And on an M4 Pro I measured ML-DSA-44 signing

OpenSSL 4.0.3, one core.                4,120/s
BoringSSL (oct 2026), one core      11,260/s
BoringSSL (oct 2026), all 12 cores  99,685/s


On Fri, Oct 16, 2026 at 3:26 PM Bas Westerbaan <[email protected]> wrote:

> Nice.
>
> The OpenSSL implementation of ML-DSA-44 is not optimal yet. On an old
> i7-4790 I get on a single core mean for signing
>
> OpenSSL 3.5.5              1,415/s
> BoringSSL (oct 2026)    4,000/s
>
> Throughput of GPU is great, but latency is probably worse. That's ok for
> offline signing, but is of course less than ideal for verification. Did you
> record latency numbers?
>
> Best.,
>
>  Bas
>
> On Fri, Oct 16, 2026 at 3:04 PM Mukund Sivaraman <[email protected]> wrote:
>
>> Hi all
>>
>> ML-DSA-44 signs are an order of magnitude slower than ECDSA P-256. For
>> some months now, Loop has included a Vulkan GPU crypto library and its
>> nameserver has been able to start a Vulkan manager thread. MLDSA44
>> support has been added to it, and Loop's dnssec-signzone and
>> dnssec-verify have been updated to use the async DNSSEC and DST library
>> interfaces that make it possible for end-users to choose GPU signing. I
>> am sharing some notes and measurements from this work.
>>
>> Some OpenSSL 3.5 and Loop GPU library benchmarks follow (extract from a
>> spreadsheet with other algorithms removed):
>>
>> | ML-DSA-44 sign and verify ops/s                          |   signs/s |
>> vs. 24-CPU | verifies/s (*) | vs. 24-CPU |
>>
>> |----------------------------------------------------------|----------:|-----------:|---------------:|-----------:|
>> | OpenSSL 3.5.9 on Intel Core Ultra 9 275HX with -multi 24 |    54,859 |
>>      100% |        272,280 |       100% |
>> | OpenSSL 3.5.9 on AMD Ryzen 5 220 with -multi 12          |    11,656 |
>>       21% |         59,143 |        22% |
>> | OpenSSL 3.5.7 on Raspberry Pi 5 with -multi 4            |     3,524 |
>>        6% |         17,819 |         7% |
>>
>> |----------------------------------------------------------|----------:|-----------:|---------------:|-----------:|
>> | Loop library on NVIDIA RTX 5070 laptop GPU @85W          |   321,920 |
>>      587% |      1,100,688 |       404% |
>> | Loop library *projection* for 2x NVIDIA RTX 5080 GPU (+) | 1,429,565 |
>>    2,606% |      4,953,098 |     1,819% |
>>
>> These are raw library ML-DSA-44 performance numbers without DNSSEC. The
>> Vulkan implementation uses hedged signing and implements the
>> requirements of sections 3.6.x of FIPS 204. Section 3.6.3 requirements
>> are implemented across the program, library, and Vulkan code.
>>
>> (*) OpenSSL's speed benchmark program uses a single key (whose context
>> caches precomputed fields). A single signature is verified
>> repeatedly. It's highly unlikely for a network service to perform many
>> verifications with a single key. For verification, this can greatly skew
>> the results as OpenSSL caches parsed/precomputed fields (as an example,
>> in the case of RSA, expensive per-key precomputation of R^2 mod m used
>> in Montgomery arithmetic) in its key context. For comparison, the <Loop
>> library on NVIDIA 5070 laptop GPU @85W> result has single-key MLDSA44
>> verification measured at 3,007,385 verifies/s (1,105%). The Loop
>> benchmarks in the table used separate raw public keys without
>> precomputation per verification that reflects real-world usage. This is
>> just a clarification, not a note against the "speed" benchmark. :)
>>
>> (+) A single RTX 5090 GPU might deliver a similar result, but it's
>> currently a USD ~7,000+ part because it shares a GB202 with RTX PRO
>> cards with which makes it attractive to aftermarket modders. The RTX
>> 5080 is a USD ~1,500 part and two used together as separate Vulkan
>> physical devices could deliver similar performance as an RTX 5090 for
>> this application, with slightly higher TDP.
>>
>>
>> When a GPU is used as a coprocessor for signing and verification, it
>> frees up the CPU to run application code such as query processing. In
>> the OpenSSL benchmarks above, the CPU is fully occupied performing
>> signatures and verifies. When sharing the CPU with sign/verify ops,
>> application performance can be much lower such as for inline signers and
>> resolver validators, and vice versa. The lack of CPU overhead when using
>> a GPU as a coprocessor can be seen in the dnssec-signzone timings below
>> ("user") for a 1 million RRsets zone:
>>
>> [muks@w2 ~/tmp-signzone]$ cat mldsa44.frankenmodel.com.zone
>> $ORIGIN mldsa44.frankenmodel.com.
>> $TTL 3600
