In this guide: 21 sections
- The whole answer, by workload
- What to do with this on a listing page
- Sixteen gigabytes is now a published floor, and 8GB has slipped below it
- “Percent in use” is the wrong gauge, and the right ones are counters
- Game requirements have settled at 16GB, and the tests side with the minimum
- Where 16GB runs out: a heavy title, a small graphics card or a second program
- Vendor, benchmark and workstation builder disagree on creative work by a factor of three
- Developers pay for each virtual machine twice: once in the guest and once in the default
- Local AI is sized by arithmetic: parameters times bits, plus a cache that grows with context
- Capacity decides whether a model runs, and bandwidth decides how fast
- Two modules, not four: the population rules that set a desktop’s capacity steps
- 32GB was September 2026’s cheapest DDR5 gigabyte, and the step up to it cheaper still
- A missing gigabyte costs more than a missing megatransfer
- A soldered laptop is sized once, at checkout, for its whole life
- Four home buyers, worked from the workload to the kit
- A company fleet needs memory tiers by role, and Microsoft has already written them down
- Virtual desktops are sized per session, then rounded up to fill the memory channels
- Servers are sized per core and per channel, and the cloud has published the ratios
- A database wants its working set in memory, and not a gigabyte more than its edition can use
- AI inference servers carry more host memory than GPU memory, and the cache is what grows
- When a server gigabyte costs $30 to $75, measurement is the cheapest capacity there is
Buy 16GB for office work and esports, 32GB for a new gaming or working desktop and 64GB for serious creative work; size local AI and servers by arithmetic, not by tier. Sixteen gigabytes still runs office work, browsing and nearly every current game, but in 2026 it is the floor rather than the recommendation: Microsoft will not call a laptop a Copilot+ PC with less than 16GB, even as 8GB machines came back into the stores this year with Apple’s MacBook Neo and Microsoft’s 8GB Surface models. Four things decide the real number: a graphics card with 8GB or less of video memory, which pushes game data into system RAM; the rule that a desktop reaches full speed only with two modules; a “memory in use” figure that measures what was available rather than what was needed; and the price of each step, where 32GB was the cheapest DDR5 gigabyte on Tom’s Hardware’s US price index of September 14, 2026.
The whole answer, by workload
Prices throughout are as of late September 2026.
| Home workload | Floor | Target | Comfortable | What moves you up | Two-module configuration |
|---|---|---|---|---|---|
| Office, browsing, calls | 8GB on ChromeOS or a MacBook Neo; 16GB on Windows | 16GB | 32GB | Copilot+ features, many tabs, calls alongside, swap on an 8GB Mac growing toward its installed memory (8GB machines) | 2 x 8GB or 2 x 16GB |
| Esports | 8GB | 16GB | 16GB | streaming (where 16GB runs out) | 2 x 8GB |
| Current AAA games | 16GB | 32GB | 32GB | a card with 8GB or less, streaming, Mafia-class titles (games) | 2 x 16GB |
| Photo and light video | 16GB | 32GB | 64GB | Lightroom’s AI tools, After Effects (creative work) | 2 x 16GB or 2 x 32GB |
| Professional video and compositing | 32GB | 64GB | 96 to 128GB | 4K and longer timelines, Fusion (creative work) | 2 x 32GB, 2 x 48GB or 2 x 64GB |
| Development | 32GB | 32GB | 64GB | WSL, Docker, emulators (development) | 2 x 16GB or 2 x 32GB |
| Local AI up to 8B | 16GB | 32GB | 32GB | long context (model sizes) | 2 x 16GB |
| Local AI beside a graphics card | 16GB | twice the card’s memory | 64GB | a model larger than the card (offloading) | 2 x 16GB, 2 x 24GB or 2 x 32GB |
| Local AI, 70B at 4 bits | 64GB | 96GB | 128GB unified | context past about 32,000 tokens, speed (sizes, speed) | 2 x 48GB, or a unified-memory machine |
| Home server or NAS | 8GB (TrueNAS, up to eight drives) | 16 to 32GB | 64GB with virtual machines | guests, deduplication at 5GB per TB, and Proxmox’s own 1GB per TB of used ZFS storage (by platform) | every channel filled, often registered DDR4 |
| Enterprise role | Sizing rule | Published source | Worked here |
|---|---|---|---|
| Task and knowledge workers | 16GB per seat | this guide’s floor, from the Copilot+ requirement and the 8GB Surface measurement; Microsoft’s own tiers start at 4 to 8GB | 1,000 seats |
| Developers and creators | 32GB; 64GB for data work | Microsoft AVD and Windows 365 | development; fleet tiers |
| CAD workstation | 5GB plus 20 times the largest assembly | Puget Systems | 32GB for a 1.1GB assembly |
| Data analysis in pandas | 5 to 10 times the dataset, an upper bound | Wes McKinney, 2017 | 40 to 80GB for an 8GB dataset |
| Multi-session VDI | 16GB per 8-vCPU host: 48 light users at 0.33GB each to 16 heavy at 1GB | Microsoft | two hosts for 50 users |
| Single-session VDI | 4 to 16GB per desktop (more for CAD) plus 15 to 30% | Servermall; Microsoft | 768GB for 50 users; 1,152GB across three hosts with one spare |
| General virtualization | 4GiB per vCPU, rounded to full channels | AWS; Azure | 384 to 1,536GB per socket |
| SQL Server | 75% of available memory; Standard capped at a 256GB buffer pool | Microsoft Learn | 384GB host |
| PostgreSQL | shared_buffers at 25%; little gain past 40% |
PostgreSQL documentation | 64GB on 256GB |
| AI inference host | at least 2 times GPU memory (Puget, citing Nvidia); Nvidia’s own systems ship at 1.3 to 2.8 times; cache = sequences x context | Puget Systems; Nvidia product pages | 80GiB of cache for 64 sequences of a 70B model |
Three rules hold across both halves. Measure before buying, because “in use” overstates need at home and “allocated” overstates it in a data center (the gauges, servers). Reach a capacity at full speed: two modules on a desktop, every channel filled on a server (population rules). And at 2026 prices, spend on capacity before speed (capacity against speed).
What to do with this on a listing page
- Measure first (the gauges): peak commit and hard faults over a working week on Windows, the memory pressure graph on a Mac, MemAvailable and the “full” pressure line on Linux, active memory on a hypervisor.
- Choose the total from the tables above, then the module size that reaches it: two modules on a desktop, every channel on a server.
- Open the total page for that capacity, such as 32GB DDR5 kits, and judge a row against its generation’s typical price per gigabyte on Trends, not the cheapest row.
