SemiconductorsFEATURE

Inside Processing-in-Memory: How Samsung and SK Hynix Are Rewiring the Memory Wall

E
EffectStory 編輯部Editorial Team
Published · Updated
Processing-in-memory (PIM) places compute logic directly inside memory banks to cut the data movement that IBM Research identifies as AI computing's main energy cost. Samsung's HBM-PIM and SK Hynix's GDDR6-AiM are two working implementations of this concept, each reporting measured gains in performance, speed, or power efficiency.

Why has the von Neumann bottleneck become the core driver behind PIM development?

IBM Research identifies data movement between memory and compute units as the primary energy cost in AI workloadsCITE:E1. The organization states that during AI runtime, "the main energy expenditure ... is spent on data transfers — bringing model weights back and forth from memory to compute"CITE:E1. This back-and-forth is the defining cost of the von Neumann architecture, in which memory and processing are physically separated units connected by a shared data path.

What is the core concept of PIM, and how does it break the limits of the von Neumann architecture?

IBM Research defines processing-in-memory (PIM) as a non-von Neumann computing paradigm that performs computation directly inside memoryCITE:E2. Its stated goal is to perform "certain computational tasks in place in memory, thereby obviating the need to shuttle data back and forth between the processing and memory units"CITE:E2. By collapsing the distance between where data is stored and where it is processed, this approach targets the exact data-transfer cost identified above.

What is Samsung's technical approach to HBM-PIM?

Samsung Electronics (三星電子) built HBM-PIM by placing a DRAM-optimized AI engine inside each memory bankCITE:E3. Samsung describes the design as bringing "processing power directly to where the data is stored by placing a DRAM-optimized AI engine inside each memory bank — a storage sub-unit — enabling parallel processing and minimizing data movement"CITE:E3. Samsung positions this as the industry's first High Bandwidth Memory (HBM) integrated with AI processing capabilityCITE:E3.

How much performance and energy improvement does Samsung HBM-PIM deliver?

Samsung states that applying the new architecture to its existing HBM2 Aquabolt solution more than doubles system performance while cutting energy consumption by over 70%CITE:E4. In Samsung's own words, the new architecture "is able to deliver over twice the system performance while reducing energy consumption by more than 70%" when applied to AquaboltCITE:E4.

How does SK Hynix's GDDR6-AiM implement a different PIM approach?

SK Hynix (海力士) built its first PIM product, GDDR6-AiM, to pair with a CPU or GPU in place of standard DRAMCITE:E5. SK Hynix states that "a combination of GDDR6-AiM with CPU or GPU instead of a typical DRAM makes certain computation speed 16 times faster"CITE:E5. Unlike Samsung's bank-level AI engine inside HBM, SK Hynix's design centers on GDDR6 (a separate memory standard from HBM) acting as an accelerator alongside the host processor.

What power-design advantage does GDDR6-AiM offer?

SK Hynix states that GDDR6-AiM operates at 1.25V, below the 1.35V operating voltage of its existing productsCITE:E6. The company describes this directly: "GDDR6-AiM runs on 1.25V, lower than the existing product's operating voltage of 1.35V"CITE:E6.

Comparing the two PIM implementations

VendorProductMetricReported value
Samsung ElectronicsHBM-PIMSystem performance vs. HBM2 AquaboltOver 2xCITE:E4
Samsung ElectronicsHBM-PIMEnergy consumption vs. HBM2 AquaboltReduced over 70%CITE:E4
SK HynixGDDR6-AiMComputation speed vs. typical DRAM pairing16x fasterCITE:E5
SK HynixGDDR6-AiMOperating voltage1.25V, vs. 1.35V priorCITE:E6

What this means: IBM Research's framing of data movement as AI computing's core energy costCITE:E1 and its non-von Neumann in-memory computing conceptCITE:E2 describe the same underlying problem that Samsung and SK Hynix each engineered a distinct product around. Samsung embedded AI engines inside HBM banks and reported the gain primarily in system performance and energy consumption on top of an existing HBM2 product lineCITE:E3CITE:E4. SK Hynix built a separate GDDR6-based accelerator product and reported the gain primarily in computation speed and a lowered operating voltageCITE:E5CITE:E6. Both companies point to the same architectural principle — computing where data sits rather than moving it — but report improvements on different metrics and different base memory standards.

