Geekbench 5

[_mat_]

Geekbench 5 has landed today. It officially requires Windows 10 (64 bit) for the first time.

I've looked into it from a technical standpoint and it doesn't fix anything related to timers or security. It uses QPC and GetSystemTime(), both highly unreliable and easily attackable, and both will skew with bclock/reference clock on Pre-Skylake and all AMD CPUs on Windows 10. The negligence seriously amazes me.

BenchMate to the rescue! I've already integrated Geekbench 5 into BenchMate and released a new version: BenchMate 0.9.1
Geekbench 5 is also bundled with the "Big Package", no need to install it.

I've also taken the liberty to add Geekbench 5 as a beta benchmark here on HWBOT:

Geekbench5 - Single Core: https://hwbot.org/benchmark/geekbench5_-_single_core/
Geekbench5 - Multi Core: https://hwbot.org/benchmark/geekbench5_-_multi_core/
Geekbench5 - Compute: https://hwbot.org/benchmark/geekbench5_-_compute/

You can directly upload into these categories inside BenchMate or by uploading the .hwbot-file manually.

Edited September 3, 2019 by _mat_

[Splave]

Well damn that was fast

[keeph8n]

Well guess I need to hurry up and get to testing before a certain someone has all the fun 

[keeph8n]

Also noticed that if you want to support the bench itself, they are running a launch sale through the 10th.

[I.nfraR.ed]

I've bought previous Geekbench versions, but I'd rather support Mat than buying the same benchmark again for the third time.

[Mythical tech]

https://www.patreon.com/benchmate

[Guest]

1 hour ago, keeph8n said:

Well guess I need to hurry up and get to testing before a certain someone has all the fun 

Good luck speedy

[_mat_]

I've just encountered a strange synchronization bug on Ryzen with the new Geekbench 5 integration. As I've looked further I realized that instead of using the QueryPerformanceCounter() function about 2000 to 3000 times like in Geekbench 3 and 4, it's used about 14 million times during the same duration. I really don't know what to say other than that's incredibly stupid and obviously completely unnecessary.

So there is a contention bug currently showing only on AMD CPUs in BenchMate 0.9.1 + Geekbench 5. I've already fixed it and will upload an update in a few minutes or so.

[_mat_]

And here is the update: BenchMate 0.9.2

Download | Change Log | Donate

The Geekbench 5 scores are also much better now, more than 1000 points for multicore score on my 9900K@5G. That's because I'm not emulating 14 million QPC calls through the driver with HPET. So QPC uses what it is configured to use, but is regularly checked for skewing. That's the same method that is used in CINEBENCH as well as 3DMark03, 05 and 06.

On a side note: Don't ever try to enable HPET on modern CPUs and run Geekbench 5. QPC will be awefully slow and your score will tell you. You will effectively bench your HPET timer latency.

[Guest Digg_de]

6 hours ago, I.nfraR.ed said:

I've bought previous Geekbench versions, but I'd rather support Mat than buying the same benchmark again for the third time.

I get this email:

Hi, I'm excited to announce that Geekbench 5 is now available for download. Since you recently purchased Geekbench 4 I'm happy to provide you with a complimentary upgrade to Geekbench 5. Here is your Geekbench 5 license key: XXXX-XXXX-XXXX-XXXX-XXXX-XXXX-XXXX-XXXX-XXXX

[Noxinite]

47 minutes ago, Digg_de said:

I get this email:

Hi, I'm excited to announce that Geekbench 5 is now available for download. Since you recently purchased Geekbench 4 I'm happy to provide you with a complimentary upgrade to Geekbench 5. Here is your Geekbench 5 license key: XXXX-XXXX-XXXX-XXXX-XXXX-XXXX-XXXX-XXXX-XXXX

Wat. Why give new customers keys and not old customers??

[ajc9988]

8 minutes ago, Noxinite said:

Wat. Why give new customers keys and not old customers??

Depends on when you purchased it. If you just purchased it, you likely would have gone with the newer version of the software. Whereas older purchasers knew what they bought when they bought it.

[keeph8n]

I got the same email as well......

[Guest Digg_de]

41 minutes ago, ajc9988 said:

Depends on when you purchased it. If you just purchased it, you likely would have gone with the newer version of the software. Whereas older purchasers knew what they bought when they bought it.

Hmm yes.. i've bought the key maybe 6-8 weeks ago. But new key is only 4.99$ at the moment.

[ajc9988]

Just now, keeph8n said:

I got the same email as well......

Hey, have you had any issues with GB5 on Win 10? All I've seen is it either complain of a timer skew or it doesn't pop up a browser and upload the file so I cannot see the results. Don't know if you saw similar before activating your copy...

[ajc9988]

Just now, Digg_de said:

Hmm yes.. i've bought the key maybe 6-8 weeks ago. But new key is only 4.99$ at the moment. 

Yeah, some companies put it at 30 days from release, which means you would fall outside of that with this. Then again, I haven't been happy with GB multithreaded testing since GB3, and nothing I've seen has yet to change my mind there with GB5.

Further, because this is the full x64 version, whereas prior versions required a paid license for x64, which always outperformed the x32 variant, thereby guaranteeing a need to purchase the full version, you don't have to this time. So you may think of skipping it to be honest.

[cbjaust]

11 hours ago, _mat_ said:

I've just encountered a strange synchronization bug on Ryzen with the new Geekbench 5 integration. As I've looked further I realized that instead of using the QueryPerformanceCounter() function about 2000 to 3000 times like in Geekbench 3 and 4, it's used about 14 million times during the same duration. I really don't know what to say other than that's incredibly stupid and obviously completely unnecessary.

So there is a contention bug currently showing only on AMD CPUs in BenchMate 0.9.1 + Geekbench 5. I've already fixed it and will upload an update in a few minutes or so.

Is that something GB does by itself or is it to do with the BenchMate integration?

If it's the former, what's the bet our friends at Intel bought and paid for the GB update?

[_mat_]

GB itself makes the timer calls, the BenchMate integration just counts the calls and checks that they are not skewed.

Intel has nothing to do with that. Maybe they sponsored the deep learning test though, because with the new instructions in Cascade Lake-X they will rock it.

[ajc9988]

4 hours ago, _mat_ said:

GB itself makes the timer calls, the BenchMate integration just counts the calls and checks that they are not skewed.

Intel has nothing to do with that. Maybe they sponsored the deep learning test though, because with the new instructions in Cascade Lake-X they will rock it.

Yeah, the ML seems to be out of place. But the AES test seems to be highly mem sensitive. Just changing my mem interleaving to optimize for it I saw a 2000 point jump. I like the physics tests, but the lightly multithreaded crap, like HTML and their SQL (at least from what it seems), along with speech recognition, horizon line detection, etc., don't really hit the mark on heavy multithreaded tasks, IMO.

It is like a mix between GB3 multithreaded, which was good (but would need updated by this point), and GB4 multithreaded, which sucked. You get a couple tests that stress the CPU like GB3 did, but most do not, like GB4 did. But, to be fair, the distribution on scores does seem more representative of what has been seen on performance by AMD and Intel CPUs, at least as of the first gen Ryzen/TR and the 7960/7980X/E variants and the soldered 9 series which brought nothing to the table, at least to a degree.

I also think the AES test is good as it is similar to what is used in bitlocker by MS, so not so far out there that it is a single case scenario that only a couple users will get a benefit from. It's similar enough that it can show everyday performance unlocking password managers, etc. How heavily it is weighted (see the 2000 point jump) is a different story.

Overall, I'm not as hating on the benchmark as I was yesterday, but still trying to fully unpack whether I like it or not (still in the not category, but doesn't mean I won't bench it).

Edit: I'm also trying to figure out why the old AES tests used over 30GBps while the new one is only using 25GBps on my Ryzen 1950X, while using closer to the 30s for the Intel chips. Little things like that are bugging me and I will have to look through them when I get time to compare the versions more thoroughly.

Edited September 4, 2019 by ajc9988

[_mat_]

To give an update here as well: BenchMate 0.9.3 is out.

The new "feature" is that it removes Geekbench from the Big Bundle, because the dev of Geekbench rather wants to sue me than making his benchmark valid for competitive overclocking.

Full details here:

 

 

[max1024]

I have a question about "Geekbench 5". I bought Geekbench 5, did some tests, I liked this benchmark. I tried to upload the results to hwbot, but it’s bad, on the benchmark page itself there is a button that says “GEEKBENCH5 - MULTI CORE (BETA)”, but in fact only GEEKBENCH 5 - WITH BENCHMATE benchmark. See for yourself.

Please add another GEEKBENCH 5 without BENCHMATE, as was done for previous versions.

[_mat_]

I will remove the categories "with BenchMate" with the release of BenchMate 0.10. Geekbench will no longer be supported due to reasons stated in the post above yours.

What happens to Geekbench 5 here on HWBOT is nothing I really care about. It's not a safe benchmark to compare results competitively, that's for sure.

[Guest Nitro35]

I'd like an explanation.
0.93 does not pass to geekbench5
 and on 8.1 . impossible to save.

where.....

thanks

Edited September 28, 2019 by Nitro35
to keep it greyed out

[_mat_]

Nowhere in this universe.

Please read the posts above, I will not answer questions about Geekbench anymore.

[Guest]

Yup we don't need Geek anymore either.  I won't be buying #5 or any future releases.

Not even interested to bench it. Maybe if I'm bored enough