- Read the module split in the title. 2 x 16GB is not 4 x 8GB, and 96GB of desktop DDR5 should be 2 x 48GB; on DDR5 a 96GB title can also be a single server module; “96GB DDR4” is always a kit total, or a figure to check, for the reason DDR4 vs DDR5 gives.
- Start from the default view, which compares Buy It Now prices, because a
bid is not a price;
?bids=showadds bid-only rows and?auctions=noremoves auctions entirely. - Check the condition on each row, and match a pair by part number on the part pages rather than by title.
- For a laptop, confirm there are slots at all; for a server, confirm module type and population against the table of server models.
- If buying used, follow buying used RAM on eBay for testing and returns.
Sixteen gigabytes is now a published floor, and 8GB has slipped below it
Microsoft’s Windows 11 specifications page gives 4GB as the minimum, the amount needed to install and boot. Application vendors publish their floors the same way:
Windows 11 minimum 4 GB Microsoft, Windows 11 specifications page
Teams desktop client 4 GB Microsoft Learn, updated July 16, 2026
Microsoft 365 Apps 4 GB Microsoft Support, requirements page
-----
stated minimums, stacked 12 GB before a single browser tab is open
Each of those figures assumes it is the only thing on the machine, and an ordinary working day is all three at once plus a browser. The one 2026 measurement of an 8GB Windows machine bears that out. The Verge tested one of Microsoft’s new 8GB Surface models, and TechSpot reported on July 20, 2026 that with about ten Chrome tabs, Slack and Signal open, Task Manager showed around 6.7GB of the 7.6GB available in use “most of the time”, 88%, and the laptop froze for several seconds whenever someone joined a Teams call. The 8GB Mac fails more gently: in Macworld’s week with a MacBook Neo in March 2026, swap grew “to nearly 8GB” at 59 tabs and a long export ran “about 10 minutes longer” than on an M5 Max MacBook Pro; the reviewer saw no stalls and called his results “all anecdotal”. An 8GB Mac keeps working by writing its overflow to the SSD, and its memory can never be added to.
The Copilot+ requirement is the only memory floor above 8GB found from a PC platform owner; devices with less “are not Copilot+ PCs” even when the processor qualifies. The others draw their lines lower:
| Platform or product | Stated memory | Who says so, and when |
|---|---|---|
| Windows 11 installation | 4GB minimum | Microsoft, Windows 11 specifications page |
| Chromebook Plus | 8GB minimum | Google, October 2, 2023 |
| MacBook Neo | 8GB, the only option | Apple, March 4, 2026 |
| MacBook Air (M5) | 16GB base; 24GB or 32GB to order | Apple specifications page, September 2026 |
| Mac mini (M6) | 16GB base; up to 32GB | Apple Newsroom, September 22, 2026 |
| Copilot+ PC | 16GB of DDR5 or LPDDR5 | Microsoft, Windows 11 specifications page |
A 2026 base configuration, such as the 8GB MacBook Neo, launched at $599 on March 4, 2026, and raised to $699 on June 25 (Tom’s Hardware, citing Bloomberg), is where the maker’s costs stopped, not a recommendation about what you need.
“Percent in use” is the wrong gauge, and the right ones are counters
When programs ask for more than the memory holds, the operating system moves the least-used pages to the drive (the page file on Windows, swap on a Mac or Linux) and reads them back far more slowly. The usual gauge cannot tell that moment from memory that is merely busy. HP’s 2026 guide says “If usage consistently sits above 80%, your system needs more memory”; Newegg’s says 50 to 60% means you have the right amount. Both read a number the system inflates on purpose: in PC Gamer’s April 2026 test on one RTX 5070 system, Star Citizen used 29.3GB with 32GB installed and 13.7GB with 16GB. Memory in use on a large machine measures what was available, not what was needed. The signals that mean a shortage record the machine being forced to give something up.
On Windows, read commit charge against commit limit, then hard faults. The commit limit is “the sum of physical memory (RAM) and all page files combined” (Microsoft Learn, February 12, 2026), and a system-managed page file grows when the charge reaches 90% of it; Task Manager’s Performance tab shows both under “Committed”. Sustained hard faults under your normal workload signal that more memory would help:
Windows: Performance Monitor counters
\Memory\Committed Bytes commit charge: memory promised to processes
\Memory\Commit Limit physical RAM plus every page file
\Memory\Available MBytes memory free for new work
\Memory\Page Reads/sec hard faults: reads from disk, file-cache misses included
On a Mac, the gauge is the memory pressure graph, not “Memory Used”: yellow
means the Mac “might eventually need more RAM” and red that it “needs more RAM”
(Apple), and a nearly full Mac with a green graph is working as designed. On
Linux, read MemAvailable in /proc/meminfo, the kernel’s “estimate of how much
memory is available for starting new applications, without swapping”; a persistent non-zero “full”
line in /proc/pressure/memory is what its documentation calls thrashing.
Record peaks across a real working week, not an idle minute. Two illustrative machines:
Machine A: 16 GB RAM, 8 GB page file
commit limit 16 + 8 = 24.0 GB
peak commit over a week 22.4 GB
peak as share of limit 22.4 / 24.0 = 93% past the 90% growth point
Page Reads/sec sustained during the peak
verdict short of memory; 32 GB changes this machine
Machine B: 32 GB RAM, 8 GB page file
commit limit 32 + 8 = 40.0 GB
peak commit over a week 29.0 GB
peak as share of limit 29.0 / 40.0 = 73%
Task Manager "in use" 85% while gaming
Page Reads/sec near zero
verdict no signal; 64 GB changes nothing
The 80% rule would send Machine B shopping, and it needs nothing. Buy memory for sustained paging under your own workload, not for a high “in use” figure. RAM caching covers the standby list that makes a busy Windows machine look full. Paging is also writes to the drive, a reason to add memory and never to treat an SSD as spare RAM.
Game requirements have settled at 16GB, and the tests side with the minimum
Of seventeen major releases from 2025 and 2026, a selection of big titles rather than a random sample, as their Steam listings stood on September 26, 2026, fourteen ask for 16GB as the minimum. Five recommend 32GB, all from 2025, and none of the five 2026 releases checked - Resident Evil Requiem, Marathon, Crimson Desert, PRAGMATA and 007 First Light - recommends more than 16GB. Escape from Tarkov is the only title found that recommends 64GB.
