Cut memory in half, performance still 95%

🔧 AI Tools 2026-11-14 10 min read

Cut memory in half, keep 95% of the performance: AMD does it with software alone. Install half the memory, and you still keep 95% of the performance. When AMD dropped this line at the In Depth conference on the 11th, my first reaction was: here we go again?

The key point is that it didn't build new chips, and it didn't buy a memory manufacturer. In June it acquired a small software company called MEXT, and only on the 11th did it put the details on the table. What it saves is the most expensive material, yet it doesn't rely on new hard

📜 Table of Contents

Cutting Memory in Half, Keeping 95% of Performance: AMD Does It With Just Software

Install half the memory, keep 95% of the performance. When AMD dropped that line at its In Depth event on the 11th, my first reaction was: here we go again.

The key point: it didn't build a new chip, and it didn't buy a memory fab. In June it acquired a small software company called MEXT, and only on the 11th did it put the details on the table. It's saving on the most expensive component, but not through new hardware.

This stands out precisely because memory has gotten so expensive. In August this year, Korean-exported DRAM was selling at $92,000 per kilogram, up 401% year over year — roughly the price of 620 grams of gold. A single 16GB DDR4 stick was $3.20 at the end of 2024 and climbed past $62 by the end of 2025. At these prices, adding a stick makes you take a deep breath. TrendForce estimates Server DRAM contract prices will rise about 270% in 2026, and by 2027 DRAM plus NAND will eat up 68% of cloud providers' capex. Samsung, SK Hynix, and Micron are funneling their good capacity into HBM and server memory, leaving consumer-grade on the sidelines. Even the PS6 won't commit to a release date — probably not unrelated.

What this software actually does

Server memory has a counterintuitive quirk: it looks full, but a big chunk is just lying flat. A Meta study found that 50% to 80% of memory pages go untouched for five minutes, ten minutes, or even longer — even when the system shows 95% usage.

What MEXT does is move cold pages onto NVMe flash, keeping only hot data in the expensive tier.

Sounds simple enough. The catch is speed. Flash costs about 50x less per bit than DRAM and uses 30x less power, but it's 500x slower. If an application suddenly demands a page that was just moved out, it just has to sit and wait.

So AMD layered an AI predictor on top of a Linux kernel module: a Transformer-derived model that doesn't guess the next word — it guesses the next memory page, pulling that page back into DRAM before the application asks for it. Gary Smerdon, MEXT's founder and now an AMD VP, says that as long as the page is back before it's needed, it doesn't matter which page gets moved first or last.

What surprised me most is that it's not picky at all: no GPU needed, it occupies one EPYC core, first training takes under a second, and after that it learns as it runs. The system side is even more fun — memory that physically doesn't exist still shows up as "installed."

That 95% number — don't rush to share it

AMD put out a few of its own tests:

Nice, right? But put these three workloads side by side and the common thread is obvious: large data volumes, untouched most of the time, access patterns that repeat over and over. PrimeTime especially — massive datasets that are basically read-only and traversed again and again; you could say it's practically a custom exam question for the predictor. The Tom's Hardware reporter himself wrote that it's "a little too good to be true."

To its credit, AMD didn't oversell it, admitting that most ordinary workloads only save 20% to 25% of memory, with 20%–75% as the range — the closer you get to the upper bound, the pickier the workload and the more performance drops off. And it's all vendor self-testing; no third party has reproduced it yet.

The money math, though, is real. AMD first deployed it on its own 145 servers, estimates it saved about $4.9 million in memory purchases, with expansion costs 83% lower than buying all DRAM. It plans to roll out to 2,000 servers by 2027 and has partnered with HPE to sell it. Right now, phase one slaps a "MEXT Ready" label on sixth-gen EPYC Venice and reserves a dedicated NVMe drive bay; phase two is when licensing and the RAM View tool ship. In other words, this is a paid add-on, not something that comes free with the CPU.

Some cold water

First, it's betting on guessing right. Guess wrong and you lose on both ends: latency stalls, and a memory seat is wasted for nothing. For workloads that stay quiet and then suddenly want everything — like 3D rendering dragging its cache — the predictor is basically helpless.

Second, the ecosystem math isn't settled. The driver is closed-source, and it sits right at the bottom layer of memory mapping — Phoronix has already called this out; in virtualization, each VM needs another core. Out in the wider world, CXL memory pooling, MemVerge, and Linux's built-in zswap and tiered memory are all crowding onto the same road. People in the industry haven't forgotten how Intel Optane ended.

