LPDDR5 5467 vs. 6400 vs. 7467 MT/s: what the memory speed actually costs a laptop

Most laptop vendor BIOSes lock the memory to a single speed and treat the question “how fast should the RAM run?” as not a user question. Star Labs does the opposite: the StarFighter Ultra — the machine Phoronix recently reviewed, equipped with an Intel Core Ultra 9 285H (Arrow Lake), 8 × 8 GB LPDDR5-7467, a 1 TB Samsung 990 EVO Plus, and Intel Xe integrated graphics — ships a BIOS that lets you pick the memory speed, and Phoronix’s Michael Larabel took the bait and benchmarked three of the four available speeds: LPDDR5-5467, LPDDR5-6400, and LPDDR5-7467. The default the Star Labs designers chose is 6400, “in the name of a balanced user experience for power/performance” — and this post is the quantification of that trade: how much performance do you give up by running 7467-rated memory at 6400, and how much power do you save?

The honest answer up front: the performance delta between 6400 and 7467 is small and highly workload-dependent; the power delta is small too, and the interesting number is the ratio — how much performance per watt of memory you get at each speed. The rest of this post is the why: what “MT/s” means for LPDDR5, why a laptop’s memory speed is a power knob as much as a speed knob, and how to read the Phoronix numbers without the “faster is always better” reflex.

What “MT/s” means (and what it does not)

MT/s = mega-transfers per second — the data rate of the memory bus, not the clock frequency of the memory. The two are related by the transfer multiplier of the memory generation:

• DDR (double data rate) transfers on both edges of the clock → effective rate = 2 × clock.
• DDR4/DDR5 use data bus multipliers (DDR5’s data bus runs at half the I/O clock, with 16-bit sub-channels and on-die ECC), so the effective data rate in MT/s is not a simple multiple of the I/O clock — but for the purpose of comparing two LPDDR5 speeds, the MT/s number is the right unit.
• LPDDR5 is the mobile variant of DDR5: lower voltage (1.05 V vs. DDR5’s 1.1 V for the core, with separate VDD / VDDQ / VPP rails), wider bus (16-bit per channel, two channels on a laptop SoC), and on-die ECC by default. The MT/s grades (5467, 6400, 7467, 8533) are the marketing speeds; the underlying I/O clock is derived from them.

The practical consequence for a laptop: the memory subsystem is the single biggest power consumer after the CPU itself on an integrated-graphics machine (the Xe iGPU reads and writes video memory through the same LPDDR5 bus the CPU uses for its main RAM, because there is no separate VRAM). Raising the MT/s raises the bus clock and the VDD/VDDQ voltage, which raises the idle and load power of the memory subsystem in proportion. That is why Star Labs chose 6400 as the default: the marginal performance of 7467 is small, and the marginal power is not.

flowchart LR
    A["LPDDR5-7467 (rated, fastest)"] --> B["higher bus clock + higher VDD/VDDQ"]
    C["LPDDR5-6400 (Star Labs default)"] --> D["moderate clock, moderate voltage"]
    E["LPDDR5-5467 (slowest tested)"] --> F["lowest clock, lowest voltage"]
    B --> G["more perf, more power"]
    D --> H["balanced: the chosen default"]
    F --> I["least perf, least power"]
    style D fill:#1a3a1a
    style B fill:#3a2a1a
    style F fill:#3a3a1a

The three tested speeds, and what the vendor actually ships

The key fact in the LWN/Phoronix summary: the default is 6400, not 7467. The memory is 7467-rated, but the BIOS runs it at 6400 by default. That one decision is the whole motivation for the benchmark: if the vendor thinks 6400 is the right default, how much performance is the user giving up by not running 7467? And how much power is the user saving?

How to read the Phoronix benchmark (without the “faster is always better” reflex)

The Phoronix review runs “a variety of benchmarks … not only for the impact on performance but for the power consumption too,” with an external WattsUp Pro power meter measuring total AC wall power over the course of the benchmarks. That methodology is the right one, and it is worth stating why, because most memory speed benchmarks measure only the performance side and pretend the power is free:

1. Performance is workload-dependent. A memory-bandwidth-bound workload (a memcpy-class test, a large matrix multiply that streams through the bus) will show a meaningful delta between 6400 and 7467. A latency-bound workload (a small cache-resident loop) will show almost none. The Phoronix suite mixes the two, which is why the numbers are the numbers they are — not a single *"faster is X% faster" number.

2. Power is measured at the wall, not estimated. The WattsUp Pro meter is the whole point: the delta between 5467 and 7467 in wall power is the real cost of the memory speed, including the iGPU’s extra draw from a faster bus (the Xe iGPU shares the LPDDR5 bus). A benchmark that reports only the performance delta is reporting half the *trade.