Two independent tests agree. PC Gamer’s (April 10, 2026) ran a Ryzen 9 9900X and an RTX 5090 on the eight most memory-hungry of 27 games, and four of them “all run just as well with 16 GB of system memory as they do with 32 GB”. TechSpot’s (January 8, 2026) ran a Ryzen 7 9800X3D and RTX 5090 at 4K with Chrome and Discord open, and thirteen of its fourteen tests ran about as well on 16GB as on 32GB or 64GB.
| Title | Released | Minimum | Recommended | Measured by TechSpot, January 2026 |
|---|---|---|---|---|
| Kingdom Come: Deliverance II | Feb 4, 2025 | 16GB | 32GB | about 13.5GB; ran well on 16GB |
| DOOM: The Dark Ages | May 14, 2025 | 16GB | 32GB | about 13.5GB; “no real advantage” from 32GB or 64GB |
| Borderlands 4 | Sep 11, 2025 | 16GB | 32GB | up to 18GB; 16GB “roughly on par” |
| Mafia: The Old Country | Aug 7, 2025 | 16GB | 32GB | about 20GB needed; 16GB “genuinely bad” |
| Stalker 2 | Nov 20, 2024 | 16GB | 32GB | peak about 18GB; ran about as well on 16GB |
| Battlefield 6 | Oct 10, 2025 | 16GB | 16GB | about 11GB on 16GB; up to 14GB with 64GB |
| The Outer Worlds 2 | Oct 29, 2025 | 16GB | 16GB | peak about 15.5GB; no issue on 16GB |
| Clair Obscur: Expedition 33 | Apr 24, 2025 | 8GB | 16GB | 13GB at most |
| Escape from Tarkov | on Steam Nov 15, 2025 | 16GB | 64GB | tested by PC Gamer; see the next section |
The measured column is memory in use, mostly with more than 16GB installed, so a figure above 16GB does not by itself mean that 16GB failed; Mafia: The Old Country is the one title where it did. “Recommended” on a store page is not “required”. With a graphics card of 16GB or more and other programs closed, 16GB DDR5 kits run nearly everything. The next section is about everyone else, which is why this guide’s target for a new gaming desktop is 32GB, or 32GB of DDR4 on a platform that takes it.
Where 16GB runs out: a heavy title, a small graphics card or a second program
Average frame rates barely move between 16GB and 32GB. The 1% lows - the slowest frames, which a player feels as stutter - move in three named cases.
A heavy title. TechSpot called Mafia: The Old Country the first game it had tested where 16GB was not enough: “around 20GB of system memory is required for a smooth experience, meaning a 24GB to 32GB kit is recommended. The experience with 16GB was genuinely bad.”
A small graphics card. On a 12GB RTX 5070 in PC Gamer’s test, 32GB lifted the 1% lows far more than the averages:
1% lows, 32 GB against 16 GB, RTX 5070 (PC Gamer, April 2026)
Escape from Tarkov 63 / 47 = 1.34 34% higher
Stalker 2 51 / 46 = 1.11 11% higher
MSFS 2024 6 / 2 = 3.00 PC Gamer's "200% increase"
Star Citizen 28 / 15 = 1.87 87% higher
An 8GB card is worse, because what does not fit in video memory lands in system RAM. With the 8GB Radeon RX 9060 XT, TechSpot’s system consumed 23GB, and going from 16GB to 32GB lifted the card’s 1% lows in Spider-Man 2 by more than 60%, while the 16GB version was “largely unchanged”. That is the common case: in Steam’s August 2026 survey of Windows machines, 68.89% had 8GB of video memory or less (Valve’s all-platform table gives 46.35%). The strongest measured case for 32GB in gaming is a graphics card with 8GB or less of its own memory; the other fix is a card with more video memory.
A second program, which is how most people play. PC Gamer measured that on 16GB its heavy games “leave between 3 and 5 GB of memory available for anything else”, and warned that for streamers “16 GB of system memory will almost certainly be a problem if you like to use maximum quality settings in some games”. A voice client and a few tabs fit; a streaming encoder, a browser playing video and a heavy title on a mid-range card do not.
Integrated graphics borrows system memory and needs both channels: Intel’s Arc graphics branding on H-series Core Ultra systems requires “at least 16GB of system memory in a dual-channel configuration”. On a laptop with one slot filled, a second DDR5 laptop module is worth more to the graphics than a bigger first one.
Esports titles stay light: Counter-Strike 2’s only stated figure on Steam is 8GB, and Riot lists 4GB for VALORANT. For a machine that only games, 64GB buys nothing the tests could find; it earns its place when the same machine edits, compiles or runs a model. 16GB fails first when something else is running, and something else is usually running.
Vendor, benchmark and workstation builder disagree on creative work by a factor of three
Vendors’ figures are floors for one application on a short project: Adobe’s Premiere page (September 16, 2026) recommends 16GB “for HD media” and “32 GB or more for 4K and higher”. Puget Systems’ benchmark isolates capacity: its test of June 1, 2026 limited a Ryzen 9 9950X3D2 to 32GB and then 16GB, so nothing but capacity changed; 16GB lost ground in every application, and 32GB was level with 64GB in most, Lightroom’s AI tools being the exception at 15% slower. Puget’s workstation pages size for long projects with several applications open: 64GB for 1080p footage in Premiere, 96GB for 4K, 128GB for 6K, 192GB for 8K and 256GB for 12K (August 12, 2026), the same for Resolve.
| Application | Vendor’s figure | Puget benchmark, 16GB against 64GB (June 2026) | Puget workstation recommendation |
|---|---|---|---|
| Premiere | 8GB; 16GB for HD; 32GB or more for 4K (Adobe, Sep 16, 2026) | 7% slower; LongGOP 12%; RAW 14% | 64GB at 1080p; 96GB at 4K (Aug 12, 2026) |
| After Effects | 16GB; 32GB or more for 4K (Adobe, Sep 9, 2026) | 43% slower; 2D compositions 58% | none in the sources |
| Photoshop | 8GB; 16GB or more (Adobe, Sep 9, 2026) | 20% slower; filters 31% | none in the sources |
| Lightroom Classic | no Adobe figure cited | 45% slower; export 118% worse; 32GB 15% slower on AI tools | 32GB minimum; 64GB “safer” (Mar 26, 2026) |
| DaVinci Resolve | 16GB; 32GB for Fusion (ClipVerdict, Aug 2026) | 9% slower; Fusion 31% | same table as Premiere (Mar 26, 2026) |
| Blender | 8GB; 32GB recommended (Blender Foundation) | not tested | none in the sources |
| Pro Tools | 16GB (Avid, Jul 9, 2026) | not tested | none in the sources |
| Ableton Live | no official figure cited | not tested | 64GB minimum; 128GB or 256GB for many (Jul 31, 2025) |
For 4K editing in Premiere the range runs from Adobe’s 32GB to Puget’s 96GB: 96 / 32 = 3. Each end is right about its own case, and Puget says its benchmark projects do not reflect real edits. Proxy media, lower-resolution copies of the footage that cost disk rather than memory, is the free alternative Tom’s Guide recommends (March 8, 2026). Two modules at full speed top a desktop out at 2 x 64GB = 128GB; past that are four modules at the derated speed and then a Xeon W or Threadripper PRO workstation on registered DDR5.
