technologybriefs
8:32in productionCh. 1 · Origin/ 8:32 · ceiling 15 min
Semiconductors

High Bandwidth Memory

HBM doesn’t fix memory latency—it weaponises stack height to brute-force bandwidth.

HBM is a JEDEC-standardised 3D-stacked DRAM interface built on TSVs. It delivers bandwidth density unattainable with planar memory. It requires advanced packaging, runs hot, and does not lower latency. Its value is narrow but decisive: for AI and HPC chips where bandwidth-per-mm² is the limiting constraint.

SemiconductorsSemiconductorsHardwareArtificial intelligenceSemiconductors across the network →
Chapters & takeaways4
  1. 1:02
    Origin

    HBM was co-proposed by AMD and SK Hynix in 2010 and standardised as JESD235 in October 2013.

  2. 2:26
    Evolution

    HBM2 arrived in January 2016; an update in late 2018 raised its bandwidth ceiling.

  3. 3:54
    Generations

    HBM3 launched on 27 January 2022; HBM4’s preliminary spec appeared in July 2024, final in April 2025.

  4. 5:04
    Standardisation

    Every generation is a JEDEC ratification—not a vendor announcement—making it a rare consensus interface in high-performance memory.

Worth your time?

Yes. Study the whole thing.

4/ 5
What works
  • delivers scalable bandwidth via vertical stacking
  • enables JEDEC-governed interoperability across vendors
  • replaces wide I/O buses with compact, high-pin-count interfaces
What does not
  • reduce memory latency
  • lower cost per gigabyte
  • simplify thermal design
Study it if
  • chip architects
  • AI accelerator designers
  • HPC system integrators
Skip it if
  • embedded systems engineers
  • mobile SoC designers
  • cost-sensitive consumer GPU makers
The written brief1 min read

What it is and the problem it solves

HBM solves the memory bandwidth wall in accelerators by replacing wide parallel buses with dense 3D-stacked interfaces.

How it works

HBM stacks DRAM dies vertically using through-silicon vias (TSVs) to create a wide, short interconnect path between memory and processor.

What works

The JEDEC-standardised TSV-based stacking delivers predictable bandwidth scaling across generations—HBM2 (2016), HBM3 (2022), HBM4 (2025)—with increasing channel counts and transfer rates.

What does not

HBM does not reduce memory latency meaningfully compared to conventional GDDR or DDR; it trades latency for raw bandwidth and area efficiency.

What it changes

It shifts memory bottlenecks from interface width to thermal management and package integration complexity.

Is it worth your time

Yes—if you design or deploy AI accelerators, GPUs, or high-performance compute systems where bandwidth density matters more than cost or power per bit.

Same field · Semiconductors4 of 40
10:49
Acorn Archimedes1987The Acorn Archimedes is a 1987 British personal computer family built around Acorn’s ARM2 RISC processor. It delivers real architectural innovation — the first mass-market use of ARM — but remains tightly constrained by proprietary chipset design, limited software reach, and unverified performance claims. It proves RISC can work in desktop hardware, but fails to escape its national, institutional, and technical silo.
10:08
ARM architecture familyARM is a licensable RISC instruction set architecture developed by Arm Holdings. It enables low-power, low-cost, low-heat computing across portable and embedded devices — and increasingly desktops and servers. Its success rests on separating specification from implementation, enabling mass adoption without vertical control. It does not produce chips. It does not enforce microarchitecture. It does not guarantee cross-vendor binary compatibility. Its dominance — 230 billion chips since at least 2003 — reflects licensing efficacy, not technical inevitability.
9:13
Arm Holdings1990Arm Holdings is a British semiconductor design company headquartered in Cambridge, England. Its primary business is designing CPU cores implementing the ARM architecture family of instruction sets. It also designs other chips, provides software development tools (DS-5, RealView, Keil), and offers systems, platforms, SoC infrastructure and software. As a holding company, it holds shares of other companies. Since 2016, it has been majority owned by Japanese conglomerate SoftBank Group. 'ARM' originated as an acronym for Acorn RISC Machine and later Advanced RISC Machines. ARM CPUs first appeared in the Acorn Archimedes. Processors based on Arm-licensed designs or instruction set architecture implementations are used in all classes of computing devices. Arm has two GPU lines — Mali and Immortalis — the latter featuring hardware-based ray-tracing.
9:34
Broadcom1961Broadcom is a semiconductor and infrastructure software supplier—not a pure-play chipmaker or SaaS vendor.
Up next in Technology

High-speed rail

1964 · 10:49

High-speed rail is not fast trains on old tracks—it’s expensive infrastructure pretending to be transport policy.

Play nextBack to Semiconductors
10:49