Thursday, December 25th 2025

Microsoft's Hardware-Accelerated BitLocker Brings Massive Performance Gains

Microsoft has unveiled a new hardware-accelerated BitLocker encryption system for Windows 11, which shifts cryptographic operations from software to dedicated accelerator units integrated into future CPU microarchitectures. Historically, software-based BitLocker in Windows 11 has caused massive performance degradation. For example, in Windows 11 going from no BitLocker to software-based BitLocker has caused average number of cycles per I/O to skyrocket from roughly 400,000 cycles to about 1.9 million cycles. This is a 375% increase in cycles per I/O, and can cause significant storage performance degradations. However, Microsoft is finally bringing hardware-based encryption to solve this.

The new hardware acceleration, announced at Ignite 2025 in November, is now available in Windows 11 version 25H2 and Windows Server 2025 with the September update. Early testing shows that some workloads experience double the storage performance while reducing CPU usage by over 70%. The system offloads AES-XTS-256 encryption processing from the main processor to a fixed-function cryptography engine embedded within the SoC. Encryption keys are hardware-wrapped to enhance security against memory-based attacks. The initial rollout will target Intel vPro platforms with the upcoming Core Ultra Series 3 "Panther Lake" processors, but Microsoft plans to extend support to other vendors as well.
Performance data indicates that sequential read and write speeds are similar between software and hardware approaches. However, random 4K operations show significant improvements with hardware acceleration. In the RND4K Q32T1 read and write tests, hardware-accelerated BitLocker is 2.3 times faster. For single-queue random reads, hardware-based encryption is approximately 40% faster, and for single-queue random writes, it is about 2.1 times faster. These results demonstrate that hardware-based acceleration significantly enhances small-block random performance. These random access patterns are crucial for modern multitasking, which explains why the previous software-only implementation experienced the most significant slowdowns.
Source: Microsoft
Add your own comment

62 Comments on Microsoft's Hardware-Accelerated BitLocker Brings Massive Performance Gains

#51
TheinsanegamerN
CosmicWandererFor "most people", having it on by default is the better practice, because it keeps their data safe even if they don't pay attention to data safety.

It's like Windows Defender being on by default, because "most people" won't bother installing a third-party anti-virus.
Bitlocker does not make your data safer. Unless you are getting your SSD stolen by state actors, bitlocker does nothing for you.
CosmicWandererFor people who know what they want. The option is there to disable BitLocker and they can use whatever drive encryption solution they want.
For people who know what they want. The option is there to enable BitLocker and they can use whatever drive encryption solution they want. Solutions should not be forced on people.
CosmicWandererPray tell what happens when someone breaks into your home and steals your unencrypted PC? Or when someone snatches your unencrypted laptop at a coffee shop the second you look away?

Like that's never happened to people before. But no, according to you, drive encryption on a local drive "is only useful if you actually do want remote access". 😂
Bruh, how often does this happen? And you know what would happen?

First, since my accounts have two factor, they cant get into things anyway. I would change my passwords and remove my laptop as a trusted device. Now they cant do much to any of my accounts. I dont save my passwords on my laptop in the first place, that is only done on my desktop at home, so they're not getting much anyway.

None of this justifies forcing a measure on people who did not ask for it. Fundamentally that action is flawed.
Posted on Reply
#52
R-T-B
CosmicWandererFor "most people", having it on by default is the better practice, because it keeps their data safe even if they don't pay attention to data safety.
If they "don't pay attention to data safety" on by default is awful because they by definition won't be backing up the key and may end up locked out entirely.
TheinsanegamerNBruh, how often does this happen?
Even though I am physically in a place where theft is highly unlikely, I have to consider that exact scenario by law as my business handles government records.

We don't use bitlocker, we aren't even allowed to. I use opal solutions from an approved vendors list. It's honestly not a lot better, but it is faster at least.
Posted on Reply
#53
harm9963
harm9963My M.2 test showed a big hit 7000 vs 6000 , and if you make a mistake updating your BIOS ,forget to pause , long as key is on a USB DRIVE , make sure you have the lastest one , or have to go to Microsoft acount to retrieve , can be a pain , or nerve-racking ordeal for some, currentlly have off , but at some might turn on just to test .
I have it back on , with latest build Beta /Dev , read and write the same with or without , but honestly I never felt a differance ,other than benchmark .