The benchmark sets the floor for professional work at 32GB, and the builder’s figure is what long projects consume: 64GB for most professionals, as 64GB kits of two modules, and 96GB or 128GB only for 4K timelines and heavy compositing.
Developers pay for each virtual machine twice: once in the guest and once in the default
The largest default claim on a development machine’s memory is hidden: Microsoft Learn’s WSL page (September 16, 2026) says WSL 2 assigns “50% of total memory on Windows” to its Linux virtual machine unless told otherwise. Docker Desktop for Windows asks for 8GB; Android Studio lists 16GB with the Android Emulator and recommends 32GB; Visual Studio 2026 “works best with 64 GB RAM”; Epic recommends 32GB for Unreal Engine 5.8.
32 GB Windows laptop, all defaults
WSL 2 virtual machine 32 x 0.5 = 16 GB
left for Windows 32 - 16 = 16 GB
Android Studio with emulator 16 GB system minimum (Google)
plus the browser, the chat client, Docker Desktop and the IDE's indexer
result the host pages as soon as the emulator
and a container run together
Fix 1: cap WSL 2 at 8 GB
WSL 2 virtual machine 8 GB
left for Windows 32 - 8 = 24 GB
Fix 2: a 64 GB desktop, defaults left alone
WSL 2 virtual machine 64 x 0.5 = 32 GB
left for Windows 64 - 32 = 32 GB
The cap is one setting in the .wslconfig file in the Windows user profile:
[wsl2]
memory=8GB
For development, 32GB is the working floor and 64GB is where the defaults stop fighting each other; the cheapest fix before buying anything is capping WSL’s share. On a desktop the second fix is two modules from 64GB DDR5 kits rather than four from 32GB kits; on a soldered laptop only the first exists after purchase.
Local AI is sized by arithmetic: parameters times bits, plus a cache that grows with context
A language model’s memory can be calculated before buying anything. The weights take parameters times bits per weight, divided by eight. The llama.cpp project’s README lists Llama 3.1 8B at 14.96GiB in F16 and 4.58GiB in Q4_K_M, at 4.8944 bits per weight, and says “memory and disk requirements are the same”; a 70B-class model of 70.6 billion parameters (an assumption) takes 70.6e9 x 4.8944 / 8 = 43.2GB at Q4_K_M, which the README’s own 43.1GB for Llama 3.1 70B confirms.
| Model | Parameters | F16 | Q8_0 | Q4_K_M | Smallest common capacity for Q4_K_M |
|---|---|---|---|---|---|
| Llama 3.1 8B | 8B | 16.1GB (14.96GiB) | 8.5GB (7.95GiB) | 4.9GB (4.58GiB) | 16GB |
| Qwen3-32B | 32.8B | about 66GB | about 35GB | about 20GB | 32GB tightly, 48GB comfortably |
| 70B class | 70.6B, assumed | about 141GB | about 75GB | about 43GB | 64GB at short context; 96GB for long context |
| Llama 4 Scout | 109B total, 17B active | about 218GB | about 116GB | about 67GB | 96GB |
The last column adds the cache at 4,096 tokens and about 8GB for the operating system and tools, an assumption: 4.9 + 0.5 + 8 = 13.4GB for the 8B model, 20.1 + 1.1 + 8 = 29.2GB for Qwen3-32B and 43.2 + 1.3 + 8 = 52.5GB for the 70B; Scout’s row uses a 2GB placeholder for its cache, and AMD’s own figure for Scout at Q4_K_M, “roughly 96GB of VRAM”, lands on the same 96GB. At 32,768 tokens the sums grow to 17.2GB for the 8B model, past a 16GB machine, and 61.9GB for the 70B, which fills 64GB.
The cache grows with every token of context. Nvidia’s formula, with the hidden size replaced by the key-value heads times their 128 dimensions, gives 2 x layers x key-value heads x 128 x 2 bytes per token at 16 bits; Meta gives Llama 3’s 8B model 32 layers and its 70B 80, each with 8 key-value heads, and Qwen gives Qwen3-32B 64 layers and 8:
per token, 16 bits = 2 x layers x key-value heads x 128 x 2 bytes
Llama 3.1 8B 2 x 32 x 8 x 128 x 2 = 131,072 bytes = 128 KiB
Qwen3-32B 2 x 64 x 8 x 128 x 2 = 262,144 bytes = 256 KiB
Llama 3.x 70B 2 x 80 x 8 x 128 x 2 = 327,680 bytes = 320 KiB
4,096 tokens 32,768 tokens 131,072 tokens
Llama 3.1 8B 0.5 GiB 4 GiB 16 GiB
Qwen3-32B 1 GiB 8 GiB 32 GiB
Llama 3.x 70B 1.25 GiB 10 GiB 40 GiB
At long context a large model’s cache is as big as its weights: 40GiB, about 43GB, beside 43.2GB for a 70B model at four bits and 131,072 tokens. Ollama’s default context now follows video memory, reaching 256k tokens at 48GiB or more, so set the context deliberately; a q8_0 cache “uses approximately 1/2 the memory of f16”.
Mixture-of-experts models save compute, not capacity. AMD says that for Llama 4 Scout, with 17 billion parameters active per token, “the footprint is the same as a dense 109 billion parameter model”, so Scout cannot run in “about 10GB”, as some 2026 guides say: even at two bits its weights are 27.3GB.
LM Studio recommends 16GB or more for local models. A 70B model at four bits wants two 48GB DDR5 modules from the 96GB kits, fits tightly in 64GB at short context, and at eight bits (75.0 + 1.3 + 8 = 84.3GB) fits tightly in 96GB and comfortably in 128GB. Capacity follows total parameters; speed follows active parameters.