>> @       IN              SOA ns-primary.banu.com. support.banu.com. 18
>> 21600 3600 604800 3600
>>
>>                         NS ns1.banu.com.
>>                         NS ns2.banu.com.
>>
>> $GENERATE 1-1000000 host$ A 127.0.0.1
>> [muks@w2 ~/tmp-signzone]$
>>
>> (a) Loop 1.99.22 (before; CPU signing and post-sign verification)
>>
>> [muks@w2 ~/tmp-signzone]$ time dnssec-signzone -S -z -o
>> mldsa44.frankenmodel.com -e +2y mldsa44.frankenmodel.com.zone
>> Fetching ZSK 46736/MLDSA44 from key repository.
>> Verifying the zone using the following algorithms: MLDSA44.
>> Zone fully signed:
>> Algorithm: MLDSA44: KSKs: 0 active, 0 stand-by, 0 revoked
>>                     ZSKs: 1 active, 0 stand-by, 0 revoked
>> mldsa44.frankenmodel.com.zone.signed
>>
>> real    3m14.693s
>> user    17m13.125s
>> sys     0m15.748s
>> [muks@w2 ~/tmp-signzone]$
>>
>> (b) Loop 1.99.23 (after; GPU signing and post-sign verification)
>>
>> (The first run after install or version upgrade can cause the Vulkan
>> driver to compile shaders and may take some time to complete.)
>>
>> [muks@w2 ~/tmp-signzone]$ time dnssec-signzone -G -Y 1 -S -z -o
>> mldsa44.frankenmodel.com -e +2y mldsa44.frankenmodel.com.zone
>> Fetching ZSK 46736/MLDSA44 from key repository.
>> Verifying the zone using the following algorithms: MLDSA44.
>> Zone fully signed:
>> Algorithm: MLDSA44: KSKs: 0 active, 0 stand-by, 0 revoked
>>                     ZSKs: 1 active, 0 stand-by, 0 revoked
>> mldsa44.frankenmodel.com.zone.signed
>>
>> real    0m20.845s
>> user    1m58.707s
>> sys     0m38.604s
>> [muks@w2 ~/tmp-signzone]$
>>
>> The timings include zone load, dump, and pre-signing steps. The actual
>> sign and verification times were:
>>
>> Signing time in seconds: 8.462
>> Signatures per second: 236325.942
>>
>> Verification time in seconds: 7.785
>> Verifications per second: 256885.220
>>
>> The dnssec-signzone signs/s and verifies/s are lower than the raw
>> library GPU performance on the RTX 5070 laptop GPU because of lower rate
>> of submission by the program. dnssec-signzone has high RBTDB lock
>> contention writing signed records, and its verify is still
>> single-threaded in zone traversal, etc. These limit the rate at which
>> sign/verify requests are submitted to the GPU. There's work to be done
>> to address this.
>>
>> This dnssec-signzone process was run on Fedora 44 x86_64 with
>> loop-1.99.23.20261015053924.dce1b9848c-1.fc44.x86_64,
>> kernel-7.2.10-200.fc44.x86_64, openssl-3.5.9-1.fc44.x86_64,
>> vulkan-loader-1.4.341.0-1.fc44.x86_64,
>> xorg-x11-drv-nvidia-libs-615.78.08-1.fc44.x86_64 (rpmfusion).
>>
>> Although Vulkan compute shaders are used for portability across device
>> vendors and there's nothing non-portable in the code, AMD (RADV), Intel
>> (ARL), and others such as as lavapipe have trouble currently due to
>> shader compilation issues among others. NVIDIA's Vulkan driver does not
>> use Mesa and does not have these issues. During development, I've
>> noticed that shader compilers and SPIR-V tools are particularly
>> problematic, but using Vulkan is advantageous in that it can be used out
>> of the box on Linux distros. RSA, ECDSA P-256 and P-384 sign/verify
>> compute shaders are also present but they currently have a large
>> performance regression due to recent changes. Resolver validation and
>> inline signing are not yet connected to the async interfaces. These will
>> be addressed.
>>
>> On a personal note, this was brutal. I'd been developing this
>> acceleration branch since October 2025 alongside other work. Though at a
>> top-level the concept is straightforward (use GPUs to accelerate
>> DNSSEC), the implementation and considerations were extensive. The
>> Vulkan GPU manager, the design decisions to keep the units continually
>> fed, adding async APIs to the DST and DNSSEC library modules, and using
>> it in dnssec-signzone were all large changes. A/B selection and tuning
>> the methods used in the compute kernels was long and substantial
>> work. The work division approaches that perform well for ECDSA and MLDSA
>> are different at least.
>>
>> A $GENERATE 1-10000 RRsets GPU-signed mldsa44.frankenmodel.com zone is
>> published. It can be queried using any validating resolver with MLDSA44
>> support such as Cloudflare's 1.1.1.1 service.
>>
>>                 Mukund
>> _______________________________________________
>> DNSOP mailing list -- [email protected]
>> To unsubscribe send an email to [email protected]
>>
>
_______________________________________________
DNSOP mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to