Still needs work on random read and write , read and write are normal .
Posted on Reply
#54
CosmicWanderer
R-T-BIf they "don't pay attention to data safety" on by default is awful because they by definition won't be backing up the key and may end up locked out entirely.
The key is automatically backed up to the users Microsoft Account, which is a requirement for getting through the OOBE. It can't be skipped by "most users".
Posted on Reply
#55
R-T-B
CosmicWandererThe key is automatically backed up to the users Microsoft Account, which is a requirement for getting through the OOBE. It can't be skipped by "most users".
That sounds like awful encryption and is ignoring those who choose for whatever reason a local account. Yes they exist and regardless of how you feel about it, auto encrypting those is borderline criminal.
Posted on Reply
#56
lexluthermiester
R-T-Bauto encrypting those is borderline criminal.
It IS very much criminal is certain situations.
CosmicWandererwhich is a requirement for getting through the OOBE.
So is this.
Posted on Reply
#57
Vayra86
R-T-BThat sounds like awful encryption and is ignoring those who choose for whatever reason a local account. Yes they exist and regardless of how you feel about it, auto encrypting those is borderline criminal.
Mommy and daddy told me that if strangers start saying you really need something, you need to stay as far away as possible. Its really that simple. If I 'want' something, I'll have a genuine reason to want it. This supposed prevention isn't a net positive, it is simple risk mitigation that comes with a loss of control.

Its pretty amazing how little people weigh those two up against each other, but I do, on everything I am supposed to 'buy' or 'install'. Most people don't even get to this dilemma because they have no fucking clue what's happening on their system, and they can't assess the risk either.

What's most striking about all those things big tech say we 'need' is that it happily coincides with their business agenda and agile way of working. 'Oh, we pooped out this update, way beyond the moment we already told you to install this essential stuff, and now its REALLY essential'. How on earth did we ever survive without this essential update/app/insert favorite POS... Case in point being this Bitlocker shit, so we actually had a brutally underperforming application on every PC, and now we have to thank MS for making it less shit AND believe it 'was just fine' prior?
Posted on Reply
#58
R-T-B
lexluthermiesterIt IS very much criminal is certain situations.
Could be, I was speaking in a moral sense despite how it sounded lol.
Posted on Reply
#59
Lordken
_roman_I put in a few hours ago a ryzen 7700x.

The build in benchmark shows here.
Just out of curiosity i ran this on my 5600X , not sure how reliable that benchmark is but it seems odd that your numbers on 7700X are a bit low on PBKDF2 numbers (esp sha256)
btw both your tables are same
cryptsetup benchmark
# Tests are approximate using memory only (no storage IO).
PBKDF2-sha1 4476311 iterations per second for 256-bit key
PBKDF2-sha256 9208131 iterations per second for 256-bit key
PBKDF2-sha512 2818752 iterations per second for 256-bit key
PBKDF2-ripemd160 1263344 iterations per second for 256-bit key
PBKDF2-whirlpool 911805 iterations per second for 256-bit key
argon2i 12 iterations, 1048576 memory, 4 parallel threads (CPUs) for 256-bit key (requested 2000 ms time)
argon2id 13 iterations, 1048576 memory, 4 parallel threads (CPUs) for 256-bit key (requested 2000 ms time)
# Algorithm | Key | Encryption | Decryption
aes-cbc 128b 1382.2 MiB/s 4044.5 MiB/s
serpent-cbc 128b 140.9 MiB/s 945.3 MiB/s
twofish-cbc 128b 275.0 MiB/s 487.3 MiB/s
aes-cbc 256b 1047.6 MiB/s 3699.9 MiB/s
serpent-cbc 256b 141.3 MiB/s 911.2 MiB/s
twofish-cbc 256b 276.4 MiB/s 487.8 MiB/s
aes-xts 256b 5461.8 MiB/s 5444.6 MiB/s
serpent-xts 256b 835.5 MiB/s 846.5 MiB/s
twofish-xts 256b 451.2 MiB/s 467.7 MiB/s
aes-xts 512b 5032.9 MiB/s 5012.7 MiB/s
serpent-xts 512b 831.6 MiB/s 846.7 MiB/s
twofish-xts 512b 458.2 MiB/s 453.3 MiB/s
thats on uname -a
Linux nobara 6.17.12-200.nobara.fc43.x86_64 #1 SMP PREEMPT_DYNAMIC Fri Dec 12 21:00:50 UTC 2025 x86_64 GNU/Linux
Posted on Reply
#60
HasanEsa
_roman_Microsoft discovered stuff which is very long in use, use cpu instructions

As expected wiht windows 11 pro, it does not use the hardware cpu instructions. Why else would windows run according to some posters on intel 3rd generation intel core cpus.

