HWBOT 2020 Make it great again Thread:

[saltycroissant]

I don't think removing x3/x4 3D globals because its old would be fair. We still have 1/5 cores globals even tho its a thing of the past. I feel like you should be rewarded for dealing with the cancer that are older AMD driver/GPU.

And if you look at TSE/Vantage xtrm/superposition those don't really have a ton of globals points anyway.

[Leeghoofd]

Programmer called me today and wants to push the new point system to the HWBOT page as soon as possible. He will first add a lower threshold for hardware points, so hardware with less than 7 subs gets a max of 6 points iso the 15 now on UAT.

I guess we can only analyse everything once all recalculations are done and users can properly compare profiles, rankings,...

Note: Issue is that some current user profiles on HWBOT are not correct, credits to the current bugged point algorithm...

Plan is to push all the lastest scores to the UAT server so you can still compare scores if you are benching
Hoping that all will be up and running within the time margin of just a few days.

Keep you posted when it will happen!

[yee245]

Regarding globals for 3x/4x cards setups, I doubt it would be trivial to implement, but if some of the options could just be disabled, or have points disabled only for certain benchmarks that just don't actually support that many GPUs, maybe it would solve some of the issues of getting "free" global points. For example, I think it's Port Royal (and Night Raid) that just won't actually use more than 2 GPUs, even if you have a functional 3/4-way setup. Similarly, I think the Wildlife benchmarks state in their documentation that it doesn't support SLI, so any submissions with 2/3/4-way SLI are actually still effectively just a single card benchmark. WL isn't so much an issue regarding points, since it currently gives none, but I am currently benefitting from the fact that 3/4-way SLI will technically run Port Royal with older GPUs and give a score, even if it doesn't scale past 2 cards anyway, but PR does give globals for those 3/4-way configs.

Other 3D benchmarks that currently give globals do in fact support 4-way configs, like FSE, TSE, Vantage, etc., and they're filterable options in 3DMark's Hall of Fame pages.

Then, you run into other potential "issues" with removing points/globals from the multi GPU setups. For example, GPUPI v3.3 32B gives globals, and there are subs for 1/2/3/4/5/6/7/8x GPUs, which just benefits from having multiple GPUs, but not necessarily requiring functional SLI or anything to run, like the actual graphical-related benches. So, do you award globals for only 1 or 2 GPU setups, but not more, or do you just remove globals entirely from that benchmark? And same with the other 3D benchmarks above that do support 3/4 GPU configs.

[Leeghoofd]

Interesting points, however we don't remove any globals for multi card setups, till we can assess the impact of the new points correctly.

Also think that you will require a certain amount of subs to get good globals, so take eg Port Royal/Timespy/Vantage Extreme with 3 and 4 cards on UAT , it's only 25 global points. Principle remains the same, the more people bench and submit, the higher the points will get...

[yee245]

I guess I was just mainly getting at the fact that some benchmarks don't/can't actually use the extra GPU cores, even if they're there. While maybe not quite the same situation (since they're already categorized as single-threaded anyway), but it would kind of be like if something like Superpi or Pifast gave points for each different core or thread or processor count, even though it's single-threaded. You don't get to sub more categories and get extra points for Superpi just because you're using a dual- or quad-Xeon setup. As I understand it, some of those benchmarks, like Port Royal, Night Raid, etc. will only utilize 2 GPUs, even if you have more, just by the nature of the benchmark. But, instead of it being only a single ranking system for Superpi/Pifast, some of those 3D benchmarks are really only "single-threaded" or "2-threaded" GPU benchmarks. Any additional GPU cores will just sit idle. That said, I don't know how something like Different SLI may or may not help with making more GPU cores work together.

I just don't know the difficulty of selectively limiting certain of the benchmarks to only give points for certain GPU counts. Or, if it should just be left to people to decide time/effort benchmarking things that may be inefficient but still give you a score and boints (like if you were to benchmark some of those high-core-count multi-Xeons/Epyc setups without disabling cores/threads... for Time Spy or something... getting really inefficient scores).

[unityofsaints]

Am I correct in thinking that > 50 subs = 70 pts, up to 50 subs = 50 pts, lower = 35 pts for hwpts? I am warming up to this system after thinking about it a bit... but removing globals for multiway 3D is a mistake in my opinion, I see it like 3/5-core 2D globals: niche, but tons of fun.

