirement in the memory array as shown in Figure 3.

Figure 3: 3-D Stacked Memory Module
Figure 3: 3-D Stacked Memory Module

This design achieves six objectives:

  1. A method for further scaling of DRAM density.
  2. A relatively wide datapath from the memory array to the memory pins, relaxing the speed constraints on the DRAM technology.
  3. A high density connection from the memory module to the memory controller, which makes for more efficient use of power.
  4. The elimination of many of the traditional interconnect components from the electrical path.
  5. It separates the high bandwidth I/O solution from the microprocessor and memory controller power delivery path when using the top of the package for high speed I/O.
  6. The increased density eliminates the need for the electrically-challenged and energy-inefficient, multi-drop DIMM bus.

A key new challenge is introduced; we need a way to move the data from the wide datapath from the memory array to the memory device pins. The general characteristics necessary for an optimal solution are the ability to efficiently multiplex the data at a rate that matches the data rate of the increased device pins (Gb/s), rather than a rate that matches the slower, wider memory datapath, at an efficient energy level (low pJ per bit) that closely matches the characteristics of the CPU generating the memory requests. The architecture, design and implementation of the data collection function will be dependent on the usage of the 3-D memory module, ranging from specialized DRAM chips to a mix of logic process chips and DRAM process chips.

Memory Hierarchy

Given a memory of the type we describe, we must also examine the entire memory hierarchy. For example, it may be advantageous to add a level of memory to the hierarchy.

Analyzing different memory hierarchies is a huge challenge. All the metrics mentioned previously need to be evaluated in the context of the applications of interest (see “Key Metrics”). When considering additional levels of the memory hierarchy, the key decisions are where to add a level or levels in the memory hierarchy and how the levels of memory are managed.

Memory Hierarchy — Where to Add Memory

Earlier, we concluded that to meet the needs of terascale systems, designers should investigate new architectures and manufacturing techniques for DRAM, with an emphasis on 3-D stacking with TSVs. We are confident that these techniques will lead to improved DRAM products, while maintaining a low cost per bit stored. We also realize that when the new technologies are introduced, it will take time for the price per bit to drop. Therefore, early use of 3-D stacked memory as near memory, backed up by DDR-based DRAM or other low cost per bit memory technologies, may be an appealing and cost-effective choice for designers.

The policies of what data (or instructions) are placed, where they are placed as well as what is copied and shared are the key research issues facing system designers. The simple statement that data movement must be minimized will take on additional importance as terascale CPUs are built.

Summary and Conclusions

The demand for bandwidth continues to increase. Terascale CPUs will exacerbate the challenges of the memory subsystem design, including the architecture and design of memory controllers, the memory modules and memory devices themselves. DDR-based memory and interfaces will continue to be used for the markets segments where they can, but the shift to something new will begin in next few years.

To learn more, read the Intel Technology Journal, Volume 13, Issue 4, December 2009, Addressing the Challenges of Tera-scale Computing, ISBN 978-1-934053-23-2

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industry updates delivered to you every week!

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing power it brings to artificial intelligence.  Nvidia's DGX Read more…

Call for Participation in Workshop on Potential NSF CISE Quantum Initiative

March 26, 2024

Editor’s Note: Next month there will be a workshop to discuss what a quantum initiative led by NSF’s Computer, Information Science and Engineering (CISE) directorate could entail. The details are posted below in a Ca Read more…

Waseda U. Researchers Reports New Quantum Algorithm for Speeding Optimization

March 25, 2024

Optimization problems cover a wide range of applications and are often cited as good candidates for quantum computing. However, the execution time for constrained combinatorial optimization applications on quantum device Read more…

NVLink: Faster Interconnects and Switches to Help Relieve Data Bottlenecks

March 25, 2024

Nvidia’s new Blackwell architecture may have stolen the show this week at the GPU Technology Conference in San Jose, California. But an emerging bottleneck at the network layer threatens to make bigger and brawnier pro Read more…

Who is David Blackwell?

March 22, 2024

During GTC24, co-founder and president of NVIDIA Jensen Huang unveiled the Blackwell GPU. This GPU itself is heavily optimized for AI work, boasting 192GB of HBM3E memory as well as the the ability to train 1 trillion pa Read more…

Nvidia Appoints Andy Grant as EMEA Director of Supercomputing, Higher Education, and AI

March 22, 2024

Nvidia recently appointed Andy Grant as Director, Supercomputing, Higher Education, and AI for Europe, the Middle East, and Africa (EMEA). With over 25 years of high-performance computing (HPC) experience, Grant brings a Read more…

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing po Read more…

NVLink: Faster Interconnects and Switches to Help Relieve Data Bottlenecks

March 25, 2024

Nvidia’s new Blackwell architecture may have stolen the show this week at the GPU Technology Conference in San Jose, California. But an emerging bottleneck at Read more…

Who is David Blackwell?