I put in a few hours ago a ryzen 7700x. The build in benchmark shows here.

aes-xts is the default for many things. There was a light impact with my previous obosleted Asus g75vw with intel i7-3610qm, the ryzen 3 3100 or 5800x or 7600x, or now 7700x does not show any impacts in user experience. Maybe Microsoft will now also discover a better file system with binary trees and customiseable selection on different compression rates.

When your main concern is not security, you usually run a benchmark first to decide on which cipher to use. e.g.
Wheat_Nas /home/roman # cryptsetup benchmark
# Tests are approximate using memory only (no storage IO).
PBKDF2-sha1 3266591 iterations per second for 256-bit key
PBKDF2-sha256 3672770 iterations per second for 256-bit key
PBKDF2-sha512 3079518 iterations per second for 256-bit key
PBKDF2-ripemd160 1495828 iterations per second for 256-bit key
PBKDF2-whirlpool N/A
argon2i 22 iterations, 1048576 memory, 4 parallel threads (CPUs) for 256-bit key (requested 2000 ms time)
argon2id 23 iterations, 1048576 memory, 4 parallel threads (CPUs) for 256-bit key (requested 2000 ms time)
# Algorithm | Key | Encryption | Decryption
aes-cbc 128b 1695.7 MiB/s 5151.3 MiB/s
serpent-cbc 128b 165.0 MiB/s 1164.0 MiB/s
twofish-cbc 128b 337.7 MiB/s 710.2 MiB/s
aes-cbc 256b 1289.4 MiB/s 4763.2 MiB/s
serpent-cbc 256b 164.7 MiB/s 1155.7 MiB/s
twofish-cbc 256b 333.7 MiB/s 706.2 MiB/s
aes-xts 256b 7330.5 MiB/s 7267.5 MiB/s
serpent-xts 256b 1047.4 MiB/s 1044.2 MiB/s
twofish-xts 256b 655.5 MiB/s 658.7 MiB/s
aes-xts 512b 6805.4 MiB/s 6801.0 MiB/s
serpent-xts 512b 1051.0 MiB/s 1042.4 MiB/s
twofish-xts 512b 655.7 MiB/s 660.9 MiB/s

I see the only issue when micrisoft is stupid enough to use any TPM1 or TPM2 hardware functionality.

a modern cpu has lots of cpu instructions which should be used, tpm2 module is not something which should be used.

Microsoft discoveres stuff which is very long in use.

I put in a few hours ago a ryzen 7700x.

The build in benchmark shows here.

Wheat_Nas /home/roman # cryptsetup benchmark
# Tests are approximate using memory only (no storage IO).
PBKDF2-sha1 3266591 iterations per second for 256-bit key
PBKDF2-sha256 3672770 iterations per second for 256-bit key
PBKDF2-sha512 3079518 iterations per second for 256-bit key
PBKDF2-ripemd160 1495828 iterations per second for 256-bit key
PBKDF2-whirlpool N/A
argon2i 22 iterations, 1048576 memory, 4 parallel threads (CPUs) for 256-bit key (requested 2000 ms time)
argon2id 23 iterations, 1048576 memory, 4 parallel threads (CPUs) for 256-bit key (requested 2000 ms time)
# Algorithm | Key | Encryption | Decryption
aes-cbc 128b 1695.7 MiB/s 5151.3 MiB/s
serpent-cbc 128b 165.0 MiB/s 1164.0 MiB/s
twofish-cbc 128b 337.7 MiB/s 710.2 MiB/s
aes-cbc 256b 1289.4 MiB/s 4763.2 MiB/s
serpent-cbc 256b 164.7 MiB/s 1155.7 MiB/s
twofish-cbc 256b 333.7 MiB/s 706.2 MiB/s
aes-xts 256b 7330.5 MiB/s 7267.5 MiB/s
serpent-xts 256b 1047.4 MiB/s 1044.2 MiB/s
twofish-xts 256b 655.5 MiB/s 658.7 MiB/s
aes-xts 512b 6805.4 MiB/s 6801.0 MiB/s
serpent-xts 512b 1051.0 MiB/s 1042.4 MiB/s
twofish-xts 512b 655.7 MiB/s 660.9 MiB/s

As expected wiht windows 11 pro, it does not use the hardware cpu instructions. Why else would windows run according to some posters on intel 3rd generation intel core cpus.