Capacity decides whether a model runs, and bandwidth decides how fast
Generating each token means reading every active weight once, so peak memory bandwidth divided by the bytes read per token is a hard ceiling on tokens per second. A dual-channel desktop at DDR5-6000 has 6,000 x 8 bytes x 2 channels = 96GB/s (the speed guide derives it). The machines built for local AI break the two-channel limit with memory soldered next to the processor; a graphics card is the other route, limited by capacity, since Nvidia’s RTX 5090 has 32GB and a 43GB model does not fit.
| Machine | Memory | Peak bandwidth | Ceiling, 70B at Q4_K_M (43.2GB per token) | Ceiling, Scout at Q4_K_M (10.4GB per token) | Holds Scout (66.7GB)? |
|---|---|---|---|---|---|
| AM5 desktop, 2 x 64GB at rated DDR5-5600 | 128GB | 89.6GB/s | 2.1 tokens/s | 8.6 tokens/s | yes |
| AM5 desktop, 2 x 64GB at EXPO DDR5-6000 | 128GB | 96GB/s | 2.2 | 9.2 | yes |
| AM5 desktop, 4 x 64GB at rated DDR5-3600 | 256GB | 57.6GB/s | 1.3 | 5.5 | yes |
| Ryzen AI Max+ 395 | 128GB | 256GB/s | 5.9 | 24.6 | yes |
| DGX Spark | 128GB | 273GB/s | 6.3 | 26.3 | yes |
| MacBook Pro, M5 Max | 128GB | 614GB/s | 14.2 | 59 | yes |
| Mac Studio, M5 Ultra | up to 512GB (512GB from late October) | 1,200GB/s | 27.8 | 115 | yes |
Every ceiling is bandwidth divided by bytes per token (96 / 43.2 = 2.2), and Scout’s 10.4GB is its 17 billion active parameters at Q4_K_M. These are upper bounds, not measurements, and real systems land below them: AMD’s figure of up to 15 tokens per second for Scout on the Ryzen AI Max+ 395, given without a quantization, is 61% of the ceiling at the Q4_K_M its own FAQ uses for Scout.
Beside a graphics card, system memory holds whatever the card cannot. Puget Systems’ “first rule of thumb” for machine-learning workstations (June 30, 2026) is at least double the system memory of the total video memory: 32GB beside a 16GB card, 48GB beside a 24GB card and 64GB beside a 32GB card. When the model does not fit, the layers left over are read from system memory for every token, and that part alone caps the rate:
70B class at Q4_K_M, 43.2 GB of weights, DDR5-6000 system memory at 96 GB/s
24 GB card, weights kept on the card (assumption) 22 GB
left in system memory 43.2 - 22 = 21.2 GB
ceiling from system memory alone 96 / 21.2 = 4.5 tokens/s
system memory needed 21.2 + 8 (OS, tools) = 29.2 GB: 32 GB, barely
32 GB card, weights kept on the card (assumption) 30 GB
left in system memory 43.2 - 30 = 13.2 GB
ceiling from system memory alone 96 / 13.2 = 7.3 tokens/s
system memory needed 13.2 + 8 = 21.2 GB: 32 GB
no card, the whole model in system memory 96 / 43.2 = 2.2 tokens/s
Offloading a 70B model to a 24GB card roughly doubles a desktop’s ceiling, and a 32GB card more than triples it. For an 8B model the desktop is fine: 96 / 4.9 = about 20 tokens per second as a ceiling.
Two modules, not four: the population rules that set a desktop’s capacity steps
A desktop processor has two memory channels and most boards four slots, and filling all four costs speed. AMD rates its Ryzen 9000 processors, the Ryzen 7 9800X3D and Ryzen 5 9600X among them, at DDR5-5600 with two modules and DDR5-3600 with four, with 256GB as the maximum; Intel’s Core Ultra 200S drops from DDR5-5600 to 4800 or 4400 at two modules per channel.
| Population | Rated speed | Peak bandwidth | Against two modules |
|---|---|---|---|
| AMD Ryzen 9000, two modules | DDR5-5600 | 89.6GB/s | baseline |
| AMD Ryzen 9000, two modules on an EXPO 6000 profile | DDR5-6000, an overclock | 96.0GB/s | 7% more |
| AMD Ryzen 9000, four modules | DDR5-3600 | 57.6GB/s | 36% less |
| Intel Core Ultra 200S, two standard modules | DDR5-5600 | 89.6GB/s | baseline |
| Intel, four single-rank modules | DDR5-4800 | 76.8GB/s | 14% less |
| Intel, four dual-rank modules | DDR5-4400 | 70.4GB/s | 21% less |
Capacity comes from module size, not module count. DDR5 has 24GB and 48GB modules, so 48GB and 96GB are two-slot configurations at full speed, and two 64GB modules reach 128GB; past that, every total takes four at the derated speed. DDR4 vs DDR5 covers the derating by generation, and ranks and 3DS the rank rules.
A 2 x 8GB kit is hard to grow. Newegg’s September 2026 guide: “8GB modules are what you will end up replacing rather than adding to.” A second pair means four modules and the lower rating, and Tom’s Hardware warned on July 31, 2026 that modules bought separately are not validated together. In that Tom’s Hardware test, a single 16GB module cost around 8 to 10% against two across 13 games on most processors (10.9% on a Ryzen 5 7600X) and 2.8% on a 9800X3D, with capacity differing too.
DDR5-6000 on AM5 is an EXPO profile above AMD’s rating, and AMD says operating outside its published specifications “will void any applicable AMD product warranty”; the speed guide covers the profiles. A V-Color “1+1” pack is not a pair but one module and an RGB dummy (Tom’s Hardware, Sep 4, 2026), one of the offers that only look cheaper. Pick the total, then the module size that reaches it in two slots: the 24GB, 48GB and 64GB DDR5 module pages list the routes to 48, 96 and 128GB, and how to check what RAM fits covers what a given machine takes.
32GB was September 2026’s cheapest DDR5 gigabyte, and the step up to it cheaper still
Tom’s Hardware’s US RAM price index of September 14, 2026 lists the single cheapest US listing it found for each kit, a floor rather than a median:
| Kit (DDR5-6000 unless stated) | Best US price | Per GB | Step | Step cost | Per GB of the step |
|---|---|---|---|---|---|
| 16GB | $239 | $14.94 | |||
| 32GB | $409 | $12.78 | 16 to 32GB | $170 | $10.63 |
| 48GB | $661 | $13.77 | 32 to 48GB | $252 | $15.75 |
| 64GB | $889 | $13.89 | 32 to 64GB | $480 | $15.00 |
| 96GB | $1,699 | $17.70 | 64 to 96GB | $810 | $25.31 |
| 128GB (DDR5-6400) | $2,339 | $18.27 | 96 to 128GB | $640 | $20.00 |
The 16GB kit costs 14.94 / 12.78 = 17% more per gigabyte than the 32GB kit. The sixteen gigabytes that take a machine from 16GB to 32GB cost $170 / 16 = $10.63 each, less than the average gigabyte of either kit. The 128GB row is the least stable: the same index showed $3,399 for it on August 17, $1,060 more. The shape is not universal (in Germany, PC Games Hardware found in August 2026 that the usual discount for doubling capacity had disappeared), so check it on live prices: this site’s 32GB DDR5 kits, 16GB and 64GB pages state the cheapest price per gigabyte, and Trends shows how far each generation’s typical gigabyte has moved since the September index.
