The 4^th Data Prefetching Championship (DPC4)

Co-located with HPCA 2026

​

Organizers: Rahul Bera, Konstantinos Kanellopoulos, and Onur Mutlu

​

February 1, 2026

Sydney, Australia

Why Do We Care about Prefetching?

2

Arm Neoverse v2, HotChips 2023

https://hc2023.hotchips.org/assets/program/conference/day1/CPU1/HC2023.Arm.MagnusBruce.v04.FINAL.pdf

Why Do We Care about Prefetching?

3

Arm Neoverse v2, HotChips 2023

https://hc2023.hotchips.org/assets/program/conference/day1/CPU1/HC2023.Arm.MagnusBruce.v04.FINAL.pdf

Prefetching is important for providing performance

Why Another DPC?

• It’s been a while ...

4

DPC1

HPCA 2009

DPC2

ISCA 2015

DPC4

HPCA 2026

ML Prefetching

ISCA 2021

DPC3

ISCA 2019

Once in ~4 years

~7 years

What Changed since Last DPC?

5

AI/ML inference,

datacenter-class workloads,

complex graph-based apps

Workloads

Ever increasing working dataset size

Data Footprint

32+ core datacenter-class processors with limited

per-core DRAM bandwidth

System Configuration

Goal of DPC4

6

Evaluate innovative ideas on prefetching

under a common evaluation framework

to tangibly advance the state-of-the-art

Format of DPC4

• Simulator: ChampSim

​

• Participants are free to implement prefetchers at L1/L2/LLC

​

• Limited hardware budget
• 32KB for L1 data cache prefetcher
• 128KB for L2 prefetcher
• 256KB for LLC prefetcher

​

• No limitation in complexity

​

• System-level feedback are available to enable dynamic adaptation
• Prefetcher accuracy, memory bandwidth usage, rolling-window IPC, ...

7

New Things in DPC4

8

• State-of-the-art AI/ML inference
• LLama2 (LLM), ViT (image generation), Whisper (text-to-speech), ...
• Real workloads from Google’s warehouse-scale computers
• Graph mining workloads

New Workloads representative of our time

• Every submission was evaluated on three system configurations:
• Single-core with ample DRAM bandwidth (FullBW score)
• Single-core with limited DRAM bandwidth (LimitBW score)
• Four-core random mixes (MC score)
• Dynamic system-level feedback, instead of static knobs (e.g., bandwidth usage)

Broader evaluation across diverse system configurations

DPC4 at a Glance

9

Overall

8

competing teams across the globe

Evaluated across

610

single-core workloads and

484

four-core workload mixes, totaling

3TB

of space

Over

50,000

CPU core-hour

spent on evaluation

9

Program committee members,

providing three reviews

to each submission

Did We Meet the Goal?

10

Sneak Peak: DPC4 Advances the State-of-the-Art

11

2.6% across 610 traces!

2.8% across 484 mixes!

Berti@L1 + Pythia@L2: Already +15.1% over no-prefetching

5.2% across 610 traces!

Sneak Peak: DPC4 Advances the State-of-the-Art

12

27.7%

4%

4.5%

1.1%

Sneak Peak: DPC4 Advances the State-of-the-Art

13

27.7%

4%

4.5%

1.1%

More opportunities for the future

Today’s Agenda (I)

• 2 invited talks from industry veterans

​

​

​

​

​

​

​

​

​

• 8 competitors’ talks

14

Dr. Leeor Peled, Huawei Research

Title: “Is Prefetcher Research

Still Alive?”

Dr. Akanksha Jain, Google

Title: “Data Prefetching:

A Data-Center Perspective”

Today’s Agenda (II)

• Session 1 (13:45 – 15:35): 1 invited talk + 4 papers
• Invited talk by Dr. Leeor Peled
• Crossing the Boundary – Lee+, Yonsei University
• SPPAM – Merrell+, Texas A&M
• Emender -- Chen+, Tsinghua University
• sBerti – Zhou+, HKUST, Guangzhou
• Break
• Session 2 (15:50 – 17:40): 1 invited talk + 4 papers
• Invited talk by Dr. Akanksha Jain
• Pushing the Limits of Berti – Singh+, University of Murcia and University of Zaragoza
• Entangling Data Prefetcher – Torres+, University of Zaragoza , University of Murcia, and IIT Bombay
• uMAMA – Block+, UIUC
• Global Berti – Posluns+, University of Toronto

15

The 4^th Data Prefetching Championship (DPC4)

Co-located with HPCA 2026

​

Organizers: Rahul Bera, Konstantinos Kanellopoulos, and Onur Mutlu

​

February 1, 2026

Sydney, Australia