Skip to content

Memory Wall

Compute outran memory. This is the index.

Wm. A. Wulf and Sally A. McKee, “Hitting the memory wall: implications of the obvious.” ACM SIGARCH Computer Architecture News 23(1), March 1995, pp. 20–24. doi:10.1145/216585.216588

Five pages of arithmetic: if processor speed and DRAM latency improve at different exponential rates, average memory access time comes to dominate runtime regardless of cache hit rate. The wall follows from the growth rates alone. It is not a design mistake anyone can correct, and thirty years later it is still the organizing constraint of the field.

The interplay is the point

Three lanes, one axis. HBM appears in the interface lane bridging DRAM and compute, exactly when compute is straining against GDDR. NVMe appears when SSDs have outrun a protocol built for one slow mechanical actuator. CXL bridges all three at once.

Select a milestone to see its summary and its sources.

NPU / computeDRAMStorageInterfaces16 milestones, 1968–2024

Loading timeline…

Select a milestone to see its summary and sources.

What this is

An index, not an encyclopedia. There are already many good introductions to DRAM, and this is not another one. The valuable artifacts are the ones nobody has assembled: sourced generation timelines with real spec deltas, an annotated bibliography of the papers that mattered, a normalized accelerator comparison that says out loud how each vendor constructs its TOPS/W claim, the readable history of the industry, and honest maps of where public knowledge stops.

Most of the underlying material is genuinely public — NVM Express publishes openly, Micron publishes real datasheets, much of the JEDEC catalogue is free after registration, and fifteen years of academic DRAM characterisation sits on arXiv. It has just never been assembled in one place.

Where that stops being true, we say so. The mainline DDR5 and HBM standards have moved to non-member purchase since this project started, while HBM3 and HBM4 remain free — a boundary that is recorded and sourced rather than smoothed over.

Verification over generation

Generated prose is nearly free and therefore nearly worthless, and this domain is maximally hostile to unverified numbers. Every figure here carries a source ID in structured data. Every source ID must resolve to a registry record with an access date and a real archive snapshot. A pull request that adds an unsourced number fails CI.

Where a fact has been extracted but not verified by a human, it sits in quarantine/ and does not appear on this site.

Sources

paperPaywalled

Hitting the memory wall: implications of the obvious

Wm. A. Wulf, Sally A. McKee · ACM SIGARCH Computer Architecture News 23(1), pp. 20-24, 1995

The project's founding document and the origin of the phrase. Five pages of arithmetic showing that if processor speed and DRAM latency improve at different exponential rates, average memory access time comes to dominate total runtime regardless of hit rate — so the wall is a consequence of the growth rates alone, not of any fixable design choice. Cite it for the framing and the argument, not for device numbers, which are long obsolete. Publisher of record dates it March 1995; it is frequently cited as 1994 from its drafting and circulation.

standardPurchase

JESD79-5 - DDR5 SDRAM

JEDEC Solid State Technology Association · JESD79-5, 2020

The DDR5 base standard. Authoritative for bank grouping, on-die ECC, the move of power management onto the module, and the timing parameter definitions that every DDR5 datasheet then instantiates. Later revisions change numbers that matter; do not cite this record for a figure taken from a revision without recording which revision. This record no longer describes a free document. The JESD79-5 line has moved to non-member purchase and the page quotes $423.00, which corrects the position this project stated at bootstrap. It is also a series URL: it serves JESD79-5D, Version 1.41, published November 2025, not the 2020 base edition this record is dated to. Read the document number and release version off the page before attributing any figure to an edition. Where a reader who will not pay should go instead: the Micron component datasheets in this registry carry full timing tables, mode register maps and refresh behaviour for shipping DDR5 parts, free and without registration. They are not the standard and cannot settle what the standard requires, but for concrete timing figures they are the better source anyway.

standardPurchase

JESD235D - High Bandwidth Memory (HBM) DRAM

JEDEC Solid State Technology Association · JESD235D, 2021

The current revision of the original HBM line, and the document to cite for anything in the JESD235 series read today rather than from the 2013 edition. The landing page states the document number and a publication date of March 2021. Where it stops, for this project's purposes: the page carries an abstract and the committee assignment, not the parameter tables. Bus width, channel and pseudo-channel organisation, per-pin data rate, per-stack bandwidth, stack height and supply voltage all live inside the PDF, which is $247.00 for non-members. No figure in the HBM generation table may be filled from this record until someone has actually opened the document. Note the slug. JEDEC serves the current revision of a series at a stable series URL, so docs/jesd235a serves JESD235D and will serve JESD235E when there is one. A JEDEC URL identifies a series, not an edition, and an archive snapshot of one captures whichever revision was current on that date. Read the document number off the page rather than off the URL.

standardFree, registration required

JESD238B.01 - High Bandwidth Memory (HBM3) DRAM

JEDEC Solid State Technology Association · JESD238B.01, 2025

HBM3, and the current revision of it. The landing page states the document number and a publication date of April 2025, and offers a free download after registration — no purchase block, unlike the JESD235 line. That contrast is the useful thing this record carries. HBM3 broke out of the JESD235 series into its own document number, and it remained free while the older series moved to paid. Anyone who concluded from a JESD235 paywall that the HBM standards are closed would be wrong about the generation they most likely want. Authoritative for the HBM3 interface as standardised: the page's own abstract states the 64-bit-per-channel data bus at double data rate, against the 128-bit channel of the original HBM. Every other parameter is inside the PDF and none has been read yet, so this record supports no figure in the generation table beyond the two the page itself prints. The slug caveat applies: docs/jesd238 is a series URL and serves whatever revision is current. It served the original JESD238 when the 2022 snapshot was taken and serves B.01 today.