[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20211026155155.GB3954@fantomas.sk>
Date: Tue, 26 Oct 2021 17:51:55 +0200
From: Matus UHLAR - fantomas <uhlar@...tomas.sk>
To: john-users@...ts.openwall.com
Subject: Re: GPU benchmarks?
>>On Mon, Oct 25, 2021 at 07:19:04PM +0200, Matus UHLAR - fantomas wrote:
>>>BTW, do LWS and GWS matter here? shopuld they be at least shown in the
>>>output?
>On 25.10.21 19:43, Solar Designer wrote:
>>Yes, they matter, and they should probably be included.
>>
>>Ditto for GPU driver version and OpenCL backend version. For NVIDIA
>>GPUs, you'd include things like these:
>>
>>Driver Version: 418.39 CUDA Version: 10.1
>>
>>as printed by nvidia-smi.
>
>I have one nvidia card and laptop with intel and amd GPUs.
>
>I guess these could be important:
>
>% clinfo | grep -e 'Platform Version' -e 'Device Name' -e 'Driver Version' -e 'Max compute units' -e 'Max clock frequency'
> Platform Version OpenCL 1.2 CUDA 11.2.162
> Device Name GeForce GT 1030
> Driver Version 460.91.03
> Max compute units 3
> Max clock frequency 1430MHz
>
>>>so we could start with same hashes as CPU tables have..
>>
>>I think that choice is not ideal. We'd also need to add sha*-crypt to
>>all tables, whereas our LM is in a poor state on CPU (doesn't scale to
>>many threads, except with "--fork") and our bcrypt-opencl is in a poor
>>state on NVIDIA GPUs.
>
>how else are we supposed to see the progress if they are optimised? :)
>
>> However, if you want to show JtR shortcomings
>>along with CPU/GPU differences, then these are fine, they just should
>>not be misinterpreted as indicating the CPUs and GPUs can only achieve
>>roughly those speeds at these two hash types.
On 25.10.21 20:57, Matus UHLAR - fantomas wrote:
>I put a note about this to the wiki:
>
>https://openwall.info/wiki/john/benchmarks#GPU-benchmarks
I have tried those on my laptop with intel and AMD graphics, so far a few
observations:
- it's not as easy to get driver version with opencl
- there are bigger differences real/virtual than one/many salts
descrypt on intel:
Many salts: 5480K c/s real, 7580K c/s virtual
Only one salt: 5447K c/s real, 5447K c/s virtual
descrypt on AMD:
Many salts: 33242K c/s real, 3440M c/s virtual
Only one salt: 28356K c/s real, 268101K c/s virtual
mdcrypt on intel:
Many salts: 173070 c/s real, 244132 c/s virtual
Only one salt: 173070 c/s real, 244132 c/s virtual
mdcrypt on AMD:
Many salts: 284928 c/s real, 2590K c/s virtual
Only one salt: 279552 c/s real, 1996K c/s virtual
bcrypt on intel:
Raw: 364 c/s real, 18618 c/s virtual
bcrypt on AMD:
Raw: 748 c/s real, 273066 c/s virtual
so I better put all of those to the benchmark wiki.
... LM on AMD failed so no benchmark for now:
ac_rtld error: symbol lm_loop: unknown
LLVM failed to upload shader
FAILED (cmp_all(0))
--
Matus UHLAR - fantomas, uhlar@...tomas.sk ; http://www.fantomas.sk/
Warning: I wish NOT to receive e-mail advertising to this address.
Varovanie: na tuto adresu chcem NEDOSTAVAT akukolvek reklamnu postu.
"The box said 'Requires Windows 95 or better', so I bought a Macintosh".
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