[chispy]

Please do not remove globals for multi-gpu , it is not as easy as it seems , difficulty level x4 gpus goes up exponentially , like my amigo @saltycroissant said , try to get amd drivers to work on legacy quad crossfire is a battle in itself ? and we should be rewarded for all the trouble and dificulties just to make it run ( ahheemm... amd 5970 x2 quadfire pita ). But that's where the points are and also the fun to tweak it to get good efficiency. So , please do not remove global for multi gpu. My 2 ? cents ...

 

The sooner we get this new point system in place the better , thank you @Leeghoofd , let's go go go !

Edited January 27, 2022 by chispy

[yosarianilives]

If I can get globals by running stupid server cpus that have 15 cores each in multi socket I don't see any reason not to give multi gpu globals. The only exception I can think of is geek 4 compute as it doesn't use sli/xfire and it does not support multi gpu at all

[Leeghoofd]

To configure benchmarks to be multithreaded or not is easy, but limited in options as you can see in the below image

 

[der8auer]

We will address the multi GPU "issue" at a certain point but right now we first have to adjust the point system itself to make the website faster and also easier to understand. Plus there should be higher rewards in general like back in the days. 

I think we can all agree on the fact that SLI/NVLink is pretty much dead in general and we also don't want to emphasize categories which are mainly ruled by "more money". We are already in a business where money buys points to a certain degree if you can bin more stuff. So if we can make benching cheaper or at least less expensive it should be better.

But as I already said this has to be adjusted/decided later. Feel free to post ideas on this.

 

 

[Leeghoofd]

Progress:

Globals done

AM4 done
TR4 done
SWRX8 done
STRX4 done
SP4R2 done
SP4 done
SP3 done

LGA4189 done
LGA3647 done
LGA2066 done
LGA2011v3 done
LGA2011v1 done
LGA2011 almost completed
LGA1700 done
LGA1200 done
LGA1151v2 done
LGA1151 up next

Geforce 3000 series done
Geforce 2000 series done
AMD RDNA2 done

 

 

 

 

[saltycroissant]

6 hours ago, der8auer said:

We will address the multi GPU "issue" at a certain point but right now we first have to adjust the point system itself to make the website faster and also easier to understand. Plus there should be higher rewards in general like back in the days. 

I think we can all agree on the fact that SLI/NVLink is pretty much dead in general and we also don't want to emphasize categories which are mainly ruled by "more money". We are already in a business where money buys points to a certain degree if you can bin more stuff. So if we can make benching cheaper or at least less expensive it should be better.

But as I already said this has to be adjusted/decided later. Feel free to post ideas on this.

 

 

Because its less popular doesn't mean we should remove it imo. Multi GPU isn't really something you have success at with $$$. You have 01SE/03/vantage xtrm/3D11 that don't require new GPU, you can get in the top 5 for those with half the cost of a RTX 3080. 1/3/5 cores is far more dead then multi GPU, even Z690 has support for SLI/Crossfire.

Ppl complain mainly because its hard to maintain your globals on 3D because with each new release someone can just bench with the latest CPU and possibly beat it for legacy benchmark. Unless we change 3D to only GPU score you will never make those ppl happy. If they have to rerun their GPU every once in a while so be it.

3D multi gpu globals issue isn't $$$, its ppl giving up on getting the drivers to behave.

(Most ppl still want to keep 3D the way they are, wouldn't be fair to change them for 1 or 2 person( some sponsored btw)).

[Rauf]

I think the new points look good overall!

The one imbalanced thing is hw points for more than 100 subs. Place 50 gets 46,x points. Improving to place 10 gets 50,x points. That's not much of a reward in points compared to the leap in ranking. Too easy points in my opinion, there should be a steeper drop from place 4 to 10 or something in the +100 subs rankings.

[Matsglobetrotter]

I would not forget the SLI/crossfire as indeed its easy and does not cost a fortune to get some older cards to have fun with. Just got some nice GTX 980Tis and hd 4870s just to play around with. considering its then possible to use such on the newer platforms its fun to see the scores that can be done.. i mean who would otherwise all the big guns keep playing with 3d3 on various older cards on new CPU's. single or multi gpus is then fun.

[moi_kot_lybit_moloko]

i hate this new point system, and now i have motivation for DOWNclock hardware and make stupid subs. 30-40 hwboitns for downclocked cpus, lol, what?

[alexmaj467]

What is the reason for the minimum score for the last place = 11.6 HW   91 result in the category.

CPU Frequency 1xCPU Ranking with a Celeron D 320 (2.40GHz, s478)

What is the reason for the minimum score for the last place = 40 HW   97 result in the category.

CPU Frequency 1xCPU Ranking with a Pentium III EB 800Mhz (s370)

 

I don't have enough logical assumptions.