Third, and most important, this has nothing to do with your PC. It only serves EPYC data centers. In the short term it won't make memory sticks cheaper — if anything, it shows capacity is being siphoned off by data centers bite by bite. For those of us who buy two sticks at a time, any benefit depends on whether new fabs ramp up around 2028.

Still, I think it's worth watching

Because the approach has changed.

Before, when memory ran short, there was only one fix: add a stick. Now someone is working on letting AI manage where data should live. Cheap flash anyone can buy; the hard part is knowing when to pull which page back ahead of time. With hardware squeezed from both ends — physical limits and capacity — software has squeezed out a gap. SK Hynix says this round of tightness could drag on to 2030, and this kind of trick — trading scheduling for capacity — will likely spawn several more players.

As for when it reaches ordinary gamers like us, I'm not holding my breath for now. Let's just watch.


Appendix 1: Image List (not counted in the body)

Reordered for the news angle: principle → performance data → background → official evidence. All 4 use bright, high-saturation colors with dark text, 1280×720 @2x (output 2560×1440 PNG). Source files 配图源文件.html + re-render script render.py; edit the text and run once.

Publish order File Content Purpose
1 figure-1.png MEXT principle: DRAM hot pages ↔ NVMe cold pages + AI prediction arrows, labeled "500x slower" with four tech capsules Main image, explains how the tech works
2 figure-2.png Bar comparison of three workloads at 50% / 25% DRAM, corner tag "AMD internal testing · no third-party reproduction yet" Data backing the 95%
3 figure-3.png 1 kg DRAM ≈ 620 g gold, plus 16GB DDR4 $3.2→$62 Background board, pairs with the price-hike paragraph
4 figure-4.png AMD official RAM View screenshot (courtesy of Peng Shu) embedded in a bright card, with 4 Chinese annotations on the right: 1.5TB / ~768GB / idle-time color coding / optimizable range Official evidence, explains where "cold pages" come from

Assets: 官方图-RAM-View.webp (copied and archived from desktop, original 1200×675). The official image is dark, so it's wrapped in a bright card + bright annotations; the overall look stays bright-dominant, keeping to the color rules.

Covers (two versions, tech version is primary): - 封面-技术版.png: 1536×1024, bright high-saturation, main visual "Cut memory in half, keep 95% of performance" + full DRAM ↔ 50% DRAM + flash comparison + four tech capsules. Drawn in HTML/CSS (封面技术版.html / render_cover.py), zero typos in Chinese. Correct news angle, main cover. - 封面.png: vivid + high generated-image version, main copy "Memory costs more than gold," green memory sticks = gold bars. Tool watermark removed with make_cover.py, original archived as 封面-原始生图.png. More of a background hook, backup / secondary image.

⚠️ Publishing focus (corrected by Peng Shu on 11-12, revisions completed accordingly): the news angle of this piece is AMD using software to halve memory while keeping 95% of performance; "memory costs more than gold" is only era background and must not take center stage. Information hierarchy: primary = counterintuitive technical fact; secondary = storage price-hike background; footnote = 145 servers / $4.9M / 2,000 servers by 2027.

WeChat formatting (red-white template): 公众号版(红白).html is a pure WeChat-native fragment of <section> + <p> + <span leaf="">, same template as the "published" draft; 公众号版.html is a preview page with a "Copy to WeChat" button. Both are generated from this file's body by build_wechat.py; edit the text and rerun to sync. Usage: open 公众号版.html → click "Copy to WeChat" at the top → Ctrl+V in the WeChat backend. Verified that the clipboard contains both text/html (~480k characters, preserving sections, all inline styles, and embedded images) and text/plain; the toolbar won't be copied over. ✅ All 4 images embedded in the layout: at generation they're compressed to 1080px-wide JPEGs and written into <img> as base64; visible in preview, and the one-click-copy rich text carries the images too (all 4 confirmed in testing). Insertion order: figure 03 background (opening) → figure 01 principle + figure 04 official evidence (principle section) → figure 02 performance data (performance section), each with a light-gray caption below. If the WeChat editor doesn't auto-transfer base64 images, manually replace with the same-named PNGs; upload 封面-技术版.png separately as the cover. 公众号版预览.png is a full-length image of the overall effect.