Whether to buy 16GB now and add more later, on AM5 at the index’s prices:
Route A: 32 GB now, as 2 x 16 GB $409
Route B: 2 x 8 GB now, a second 2 x 8 GB later $239 + $239 = $478
$69 more than A, and four modules rated at DDR5-3600 instead of 5600
Route C: 2 x 8 GB now, sold later, 2 x 16 GB later $239 + later price - resale
cheaper than A only if later price - resale is under $170
Route B loses on both counts, and Route C wins only if prices fall far enough before the second purchase, which is the subject of buy RAM now or wait, with the causes in the 2026 DRAM shortage. At September 2026 prices, capacity you will need within two years costs less as one kit now than as two kits later.
A missing gigabyte costs more than a missing megatransfer
At the same index a 32GB DDR5-5600 kit cost the same $409 as DDR5-6000, while DDR5-6600 cost $655: a speed premium of 655 - 409 = $246, within $6 of the 661 - 409 = $252 step from 32GB to 48GB.
| Change | Cost at the Sep 14, 2026 index | Measured effect | Source |
|---|---|---|---|
| DDR5-5600 to 6000, 32GB | $0 | about 5% in one CPU-bound game (Ryzen 7 9700X) | TechSpot, Apr 7, 2025 |
| DDR5-6000 to 6600, 32GB | $246 | not measured at this step | - |
| One module to two | the second module | 8 to 10% (10.9% on a 7600X); 2.8% on a 9800X3D | Tom’s Hardware, Jul 31, 2026 |
| 16GB to 32GB | $170 | nothing in most averages; 34% in Tarkov’s 1% lows on a 12GB card; over 60% with an 8GB card | PC Gamer, Apr 10, 2026; TechSpot, Jan 8, 2026 |
A speed shortfall costs a few percent, mostly on processors without the large cache of AMD’s X3D parts; a capacity shortfall costs paging. Buy the speed your platform is rated for, spend the rest on capacity, and never pay a speed premium while you are short of a gigabyte you need. The same trade makes 32GB of DDR4-3200 at $190 a better buy on capacity alone than 16GB of DDR5-6000 at $239, for the 11 to 14% of gaming performance Tom’s Hardware measured DDR4-3200 giving up to DDR5-7200 on Intel’s LGA 1700 processors (July 28, 2026); DDR4 vs DDR5 weighs the platform, and CAS latency the timings.
A soldered laptop is sized once, at checkout, for its whole life
On a machine with soldered or on-package memory there is no second chance, and the makers price the configure-to-order step accordingly:
Apple, 14-inch MacBook Pro, 24 GB to 48 GB $600 / 24 GB = $25.00 per GB
($400 in March 2026 per 9to5Mac; after Apple's June 25 rise, $2,499 at
24 GB per Thurrott and $3,099 at 48 GB per AppleInsider, July 23, 2026)
Apple, M6 Mac mini, 16 GB to 24 GB $200 / 8 GB = $25.00 per GB
(Tom's Hardware, August 25, 2026)
Framework, socketed DDR5 SO-DIMMs $13 to $18 per GB
(Framework, March 2026)
desktop 32 GB DDR5-6000 kit, best price $409 / 32 GB = $12.78 per GB
(Tom's Hardware index, September 14, 2026)
Apple's upgrade against the desktop kit 25 / 12.78 = 1.96 times
a $600 step spread over a five-year life 600 / 60 months = $10.00 a month
The ceilings are fixed too: Intel lists its Core Ultra 7 258V with a 32GB maximum. On a soldered machine the capacity chosen at purchase is the capacity for life, so size it for year four or five: 16GB for office work and study, 32GB for development, creative work or local AI, and not 8GB on Windows. How to check what RAM fits covers telling soldered memory from slots before you buy.
Four home buyers, worked from the workload to the kit
A student’s Windows laptop, meant to last five years, running Office, Teams, a browser and whatever the course adds:
8 GB Surface, measured (The Verge, via TechSpot) 88% in use with ten tabs and two apps
stated minimums, Windows + Teams + Microsoft 365 12 GB
Copilot+ PC requirement (Microsoft) 16 GB
buy 16 GB; 32 GB if the course
involves development or video
the extra 16 GB at $13 to $25 per GB 16 x 13 to 16 x 25 = $208 to $400
(Framework's modules; Apple's upgrades)
A gamer with an 8GB graphics card who streams, buying two modules from 32GB DDR5 kits:
game with an 8 GB card (TechSpot) 23 GB occupied when there is room
encoder, voice client, browser not measured; above zero
16 GB cannot hold 23 GB
32 GB as 2 x 16 GB 32 - 23 = 9 GB left for the encoder,
voice client and browser
step cost at the September 14 index $409 - $239 = $170
A 4K editor in DaVinci Resolve who uses Fusion, buying 64GB, which Puget names as the threshold for Fusion; 32GB is a sound budget choice:
Resolve 21 with Fusion (ClipVerdict) 32 GB
Puget benchmark, Fusion score against 64 GB 16 GB 31% lower; 32 GB 5% lower,
within Puget's normal variation
Puget workstation recommendation for 4K 96 GB
64 GB as 2 x 32 GB at the index $889
96 GB as 2 x 48 GB at the index $1,699
the extra 32 GB $810, or 810 / 32 = $25.31 per GB
A home server running Proxmox with a ZFS pool and a few virtual machines, a machine often built from ex-server parts:
Proxmox VE, OS and services (admin guide, Sep 18, 2026) 2 GB
ZFS, about 1 GB per TB of used storage: 16 TB (assumption) 16 GB
guests: a Windows VM and a Linux VM for containers 24 GB assumption
total 42 GB: 64 GB
64 GB of DDR4 ECC RDIMM at eRacks' $5.56 per GB 64 x 5.56 = $356
64 GB of DDR5 RDIMM at eRacks' $31.05 per GB 64 x 31.05 = $1,987
64 GB as a desktop DDR5-6000 kit (index, September 14) $889
A plain file server needs far less: TrueNAS’s hardware guide (September 11, 2026) asks for “at least 8 GB of RAM for basic TrueNAS operations with up to eight drives”, 1GB for each drive after that, at least 16GB and preferably 32GB when the pool serves virtual machines over iSCSI, and “the suggested 5 GB per TB of storage for deduplication”. Spare memory becomes read cache (RAM caching covers the ZFS ARC), and TrueNAS notes that most users “strongly recommend ECC RAM” (ECC vs non-ECC weighs that). Every platform’s figures, TrueNAS, Proxmox, Unraid and each Synology model, are in how much RAM for a NAS or homelab. A home server is sized from its pool and its guests, and registered DDR4 is usually the cheapest gigabyte it can take. The modules are on registered DDR4, and which ones a server accepts, a Dell PowerEdge R730 or an HPE ProLiant DL380 Gen9 for instance, is on the table of server models. The pool itself is sized in the sister site’s how much storage you need, and its drives chosen in drives for a NAS.