March 22, 2024

During GTC24, co-founder and president of NVIDIA Jensen Huang unveiled the Blackwell GPU. This GPU itself is heavily optimized for AI work, boasting 192GB of HB Read more…

Nvidia Looks to Accelerate GenAI Adoption with NIM

March 19, 2024

Today at the GPU Technology Conference, Nvidia launched a new offering aimed at helping customers quickly deploy their generative AI applications in a secure, s Read more…

The Generative AI Future Is Now, Nvidia’s Huang Says

March 19, 2024

We are in the early days of a transformative shift in how business gets done thanks to the advent of generative AI, according to Nvidia CEO and cofounder Jensen Read more…

Nvidia’s New Blackwell GPU Can Train AI Models with Trillions of Parameters

March 18, 2024

Nvidia's latest and fastest GPU, codenamed Blackwell, is here and will underpin the company's AI plans this year. The chip offers performance improvements from Read more…

Nvidia Showcases Quantum Cloud, Expanding Quantum Portfolio at GTC24

March 18, 2024

Nvidia’s barrage of quantum news at GTC24 this week includes new products, signature collaborations, and a new Nvidia Quantum Cloud for quantum developers. Wh Read more…

Houston We Have a Solution: Addressing the HPC and Tech Talent Gap

March 15, 2024

Generations of Houstonian teachers, counselors, and parents have either worked in the aerospace industry or know people who do - the prospect of entering the fi Read more…

Alibaba Shuts Down its Quantum Computing Effort

November 30, 2023

In case you missed it, China’s e-commerce giant Alibaba has shut down its quantum computing research effort. It’s not entirely clear what drove the change. Read more…

Nvidia H100: Are 550,000 GPUs Enough for This Year?

August 17, 2023

The GPU Squeeze continues to place a premium on Nvidia H100 GPUs. In a recent Financial Times article, Nvidia reports that it expects to ship 550,000 of its lat Read more…

Shutterstock 1285747942

AMD’s Horsepower-packed MI300X GPU Beats Nvidia’s Upcoming H200

December 7, 2023

AMD and Nvidia are locked in an AI performance battle – much like the gaming GPU performance clash the companies have waged for decades. AMD has claimed it Read more…

DoD Takes a Long View of Quantum Computing

December 19, 2023

Given the large sums tied to expensive weapon systems – think $100-million-plus per F-35 fighter – it’s easy to forget the U.S. Department of Defense is a Read more…

Synopsys Eats Ansys: Does HPC Get Indigestion?

February 8, 2024

Recently, it was announced that Synopsys is buying HPC tool developer Ansys. Started in Pittsburgh, Pa., in 1970 as Swanson Analysis Systems, Inc. (SASI) by John Swanson (and eventually renamed), Ansys serves the CAE (Computer Aided Engineering)/multiphysics engineering simulation market. Read more…

Choosing the Right GPU for LLM Inference and Training

December 11, 2023

Accelerating the training and inference processes of deep learning models is crucial for unleashing their true potential and NVIDIA GPUs have emerged as a game- Read more…

Intel’s Server and PC Chip Development Will Blur After 2025

January 15, 2024

Intel's dealing with much more than chip rivals breathing down its neck; it is simultaneously integrating a bevy of new technologies such as chiplets, artificia Read more…

Baidu Exits Quantum, Closely Following Alibaba’s Earlier Move

January 5, 2024

Reuters reported this week that Baidu, China’s giant e-commerce and services provider, is exiting the quantum computing development arena. Reuters reported � Read more…

Leading Solution Providers

Contributors

Comparing NVIDIA A100 and NVIDIA L40S: Which GPU is Ideal for AI and Graphics-Intensive Workloads?

October 30, 2023

With long lead times for the NVIDIA H100 and A100 GPUs, many organizations are looking at the new NVIDIA L40S GPU, which it’s a new GPU optimized for AI and g Read more…

Shutterstock 1179408610

Google Addresses the Mysteries of Its Hypercomputer 

December 28, 2023

When Google launched its Hypercomputer earlier this month (December 2023), the first reaction was, "Say what?" It turns out that the Hypercomputer is Google's t Read more…

AMD MI3000A

How AMD May Get Across the CUDA Moat

October 5, 2023

When discussing GenAI, the term "GPU" almost always enters the conversation and the topic often moves toward performance and access. Interestingly, the word "GPU" is assumed to mean "Nvidia" products. (As an aside, the popular Nvidia hardware used in GenAI are not technically... Read more…

Shutterstock 1606064203

Meta’s Zuckerberg Puts Its AI Future in the Hands of 600,000 GPUs

January 25, 2024

In under two minutes, Meta's CEO, Mark Zuckerberg, laid out the company's AI plans, which included a plan to build an artificial intelligence system with the eq Read more…

Google Introduces ‘Hypercomputer’ to Its AI Infrastructure

December 11, 2023

Google ran out of monikers to describe its new AI system released on December 7. Supercomputer perhaps wasn't an apt description, so it settled on Hypercomputer Read more…

China Is All In on a RISC-V Future

January 8, 2024

The state of RISC-V in China was discussed in a recent report released by the Jamestown Foundation, a Washington, D.C.-based think tank. The report, entitled "E Read more…

Intel Won’t Have a Xeon Max Chip with New Emerald Rapids CPU

December 14, 2023

As expected, Intel officially announced its 5th generation Xeon server chips codenamed Emerald Rapids at an event in New York City, where the focus was really o Read more…

IBM Quantum Summit: Two New QPUs, Upgraded Qiskit, 10-year Roadmap and More

December 4, 2023

IBM kicks off its annual Quantum Summit today and will announce a broad range of advances including its much-anticipated 1121-qubit Condor QPU, a smaller 133-qu Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire