← Latest papers
🤖 machine learning

DARTH-PUM: A Hybrid Processing-Using-Memory Architecture

The paper proposes DARTH-PUM, a hybrid architecture that integrates analog and digital Processing-using-Memory (PUM) techniques with optimized peripheral circuitry and software support to enable energy-efficient, universal in-memory computing for diverse applications such as encryption and deep learning.

Original authors: Ryan Wong, Ben Feinberg, Saugata Ghose

Published 2026-05-06
📖 5 min read🧠 Deep dive

Original authors: Ryan Wong, Ben Feinberg, Saugata Ghose

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you run a massive library where you must solve complex mathematical problems. Traditionally, you would have to fetch books (data) from the library, carry them to a separate office (the CPU) to do the math, and then return them. This back-and-forth running is slow, exhausting, and wastes a great deal of energy.

Processing-Using-Memory (PUM) is a new idea that says: "Let's do the math right on the library shelves."

The paper introduces DARTH-PUM, a new "hybrid" library system that combines two different methods for performing math within the shelves to get the best of both worlds.

Here is how it works, using simple analogies:

1. The Two Types of Librarians (Analog and Digital)

The paper explains that there are two main methods for performing math in memory, and each has a superpower and a weakness:

  • The "Analog" Librarian (The Sprinter):

    • How he works: He uses electric current flowing through wires (like water flowing through pipes) to perform massive amounts of multiplication and addition simultaneously. It is like a huge choir where everyone sings a note at the same time to instantly create a chord.
    • Superpower: Incredible speed for specific tasks like Matrix-Vector Multiplication (MVM), which represents the heavy lifting in AI and encryption.
    • Weakness: He is somewhat "noisy" and imprecise. He cannot perform complex logic (like "if this, then that") or easily handle negative numbers. He is great for heavy lifting but bad for fine details.
  • The "Digital" Librarian (The Logic Master):

    • How he works: He uses standard zeros and ones (like a normal computer) to perform logic step-by-step. He is like a precise accountant who checks every single number.
    • Superpower: He is perfect for logic, decision-making, and handling negative numbers. He is very accurate.
    • Weakness: He is slow at heavy multiplication tasks because he must perform them step-by-step.

2. The Problem with Previous Attempts

Before DARTH-PUM, researchers tried using only the Analog Librarian for large tasks (like AI). However, since the Analog Librarian cannot perform logic, the system had to send the data back to the main CPU to complete the task. This was like hiring a sprinter for a marathon but forcing them to stop at every mile to run to another city to get a stamp and then run back. This wasted all the speed.

Other attempts tried to build a custom "Special Function Unit" (a specialized robot) to help the Analog Librarian, but these were expensive, difficult to build, and worked only for a specific type of math.

3. The DARTH-PUM Solution: The Hybrid Team

DARTH-PUM (Digital–Analog Resistive Tiles for Hybrid Processing-Using-Memory) is a single chip that houses both the Sprinter (Analog) and the Logic Master (Digital) right next to each other in the same room.

  • The Workflow:

    1. When a large mathematical problem (like multiplying a huge matrix) arrives, the Analog Librarian immediately performs the heavy lifting.
    2. Since the Analog Librarian is somewhat "noisy" or needs to combine results, the data is passed directly to the Digital Librarian, who sits right next door.
    3. The Digital Librarian quickly handles the "cleanup work" (shifting bits, adding numbers, checking logic) without the data ever leaving the library.
  • The "Traffic Cop" (Arbiter):
    The paper describes a special hardware manager that acts like a traffic cop. He ensures the Sprinter and the Logic Master do not get in each other's way. He plans who works when, so the Analog Librarian can finish his big math burst and the Digital Librarian can immediately take over the parts to complete the task.

4. Real-World Examples (What They Tested)

The authors tested this system on three very different types of tasks to prove its flexibility:

  • AES Encryption (Locking a Safe):

    • The Task: Encrypting data to keep it secret. This requires a lot of logic (shifting rows, swapping bytes) and one large multiplication step.
    • The Result: The Analog Librarian took over the multiplication, and the Digital Librarian took over the shifting and swapping. The system was 59.4 times faster than the old method using a CPU plus an analog accelerator.
  • Convolutional Neural Networks (Recognizing Images):

    • The Task: Identifying objects in photos (like ResNet). This requires mainly heavy multiplication but also some logic for "activation" (deciding if a pixel is important).
    • The Result: The system was 14.8 times faster than the baseline. It saved a huge amount of energy because it did not have to move data back and forth.
  • Large Language Models (LLMs – Chatbots):

    • The Task: Understanding and generating human language. This requires complex math and logic.
    • The Result: The system was 40.8 times faster than the baseline.

5. Why This Matters

The paper claims that DARTH-PUM is a "best-of-both-worlds" solution.

  • It does not require expensive, custom-made robots (Special Function Units) for every new application.
  • It does not waste energy moving data between the CPU and memory.
  • It can handle everything from simple math to complex logic on the same chip.

In short: DARTH-PUM is like building a library where the fastest runners and the most precise accountants work side-by-side in the same room. They pass the work back and forth immediately, so the entire team completes the task much faster and with less effort than if they had to run to different buildings to get the work done.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →