Catalyst v14.9 vs 14.12

[atom]

---

Performance

---

I made a small comparison chart, see here how bad that 14.12 really is

https://docs.google.com/a/hashcat.net/sp...edit#gid=0

Not sure what to do now...

---

Memory

---

Here's some strange stuff happending to memory size query. This also happens with clinfo which is written from AMD themself:

When X11 is running:

root@et:~/oclHashcat-1.32# clinfo | grep Global
Global memory size: 3167748096
Global memory size: 3168796672
Global memory size: 3168796672

Those 3 GPU's are all the same type. Does not happen with 14.9.

When X11 is stopped:

root@et:~/oclHashcat-1.32# clinfo | grep Global
Global memory size: 3182428160
Global memory size: 3182428160
Global memory size: 3182428160

Not sure if we can call this a bug

---

Performance 2

---

The first GPU is running slower than the next ones:

root@et:~/oclHashcat-1.32# ./oclHashcat64.bin -b -m 0 --markov-disable --restore-disable --outfile-autohex-disable --potfile-disable --logfile-disable --gpu-temp-disable
oclHashcat v1.32 starting in benchmark-mode...

Device #1: Tahiti, 3035MB, 1000Mhz, 32MCU
Device #2: Tahiti, 3035MB, 1000Mhz, 32MCU
Device #3: Tahiti, 3035MB, 1000Mhz, 32MCU

Hashtype: MD5
Workload: 1024 loops, 256 accel

Speed.GPU.#1.: 3615.9 MH/s
Speed.GPU.#2.: 8020.9 MH/s
Speed.GPU.#3.: 8023.1 MH/s
Speed.GPU.#*.: 19659.9 MH/s

Started: Mon Dec 29 17:37:42 2014
Stopped: Mon Dec 29 17:38:01 2014

This is clearly a bug. X11 is not running here. Limiting to -d 1 does not have any effect.

While the benchmark is running I was testing clocks to make sure its not related to them but it's not:

root@et:~/oclHashcat-1.32# amdconfig --odgc --adapter=all

Adapter 0 - AMD Radeon HD 7900 Series
Core (MHz) Memory (MHz)
Current Clocks : 1000 1375
Current Peak : 1000 1375
Configurable Peak Range : [300-1200] [150-1650]
GPU load : 99%

Adapter 1 - AMD Radeon HD 7900 Series
Core (MHz) Memory (MHz)
Current Clocks : 1000 1375
Current Peak : 1000 1375
Configurable Peak Range : [300-1200] [150-1650]
GPU load : 99%

Adapter 2 - AMD Radeon HD 7900 Series
Core (MHz) Memory (MHz)
Current Clocks : 1000 1375
Current Peak : 1000 1375
Configurable Peak Range : [300-1200] [150-1650]
GPU load : 99%

Utilization is fine, too...

Workaround found! See below

---

GPU Ordering

---

For some reason the GPU ordering is "broken". This causes a real problem as there's no unique identifier between OpenCL devices and ADL devices (the runtime that is used for querying Fan speed, GPU temp, etc), therefore we're forced to use some fuzzy logic that tries to guess the correct assignment. But then, when the opencl runtime now randomly returns the devices we have a bigger problem...

Here's how to reproduce:

X11 is running:

$ clinfo | grep " Name"
Name: Hawaii
Name: Juniper
Name: AMD FX(tm)-8120 Eight-Core Processor

And now I stop X11 and run the command again:

$ clinfo | grep " Name"
Name: Juniper
Name: Hawaii
Name: AMD FX(tm)-8120 Eight-Core Processor

GPU's are switched!

---

Major problem, some stuff does not get cracked even with correct password!

---

So this problem is very evil, especially for people using the R9 290x. So far I can 100% reproduce this with my R9 290x while all other cards I tested (hd5770 and hd7970) are not affected. Again, when I switch back to 14.9 everything works fine again.

To test on your system, simply download the example WPA2 .hccap: http://hashcat.net/misc/example_hashes/hashcat.hccap

R9 290x:

root@sf:~/oclHashcat-1.32# ./oclHashcat64.bin -m 2500 hashcat.hccap -a 3 'hashcat!' -d 1 --quiet --potfile-disable

+------------------------------------------------------------------------------+
! ATTENTION! The installed GPU driver in your system is known to run very slow !
! Some algorithms are 13 times slower with the driver installed in your system !
! See here for details: https://hashcat.net/forum/thread-3915.html !
! We encourage you to use a different driver like AMD Catalyst v14.9 !
+------------------------------------------------------------------------------+

hd5770:

root@sf:~/oclHashcat-1.32# ./oclHashcat64.bin -m 2500 hashcat.hccap -a 3 'hashcat!' -d 2 --quiet --potfile-disable

+------------------------------------------------------------------------------+
! ATTENTION! The installed GPU driver in your system is known to run very slow !
! Some algorithms are 13 times slower with the driver installed in your system !
! See here for details: https://hashcat.net/forum/thread-3915.html !
! We encourage you to use a different driver like AMD Catalyst v14.9 !
+------------------------------------------------------------------------------+

hashcat.net:0025cf2db489:b0487ad676e2:hashcat!

**UPDATE** In case AMD dev's see this, this problem occours only in binary mode clCreateProgramWithBinary(). That is, when using SPIR binaries. In sourcecode-mode clCreateProgramWithSource() it works fine.

---

Good Driver Downloads

---

It's important to have set the correct referer, otherwise download will fail

Linux:

wget --referer=http://support.amd.com/de-de/download/desktop?os=Linux+x86 http://www2.ati.com/drivers/linux/amd-ca...x86-64.zip

Windows: http://support.amd.com/en-us/download/de...4&rev=14.9

[epixoip]

On the plus side, I'm pleased with the gains for Office97, that actually brings it up to 980 performance.

But I think some investigation is needed. It's strange to me that DEScrypt would go down so much, yet LM basically stayed the same and Oracle 7g went up. Similarly, raw SHA512 only went down a little bit, but HMAC-SHA512 and sha512crypt especially took massive hits. Something doesn't add up there.

[boyferret]

I made a small comparison chart, see here how bad that 14.12 really is

https://docs.google.com/a/hashcat.net/sp...edit#gid=0

Not sure what to do now...

I noticed that that your results for the r9 290x are not the same as what it shows on the front page, is that due to running Ubuntu?

[epixoip]

The front page only has benchmarks for 8x 290X, not a single 290X.

[atom]

I did some more tests with 14.12 while looking for workarounds. Results edited on first post

[mariaki]

Report them to AMD so they can fix all problems

[atom]

Report them to AMD so they can fix all problems

HAHA, no. You have no idea how often I reported problems to AMD and they never got fixed.

[atom]

SUCCESS!

At least I found a way to workaround the multiple-gpu bug:

export GPU_NUM_COMPUTE_RINGS=1

Speed.GPU.#1...: 8053.8 MH/s
Speed.GPU.#2...: 8062.5 MH/s
Speed.GPU.#3...: 8064.0 MH/s

[epixoip]

Nice job!

[atom]

More crap with catalyst 14.12 found! GPU ordering is broken! See first post for details