A company fleet needs memory tiers by role, and Microsoft has already written them down
Windows 365’s size guidance (June 2, 2026) is the closest thing to a published fleet standard: 4GB for frontline and call-center staff, 8GB for home and consultant users, 16GB for finance and healthcare, 32GB for developers, engineers and content creators, 64GB for software development and data analysis, and 128GB for AI model development that does not lean heavily on a GPU. Microsoft’s Azure Virtual Desktop guidance gives 8GB, 16GB and 32GB for light, medium and heavy work, the heavy tier meaning software engineers and content creators. Two 2026 facts move the bottom rung up: a Copilot+ PC needs 16GB, and Microsoft’s own 8GB Surface froze on Teams calls.
An illustrative 1,000-seat fleet of 200 task workers, 650 knowledge workers and 150 developers or creators shows what tiering is worth, priced at Framework’s March 2026 retail range for DDR5 SO-DIMMs, $13 to $18 per gigabyte, since no fleet price for memory is published:
| Policy | Memory across 1,000 seats | At $13 per GB | At $18 per GB |
|---|---|---|---|
| Microsoft’s tiers, 8 / 16 / 32GB | 16,800GB | $218,400 | $302,400 |
| This guide’s tiers, 16 / 16 / 32GB | 18,400GB | $239,200 | $331,200 |
| Everyone at 16GB | 16,000GB | $208,000 | $288,000 |
| Everyone at 32GB | 32,000GB | $416,000 | $576,000 |
Everyone on 32GB costs $176,800 to $244,800 more than this guide’s tiers, and everyone on 16GB leaves 150 developers below Microsoft’s heavy tier. Buy 16GB as the floor for every Windows seat bought in 2026 and 32GB for the named heavy personas.
Workstation seats are sized by the size of the work. Puget Systems’ SOLIDWORKS rule (April 29, 2026) is “about 5GB of RAM for Solidworks itself, then at least 20 times the largest assembly size you work with”. For data analysis, Wes McKinney, the author of pandas, wrote in September 2017 that “you should have 5 to 10 times as much RAM as the size of your dataset”, an upper bound now that pandas has Arrow-backed types. Machine-learning seats take Puget’s rule of at least twice the graphics card’s memory.
CAD seat, largest assembly 1.1 GB (Puget's rule) 5 + 20 x 1.1 = 27 GB: 32 GB
CAD seat, largest assembly 2.5 GB 5 + 20 x 2.5 = 55 GB: 64 GB
analyst, 8 GB dataset in pandas (McKinney, 2017) 8 x 5 to 8 x 10 = 40 to 80 GB
ML seat with two 16 GB cards (Puget) 2 x (2 x 16) = 64 GB
Past 128GB a seat needs a Xeon W or Threadripper PRO workstation on registered DDR5. Size a workstation seat from its largest file, not from its job title.
Virtual desktops are sized per session, then rounded up to fill the memory channels
Virtual desktops come in two models whose memory needs differ by about an order of magnitude. Multi-session hosts share one operating system among many users: Microsoft suggests at most 6 users per vCPU for light work, 4 for medium, 2 for heavy and 1 for power users, on a minimum virtual machine of 8 vCPUs and 16GB for the first three and 6 vCPUs with 56GB for power users; its example instance, Azure’s D8s_v5, has 32GiB. Single-session desktops give each user a whole virtual machine: Servermall’s VDI sizing guide (May 20, 2026) gives 4 to 6GB per desktop for office work, 8 to 12GB with video calls and 16 to 64GB for CAD, plus service machines, the hypervisor and a reserve of 15 to 30%.
50 medium users
multi-session 50 / 4 users per vCPU = 12.5 vCPUs: two 8-vCPU hosts
2 x 16 GB, Microsoft's minimum = 32 GB
2 x 32 GiB, its example VM = 64 GiB
single-session 50 x 16 GB, Microsoft's tier = 800 GB
50 x 8 to 12 GB, Servermall = 400 to 600 GB
plus 24 GB of service machines and hypervisor
(assumption), 15 to 30% reserve = 488 to 811 GB
The rest is rounding: a server socket is fastest with every channel filled by identical modules. On a 12-channel socket the balanced totals are 384GB, 768GB, 1,152GB and 1,536GB, so 488 to 811GB lands on 768GB, with a 23% reserve instead of 30% at the top. A cluster must also be sized for the host it loses: with three hosts of 384GB, any two can carry all 50 users, 1,152GB in all, the cheapest design here that survives a failure. Size virtual desktops by concurrent sessions, not headcount, then round up to a channel-balanced total. Server memory capacity limits covers population and interleaving, and RDIMM, UDIMM and LRDIMM the module types.
Servers are sized per core and per channel, and the cloud has published the ratios
The public clouds publish their memory-to-processor ratios: AWS’s eighth-generation AMD instances step at 2, 4 and 8GiB per vCPU for compute-optimized, general-purpose and memory-optimized work, and Azure’s Dsv6 runs at 4GiB per vCPU and its Esv6 at 8GiB. A vCPU is usually a hardware thread, so 4GiB per vCPU is 8GiB per core on AWS’s m8i but 4GiB on the m8a, which presents whole cores. “GB per vCPU” and “GB per core” must never be mixed in one plan.
On a 12-channel socket the module size and the core count set the ratio. AMD’s reference two-socket EPYC 9965 system carries 1.5TB as 24 x 64GB across 384 cores, 4GB per core; Intel lists its 128-core Xeon 6980P with up to 3TB per socket, a ceiling of 24GB per core.
| Module, 12 channels, one per channel | Per socket | GB per core at 192 cores | at 128 cores | at 64 cores |
|---|---|---|---|---|
| 32GB | 384GB | 2 | 3 | 6 |
| 64GB | 768GB | 4 | 6 | 12 |
| 96GB | 1,152GB | 6 | 9 | 18 |
| 128GB | 1,536GB | 8 | 12 | 24 |
| 256GB | 3,072GB | 16 | 24 | 48 |
Allocated memory overstates use: Microsoft Azure’s Pond study (2022) found that “∼50% of all VMs touch less than 50% of their rented memory”. Size by measured active memory per core, then round up to a total that fills every channel. Per-socket ceilings are in server memory capacity limits, and the table of server models gives each machine’s slots: a two-socket EPYC 9005 machine such as the Dell PowerEdge R6725 has 24, one per channel.
A database wants its working set in memory, and not a gigabyte more than its edition can use
Microsoft’s guidance for SQL Server (September 10, 2026) is to set max server
memory to “75% of available system memory not consumed by other processes,
including other instances”, and SQL Server 2025 Standard caps the buffer pool at
256GB per instance. PostgreSQL’s documentation gives “25% of the memory in your
system” as a starting value for shared_buffers on a dedicated server, says more
than 40% is unlikely to work better, and warns that total work_mem use “could be
many times the value of work_mem”.
SQL Server 2025 Standard, one instance
buffer pool cap 256 GB
host memory at the 75% rule 256 / 0.75 = 341 GB
balanced single-socket host 12 x 32 GB = 384 GB
12 x 64 GB = 768 GB adds nothing to the buffer pool past 341 GB
SQL Server Enterprise on a 768 GB host
max server memory, before thread stacks 768 x 0.75 = 576 GB
PostgreSQL on a 256 GB dedicated server
shared_buffers to start 256 x 0.25 = 64 GB
little gain past 256 x 0.40 = 102.4 GB
work_mem at concurrency (illustrative) 200 sessions x 2 sorts x 64 MB = 25.6 GB
Those 200 sessions are illustrative: concurrency, not the buffer setting, is what runs a database out of memory. On a virtual machine Microsoft calls a min server memory value “essential”, which rules out overcommitting memory beneath a database guest, and on SQL Server, Page life expectancy falling while Page reads/sec rises under the real workload means a buffer pool smaller than its working set. A database is sized from its working set and its concurrency, not the size of its data. ECC vs non-ECC covers why a database host wants ECC.
AI inference servers carry more host memory than GPU memory, and the cache is what grows
Nvidia’s DGX B200 pairs 1,440GB of GPU memory with 2TB of system memory, configurable to 4TB, on two Xeon Platinum 8570 processors with 112 cores between them, and its GB200 NVL72 rack carries 17TB of LPDDR5X beside 13.4TB of HBM3E:
DGX B200
host memory against GPU memory 2,048 / 1,440 = 1.42 times
at the 4 TB option 4,096 / 1,440 = 2.84 times
host memory per CPU core 2,048 / 112 = 18.3 GB
GB200 NVL72 rack, LPDDR5X against HBM3E 17 / 13.4 = 1.27 times
Puget Systems’ rule for language-model servers (July 16, 2026), “NVIDIA (and us) recommends at least 2 x the amount of CPU system memory as there is total GPU VRAM”, sits above DGX B200’s standard 1.42 times; only the 4TB option clears it. What grows is the cache, which is linear in the number of sequences served at once:
Llama 3.x 70B, 320 KiB per token, cache at 16 bits
one sequence, 4,096 tokens 1.25 GiB
64 sequences at once 64 x 1.25 GiB = 80 GiB
one sequence, 32,768 tokens 10 GiB
16 sequences at 32,768 tokens 16 x 10 GiB = 160 GiB
cache stored at 8 bits half: 40 GiB and 80 GiB
the model's weights at four bits about 43 GB
The Register noted on April 1, 2026 that “it’s quite common for inference engines to store KV caches at FP8”, the halving in the second-last line. In inference, memory is sized from concurrency times context as well as from the weights, and at high concurrency the cache can outgrow them. The host side is built from registered DDR5, 256GB DDR5 modules and, where bandwidth matters more than capacity, MRDIMM.
When a server gigabyte costs $30 to $75, measurement is the cheapest capacity there is
What a DDR5 server gigabyte cost in September 2026 depends on who measures it:
| Measure | Figure | Per GB | Source and date |
|---|---|---|---|
| Median across memory lines of each line’s lowest street price, DDR5 RDIMM | $31.05 | eRacks, via a PR.com release, Sep 22, 2026 (a seller’s own tracking) | |
| Median of 16 in-stock Amazon US listings, DDR5 RDIMM | $41.09 | DatacenterDisk’s Server RAM Price Index, Sep 26, 2026, as relayed by Capital and Compute | |
| Module spot price, 32GB DDR5 RDIMM | $2,050 | $64.06 | DRAMeXchange, Sep 14, 2026 |
| One 256GB DDR5-6400 RDIMM at retail | about $19,000 | $74.22 | Tom’s Hardware, Sep 16, 2026 |
The same eRacks tracking put DDR4 ECC RDIMMs at $5.56 per gigabyte, down 25% since March while DDR5 RDIMMs rose 27%; Trends follows each generation’s typical gigabyte on this site. VMware’s April 2026 guidance for Advanced Memory Tiering is to “keep active memory at 50% or less of your DRAM capacity”, with one gigabyte of NVMe per gigabyte of DRAM by default on drives of endurance Class D or higher. A two-socket host shows what measuring first is worth:
two-socket host, 24 channels, one module per channel
24 x 64 GB = 1,536 GB
24 x 32 GB = 768 GB
the difference = 768 GB
at eRacks' median, $31.05 per GB 768 x 31.05 = $23,846
at the 256 GB module's $74.22 per GB 768 x 74.22 = $57,001
measured, illustrative: 350 GB active across VMs allocated 1.2 TB
VMware's rule, active at most half of DRAM 768 x 0.5 = 384 GB; 350 is under it
NVMe tier at the default 1:1 at least 768 GB of Class D NVMe
memory presented to the VMs 768 + 768 = 1,536 GB
Unmeasured, the 1,536GB configuration is bought on the allocated figure; measured, the smaller host is safe for everything except a memory-active database and the VMs VMware keeps out of tiering: latency-sensitive, fault-tolerant and encrypted-memory security VMs, and VMs of 512GB or 128 vCPUs and more. Every gigabyte not bought because it was measured idle is worth $30 to $75 in 2026. Existing hosts that take DDR4 can be filled from registered DDR4, new ones from registered DDR5, and drives for a memory tier are on the sister site’s NVMe pages.