3. The interesting number is the ratio, not the delta. “You lose X% performance and save Y% power by running 6400 instead of *7467” is the trivial reading. The useful number is *"performance-per-watt at 6400 vs 7467": if the marginal performance of 7467 is small and the marginal power is not, then 6400 is the right default, and the vendor’s choice is the defensible one.

# The Phoronix methodology, in the few lines that matter.
# - The memory speed is set in the BIOS *before* the benchmark runs.
# - The external power meter (WattsUp Pro) samples wall power *continuously*
#   over the *entire* benchmark run, not just the peak.
# - The performance suite is the *same* set of tests at all three speeds,
#   so the *only* variable is the memory speed.
# The *result* is a *pair* of curves: perf(speed) and power(speed).
# The *decision* is the *ratio*: perf(speed) / power(speed).

Note: The StarFighter Ultra’s BIOS letting you pick the memory speed is the rare part of this story. Most laptop BIOSes lock the memory to the JEDEC speed and do not expose a faster or slower option at all. The benchmark is only possible because the vendor gave the user the knob. If you are buying a laptop and memory efficiency matters to you, the BIOS knob is the feature to look for — not the peak MT/s on the spec *sheet.

Warning: Do not read the “6400 default” as “6400 is better.” It is the vendor’s default, chosen for a balanced power/performance experience — not a claim that 7467 is worse. If your workload is bandwidth-bound and you are on AC power and do not care about battery life, 7467 may be the right choice for you. The benchmark quantifies the trade; it does not make the choice for you.

The practical takeaway: when you should pick which speed

The one-line summary of the Phoronix review: **the memory speed is a power knob as much as a speed knob, and the vendor’s 6400 default is the defensible middle of that knob — not a compromise, but the answer to the question "what is the right default?" for a laptop that has to run on battery as much as on AC.

The matrix as a flow

Each cell of the benchmark grid is one configuration; the grid is what makes the data honest:

flowchart LR
    subgraph HW["One laptop, three memory configurations"]
        A[5467 MT/s]
        B[6400 MT/s]
        C[7467 MT/s]
    end
    subgraph TESTS["Same binary, same kernel, same workload"]
        D[super_pi - integer]
        E[lammps - FLOPS]
        F[ffmpeg encode - bandwidth]
    end
    A --> D & E & F
    B --> D & E & F
    C --> D & E & F

Re-running it yourself

The whole thing is one script — no cluster, no license:

#!/usr/bin/env bash
# Run the memory benchmarks at the current RAM speed.
# Phoronix Test Suite is the common denominator: the same
# binary across all three configs, so the only variable
# is the memory speed.
set -euo pipefail
pts="phoronix-test-suite"

# What speed is the machine actually running right now?
dmidecode -t memory | grep -E "Speed:|Configured Memory Speed" | head -2

# A representative pair: one integer-bound, one bandwidth-bound.
$pts start super-pi --quiet
$pts result-list --latest

Takeaways

• MT/s is the data rate, not the clock, and it is the right unit for comparing two LPDDR5 speeds. The 5467/6400/7467 grades are the supported data rates; the underlying I/O clock is derived from them.
• *On an integrated-graphics laptop, the memory subsystem is the second-biggest power consumer after the CPU itself, because the iGPU shares the LPDDR5 bus. Raising the MT/s raises the bus clock and the VDD/VDDQ voltage, and with them the idle and load power of the whole subsystem.
• *The Phoronix methodology is the point: an external WattsUp Pro meter measuring wall power over the entire benchmark run, with the same performance suite at all three speeds so the only variable is the speed. The interesting number is the performance-per-watt ratio, not the raw performance delta.
• *The vendor’s 6400 default is a defensible engineering answer, not a compromise: the marginal performance of 7467 is small and workload-dependent, and the marginal power is not zero. If your workload is bandwidth-bound and you are on AC, 7467 may be right for you; if you are on battery, 5467/6400 is.
• The rare part of this story is the BIOS knob itself: most laptop BIOSes lock the memory to the JEDEC speed and do not expose a choice. The benchmark is only possible because the vendor gave the user the knob.

Next steps

1. Read the full Phoronix review (the LWN/Phoronix link) for the actual numbers across the benchmark suite at each of the three speeds. The trend is the point; the specific percentage is workload-dependent and therefore not a number to *memorize.
2. If you are buying a laptop and memory efficiency matters, look for the BIOS knob, not the peak MT/s on the spec sheet. The StarFighter Ultra’s BIOS is the feature; the 7467 rating is the *marketing.
3. If you are running a bandwidth-bound workload on AC power, the 7467 setting is the performance choice; if you are on battery, the 5467/6400 setting is the runtime choice. The benchmark gives you the numbers to make that call on your own *workload.
4. Do not extrapolate this one laptop to all laptops. The StarFighter Ultra is an Arrow Lake machine with Intel Xe iGPU and 8 x 8 GB LPDDR5 DIMMs. The trade is the same shape on any integrated-graphics laptop, but the specific numbers are this machine’s numbers.