 

Links are breaking. Replace  %23 characters at the end before interval=20.   with one #

Edited January 31, 2022 by alexmaj467

[TaPaKaH]

The amount of 70-point hardware categories is too damn high. You can literally have a dozen overclockers fill up their profiles with sixty 70pt scores without ever colliding on the same hardware. This is not beneficial for the quality of scores as you will never have enough active overclockers to fill up these categories and make them actually difficult. We spent 15 years working on former 50/67 point categories and still, some of those are actually easy to get #1 in. You might as well remove the requirement of 60 hardware scores towards global ranks as it no longer "proves the ability of a bencher to also excel on old competitive hardware" but rather "proves that someone wasted time and money on benching 6-8 semi-obscure categories without much quality requirement".

This, and weird examples above where two shitty scores at stock clocks yielding more points than, say, #1 32M on E8600.

[Mr.Scott]

"it's too easy........it's too hard.........."

Honest to god you guys are frustrating MF'ers.

If I was Alby I'd just walk away now.

[eviledeath]

3 hours ago, TaPaKaH said:

The amount of 70-point hardware categories is too damn high. You can literally have a dozen overclockers fill up their profiles with sixty 70pt scores without ever colliding on the same hardware. This is not beneficial for the quality of scores as you will never have enough active overclockers to fill up these categories and make them actually difficult. We spent 15 years working on former 50/67 point categories and still, some of those are actually easy to get #1 in. You might as well remove the requirement of 60 hardware scores towards global ranks as it no longer "proves the ability of a bencher to also excel on old competitive hardware" but rather "proves that someone wasted time and money on benching 6-8 semi-obscure categories without much quality requirement".

This, and weird examples above where two shitty scores at stock clocks yielding more points than, say, #1 32M on E8600.

It's true you're right I didn't think if these simple to manage to do 70pts quite easily we will quickly fill the scoreboard and reach the max
I hadn't thought about it from that angle.

[Leeghoofd]

It's not about easy, this ranking is all about popularity of the hardware...

You still need to be nr1 in these semi obscure categories to get the 70 pointer and it needs to pass the sub threshold, otherwise you end up with 50, 35 or 15 points. Last but not least, this system is giving something back to the biggest crowd at HWBOT. These ambient guys can be more competitive as the gap closes with the extreme guys. 

I always wonder what makes one merit more points in benchmark A than another OCer in benchmark B ? Is it the binning of 100 cpus? Does that make you a better overclocker, meaning you must get more points. One might find pi32m challenging, another person thinks 01 is way more fun... 

What is entertaining is that this has been on the drawing board for months, no inputs whatsoever.  Now that you see it in action nothing is good... Flashbacks...  Negativity... 

I'll check the 2 examples of Alex as this is not right, seems the threshold has not been reached, yet it gives the points. 

Raufs idea is also good, will experiment with that one and stop recalcs. 

This is just a global/HW point refresh, career / SEASONAL nor team points have been touched. Not even sure if they will be touched on the current code. 

This ranking is pretty self explanatory, rank x out of a certain amount of subs get rewarded by x points... There's no complexity behind it. No need if it's harder to run 3 VGAs on Ln2 or other stuff that was tried in the past. Algorithms that made the BOT crawl and spew out incorrect rankings... 

Provide me a fix by taking the excels as example, instead of just saying I hate this point system. 

[Papusan]

7 hours ago, moi_kot_lybit_moloko said:

i hate this new point system, and now i have motivation for DOWNclock hardware and make stupid subs. 30-40 hwboitns for downclocked cpus, lol, what?

Why would you do that? From what i can see, you'll still get more points the better you do it. Not less. With your reasoning... Downclock max and try reach nr1 on the whole bot leaderboard?

Edited January 31, 2022 by Papusan

[moi_kot_lybit_moloko]

3 minutes ago, Papusan said:

Why would you do that? From what i can see, you'll still get more points the better you do it. Not less.

because this is not normal. last place, stock sub and 40 hwboints?

[Leeghoofd]

You still fail to see how and why this is implemented, read up first... or bench stock...

[alexmaj467]

1 hour ago, Leeghoofd said:

I'll check the 2 examples of Alex as this is not right, seems the threshold has not been reached, yet it gives the points. 

It's just not clear why, with the same number of results in the category, the minimum score is different by a large margin.

[Leeghoofd]

Because we use these amount of submissions thresholds Alex.
 

Currently it was set to 100 or more check out the new excel I made, we will raise the threshold to 125 or more for the 70 pointers.

 

UPDATEDHWBOT.xlsx