Appendix 2: Source List (not counted in the body)

  1. Tom's Hardware, 2026-11-11, Jake Roach, "AMD shares how it can retain 95% of server performance with just half the DRAM" (paywalled; first half used)
  2. Tom's Hardware, 2026-06-15, "AMD takes over MEXT…"; AMD official announcement 2026-06-15 "AMD Acquires MEXT to Advance Memory Optimization"
  3. SDxCentral, on-site report from AMD In Depth 2026 (Gary Smerdon interview, 50x/30x/500x, single EPYC core, <1s training)
  4. The Register (Meta cold-page study, the five-minute argument)
  5. 工商時報 2026-11-11, 中時新聞網 2026-11-11 (145 servers / $4.9M / 83% / 2,000 servers / HPE)
  6. IT之家 2026-11-12 (Spark TPC-DS 99%, 20%–75% range)
  7. 騰訊新聞·IT時代網 2026-11-12 (transformer model, Linux kernel module, Venice MEXT Ready)
  8. 新浪科技 2026-11-12 (25% DRAM scenario boundaries, 82%–83% discrepancy)
  9. TrendForce 2026-08 (Server DRAM up ~270% in 2026, storage at 68% of CSP capex in 2027)
  10. 中國經濟網 / 中國發展網 2026-09 "'Storage inflation' is spreading" (Korean export unit price, DDR4 prices, capacity migration)
  11. NAND Research "MEXT: Predictive Memory Software to Control DRAM Costs" (CXL / MemVerge / zswap competitive landscape)
  12. Phoronix (questions over closed-source Linux driver)

Note on figures: sources differ on 145 vs "100+", 99% vs 95%, 83% vs 82%, $4.9M vs ~$5M, etc. This article consistently uses the 工商時報 / IT之家 figures and labels them in the body as "AMD claims / vendor self-testing."

Appendix 3: Revision Log (not counted in the body)

I. AI-flavor audit and colloquial rewrite

# Problem in the draft Fix
1 "Not… but…" contrast sentences, 5 across the piece Cut down to the one necessary technical contrast: "doesn't guess the next word — it guesses the next memory page"
2 Template opening "Let me put it out there first: …" Changed to get straight to the point; after the focus correction, further changed so the first sentence drops "Install half the memory, keep 95% of the performance"
3 Numbered subheads "1, 2, 3, 4, 5" Changed to colloquial short headings
4 Clusters of essay-style metaphors Replaced with "lying flat," "sit and wait," "not picky at all," "haven't forgotten"
5 Rhetorical self-Q&A "See the common thread?" Changed to "Nice, right? But put these three workloads side by side"
6 Universal punchline ending Removed the parallel-structure flourish, landed on "I'm not holding my breath for now. Let's just watch."
7 4 em dashes, 8 bold passages Em dashes down to zero, bold cut to 4
8 Zero first person, zero hesitation Added "I," "probably," "here we go again?", "don't rush to"

II. Focus correction (Peng Shu's 11-12 correction, all three changed)

Location Before After
Title Memory costs more than gold, AMD wants to cut memory in half with software Cut memory in half, keep 95% of performance: AMD does it with just software (folder renamed to match)
Opening Three paragraphs of price hikes before getting to the point First sentence "half / 95%" + second paragraph "no new chip, just software"; price hikes compressed to one paragraph, explicitly demoted to background for "why this stands out right now"
Image order figure-1 gold comparison in the center Physically reordered: 1 principle / 2 performance data / 3 gold background / 4 official evidence; cover switched to the tech main-visual version

Word count control: draft 1,193 Chinese characters → colloquial 1,305 → compressed 1,230 → 1,197 → after focus correction, rechecked and still in the 1,000–1,200 range; only redundant filler words removed, no information lost. Rigor preserved: all performance numbers still labeled "vendor self-testing, no third-party reproduction"; the 20%–75% range, 145 servers / $4.9M / 83% / 2,000 servers, and the closed-source driver and competitor concerns are all retained.

Related Articles
2026-08-16
Best AI Transcription Tool in Milan 2026: 6 Tools for Fashion, Media and Agencies
2026-08-15
Best AI Language Tutor in Sydney 2026: 5 Tools for Students and Migrants
2026-07-26
Free AI Music From Lyrics 2026
2026-08-02
Deepgram API Guide 2026: Nova-3 Speech-to-Text, Pricing, and When to Use It Over Whisper
2026-08-25
Best AI Review Reply Generator in New Orleans 2026: 6 Tools for Hospitality
2026-08-23
Best AI Social Media Scheduler in Almaty 2026: 6 Tools for Kazakh E-commerce and Influencers

Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.

💬 Comments (0)

No comments yet. Be the first!

Login to comment