📊 Evidence

FAQ

Why has the von Neumann bottleneck become the core driver behind PIM development?

IBM Research identifies data movement between memory and compute units as the primary energy cost in AI workloadsCITE:E1.

What is the core concept of PIM, and how does it break the limits of the von Neumann architecture?

IBM Research defines processing-in-memory (PIM) as a non-von Neumann computing paradigm that performs computation directly inside memoryCITE:E2.

What is Samsung's technical approach to HBM-PIM?

Samsung Electronics (三星電子) built HBM-PIM by placing a DRAM-optimized AI engine inside each memory bankCITE:E3.

How much performance and energy improvement does Samsung HBM-PIM deliver?

Samsung states that applying the new architecture to its existing HBM2 Aquabolt solution more than doubles system performance while cutting energy consumption b…

📎 Sources

  1. research.ibm.com
  2. research.ibm.com
  3. semiconductor.samsung.com
  4. news.skhynix.com

Related data

Author's TakeEffectStory 編輯部

The two disclosed implementations sit on the same principle but optimize for different targets: Samsung's HBM-PIM reports its gain as a system-level pair — over 2x performance and over 70% lower energy draw — measured against its own HBM2 Aquabolt baseline, while SK Hynix's GDDR6-AiM reports a single computation-speed multiple (16x) plus a modest voltage cut from 1.35V to 1.25V. That split in what each company chose to measure is itself informative: Samsung's framing suggests HBM-PIM is being pitched as a drop-in upgrade path within an existing HBM product line, whereas SK Hynix's framing suggests GDDR6-AiM is being positioned as a discrete accelerator component paired with a host CPU or GPU rather than a replacement for an existing memory SKU. The indicator worth watching next is whether either vendor discloses a successor generation that reports the *same* metric set as its predecessor — that would show whether these are one-off proof points or an iterating product line.

E
EffectStory 編輯部Editorial Team

Related

BRIEF

CNA Launches Taiwan's First News MCP Tool, AskCNA, Priced at NT$200 a Month

Central News Agency (中央社) launched CNA MCP on August 31, 2026, Taiwan's first news tool built on Anthropic's Model Context Protocol (released November 2024), letting AI agents such as Claude, ChatGPT, and Grok retrieve and cite its archives in real time. The tool integrates nearly 5 million newswire stories, 3.5 million photos, and open data from about 150 government agencies, priced at NT$200 a month with an early-bird bonus-quota plan, and received funding from Google Taiwan's nDX Digital Innovation Grant Program.

EffectStory 編輯部 ·
BRIEF

Sony Music and Warner Chappell Sue Anthropic Over Alleged 'Brazen Campaign' of Copyright Theft

Sony Music Publishing and Warner Chappell, joined by other music publishers, sued Anthropic and co-founders Dario Amodei and Benjamin Mann in the U.S. District Court for the Northern District of California, alleging illegal torrenting, scraping, and downloading of copyrighted lyrics and sheet music. The publishers seek up to $150,000 per work and $25,000 per instance of stripped copyright data, a total that could reach several billion dollars. The filing follows Anthropic's earlier $1.5 billion settlement in the Bartz case.

EffectStory 編輯部 ·
BRIEF

Why Anthropic Turned to Nscale and Lambda for $45B and $35B GPU Compute Deals

Anthropic has assembled compute capacity across at least four NVIDIA-linked providers: a $35 billion contract with Lambda tied to a Hut 8-built Texas data center, a $45 billion, six-year deal with Nscale for a West Virginia campus running NVIDIA Vera Rubin systems, a $10 billion contract with startup Volta in Norway, and a reported (unconfirmed) tenancy at Riot Platforms' Rockdale, Texas site. NVIDIA sits inside nearly every arrangement — as investor, lessor, or chip supplier.

EffectStory 編輯部 ·