I hope microsoft solves the long term issue, the key is in DRAM. I highly doubt they can do it.
logged in just to +1 this reply; (almost) same here in R7 7700. Provided windows just recently discovered that they can use native nvme drivers for nvme drives (the wow! moment), you can readily infer anything "today" is actually "eight years after" in microsoft eco-system (unless, ofc, such things brings good hard $$$, e.g. Copilot everywhere + state-of-the-art telemetries).

EDIT: Add my own cryptsetup benchmarks:

06:46:27 admin@lunarflare-fedora ~ → cryptsetup benchmark
# Tests are approximate using memory only (no storage IO).
PBKDF2-sha1 5282498 iterations per second for 256-bit key
PBKDF2-sha256 11848316 iterations per second for 256-bit key
PBKDF2-sha512 3758336 iterations per second for 256-bit key
PBKDF2-ripemd160 1555750 iterations per second for 256-bit key
PBKDF2-whirlpool 1103764 iterations per second for 256-bit key
argon2i 14 iterations, 1048576 memory, 4 parallel threads (CPUs) for 256-bit key (requested 2000 ms time)
argon2id 14 iterations, 1048576 memory, 4 parallel threads (CPUs) for 256-bit key (requested 2000 ms time)
# Algorithm | Key | Encryption | Decryption
aes-cbc 128b 1571.6 MiB/s 4266.0 MiB/s
serpent-cbc 128b 166.5 MiB/s 1069.5 MiB/s
twofish-cbc 128b 327.8 MiB/s 603.2 MiB/s
aes-cbc 256b 1208.2 MiB/s 4010.0 MiB/s
serpent-cbc 256b 166.6 MiB/s 1069.3 MiB/s
twofish-cbc 256b 326.9 MiB/s 603.9 MiB/s
aes-xts 256b 5741.7 MiB/s 5766.6 MiB/s
serpent-xts 256b 976.9 MiB/s 964.4 MiB/s
twofish-xts 256b 631.7 MiB/s 640.2 MiB/s
aes-xts 512b 5546.7 MiB/s 5536.4 MiB/s
serpent-xts 512b 974.0 MiB/s 964.8 MiB/s
twofish-xts 512b 632.8 MiB/s 642.8 MiB/s


There is one thing I'd note though; even with hardware acceled cryptsetup, there is still significant drop in encrypted write performance on a nvme drive (~45%) unless the drive's native hardware-encryption (OPAL) is used. At least that's what appeared on benchmarks, but usual day-to-day usage doesn't seem so.
Posted on Reply
#61
chrcoluk
Its as people have said, I once started enabling dynamic disks on on my system years ago, some cool software raid enhancement, except when I booted a macrium reflect disk to do a recovery, the recovery file was on a dynamic disk and couldnt be read. I am not interested in bitlocker for similar reasons. Not to mention all the problems it has had.
Posted on Reply
#62
_roman_
LordkenJust out of curiosity i ran this on my 5600X , not sure how reliable that benchmark is but it seems odd that your numbers on 7700X are a bit low on PBKDF2 numbers (esp sha256)
btw both your tables are same
without the input command i can not compare. I highly doubt the latest kernel + gentoo patches + gcc which is a week old from gcc 15 makes a difference.

For those guys who did not use code tags: when you want to put code, it is the same forum syntax, just replace quote tags with code tags.

Before moderators delete stuff we should discuss this in the linux section. I agree stuff needs to be deleted but i hate it.

I wanted to show different ciphers have different speed. I do not see any impact on a daily usage.

I do not agree with encryption has massive impact, I use it for years. KC3000 2TB, ~2.5 years old vs Crucial BX300 120GB SATA SSD bought second same drive in 2017. So most likely a drive which orignated from 2015. that 2017 bought SATA SSD drive is a bit slower, it seems past weeks btrfs etc improved it a bit. The intial setup was a bit slow for only 50% of fill rate for that drive
Posted on Reply
Add your own comment
Nov 21st, 2026 12:48 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts