How to fix the rankings :)

[Rauf]

Ok, the new titan is out and it upsets the rankings on hwbot like nothing else. Aircooled WR's and GFP's. And then a few months later when 1080 Ti comes everything is back to normal. Why don't we make things better to remove this "downtime" which occurs every time a new titan is released?

 

Imagine if almost every 2D-benchmark was multithreaded and were not divided into categories by the amount of cores. 6950X would rule every benchmark, wouldn't you react to that?

 

Couldn't we do the same kind of categories on GPU's as well? Split up the rankings based on the number of cores/shaders/streamprocessors? Now of course we can't divide by exact number of cores because there would be too many categories. But we could make intervals 0-500, 500-1000, 1000-1500, 1500-2000, 2000-2500, 2500-3000 etc. Maybe the intervals need to be tweaked and maybe they need to be changed over time but wouldn't it be a good way to lessen the impact of the titans and at the same time make 3D-benching more popular by making it possible to compete with cheaper GPUs?

[TerraRaptor]

Streamprocessors are different from gen to gen.

[Rauf]

Streamprocessors are different from gen to gen.

Thereby the interval... or do you mean the instruction per cycle, or whatever it is called for GPUs. We can't account for that. If some generations are inefficient, too bad. Same situation as for CPUs, too bad AMD...

[TaPaKaH]

We could also go by the number of TMUs or ROPs as they don't grow as much in quantity as shaders.

This would actually be a very interesting idea since it would, say, allow people with two mid-end cards compete in the same category with one high-end card in the same way as you can compete with dual quadcores against one octacore.

Edited August 16, 2016 by TaPaKaH

[Alan_Alberino]

Maybe make something like OC-Esports divisions? Divide hardware in divisions and update them each time new GPU series release

[der8auer]

Interesting idea. Would be interesting to see how this would look

[mickulty]

I really like the sound of ROP-based rankings as Sam suggested - per-core rankings obviously work well on the CPU side and ROP count seems to follow a comparable pattern of per-ROP performance increases and a slowish increase in ROP count over time. To avoid an explosion in the number of rankings there would have to be groups - maybe 1, 2, 3-4, 5-8, 9-16, 17-32 etc? Of course one or two cards would still be slightly disadvantaged but never any more than they are by the current system.

[Rauf]

It definitely requires some thought on how to do the split. ROP, cores/shaders or something else or some kind of ratio or product between rop and shaders. If we do ROP we see that titan pascal has the same amount as 980 Ti, which makes it less than ideal. We also see that 1070 and 1080 would be placed in same category. On the other hand, for AMD they have a large amount of shaders but low ROP on the Furys, which would be unfair towards them if we do just shaders. Maybe ROP x Shaders gives the best way to categorize GPUs?

[TaPaKaH]

Doesn't Titan XP have 96 ROPs versus 64 for GTX1080? And there's nothing wrong in placing 1070 and 1080 in the same category, we've been doing this with 2500K/3570K/46x0K/6600K for a while now. With some smart per-ROP partitioning (.., 30-50, 50-80, ..) the RX480 and GTX1060 will end up in the same category making AMD relevant again.

Edited August 18, 2016 by TaPaKaH

[K404]

Ok, the new titan is out and it upsets the rankings on hwbot like nothing else. Aircooled WR's and GFP's. And then a few months later when 1080 Ti comes everything is back to normal. Why don't we make things better to remove this "downtime" which occurs every time a new titan is released?

 

Imagine if almost every 2D-benchmark was multithreaded and were not divided into categories by the amount of cores. 6950X would rule every benchmark, wouldn't you react to that?

 

Couldn't we do the same kind of categories on GPU's as well? Split up the rankings based on the number of cores/shaders/streamprocessors? Now of course we can't divide by exact number of cores because there would be too many categories. But we could make intervals 0-500, 500-1000, 1000-1500, 1500-2000, 2000-2500, 2500-3000 etc. Maybe the intervals need to be tweaked and maybe they need to be changed over time but wouldn't it be a good way to lessen the impact of the titans and at the same time make 3D-benching more popular by making it possible to compete with cheaper GPUs?

 

If there is a problem that comes down to a hardware category or class...... exclude it from the rankings?

 

What is it that people REALLY object to? The fact that it doesn't OC better under LN2?

[Rauf]

If there is a problem that comes down to a hardware category or class...... exclude it from the rankings?

 

What is it that people REALLY object to? The fact that it doesn't OC better under LN2?

It's the prize of course. Same with 6950X, most would surely be glad if we excluded it. But excluding hw doesn't seem right. CPU categories are working well, why not do it for the GPUs as well?

[K404]

So... instead of number of GPUs being the key to global GPU-scores, it would be sub-divided by some ROP/SP count?

 

It brings more cards into play, but I don't think it solves the problem.

 

Best scores in a class/sub-division might be two GPUs under LN2. That's a hard sell.

[Rauf]

So... instead of number of GPUs being the key to global GPU-scores, it would be sub-divided by some ROP/SP count?

 

It brings more cards into play, but I don't think it solves the problem.

 

Best scores in a class/sub-division might be two GPUs under LN2. That's a hard sell.

Yes, I looked at some figures. Multi-GPU setups will beat single GPUs if we categorize by ROP. Haven't looked at shaders but I doubt multi GPUs will beat any singles there.

 

But we can still have SLI/CF like it is today, it can be linked to the number of ROPs/shaders a single card has, and then we have 2way, 3way etc for each rop/shader class.

 

The goal is to lessen the impact of ultra high end hw, and to bring cheaper GPUs into play.

[K404]

Yes, I looked at some figures. Multi-GPU setups will beat single GPUs if we categorize by ROP. Haven't looked at shaders but I doubt multi GPUs will beat any singles there.

 

But we can still have SLI/CF like it is today, it can be linked to the number of ROPs/shaders a single card has, and then we have 2way, 3way etc for each rop/shader class.

 

The goal is to lessen the impact of ultra high end hw, and to bring cheaper GPUs into play.

 

 

IMO, that would be too many rankings/leaderboards and too many sets of "global" points

 

I can get onboard with bringing more GPUs into relevance, no doubt with the extra necessity of e-Power etc for some/most.

[Massman]

It's not impossible, but this will create a LOT of new global rankings. I just did a quick check with 3DMark03 and without verifying if all the graphics cards have their SP filled out correctly, there would be ~80 global rankings (instead of the 4 we have now). We'd also have to find a solution to group the older architectures that don't have the ROP/SP layout.

[Alan_Alberino]

We'd also have to find a solution to group the older architectures that don't have the ROP/SP layout.

 

Old GPUs don't give globals now, only last generations mid-high or high end cards give GP, so I think it's not necessary to include them...

 

Maybe 10 categories divided to include from 4-way high end till X60 cards, cards slower than gtx x60 or AMD equivalent don't give globals now so I don't see a reason to include them in new system... Maybe 9 categories + a big categorie with ALL lower cards than actual gen GTX x60 cards to include all the others, but not more than that...

[elmor]

I'd like to address the complaints about the Titan XP. IMO the situation is the most open and fair it's been in a long time. Buy a card, work out the mods and bench it on ln2 = #1 in any 3D bench that's not CPU bound. I'm interested in why this is not happening.

 

Cost is too high? It's around the same as Extreme Edition CPUs.

Can't work out the mods? Ask for help here at the forums.

No samples? Maybe some will once again fear killing hardware.

[TaPaKaH]

Why not have, say, 5 categories?

• 0-31 ROPs: GTX750Ti (3363 on LN2), GTX760-192bit (2695 on air), R7-360 (4387 on air), RX460.
  
• 32-63 ROPs: GTX780Ti (8950 on LN2), GTX1060 (7151 on stock), RX480 (7698 on LN2).
  
• 64-95 ROPs: GTX980 (10459 on LN2), GTX1080 (13161 on LN2), FuryX (10556 on LN2), 2xRX480 (11863 on air).
  
• 96-128 ROPs: 2xGTX780Ti (15011 on LN2), 2xGTX980 (17198 on LN2), GTX980Ti (14432 on LN2), TitanXP (16604 on LN2), 2xFuryX (18184 on LN2)
  
• 129+ ROPs: 3/4-way stuff

I've put the top FSE score in brackets behind each card. Except for GTX780Ti monstering the 32-63 ROP rank (then again, how many of these cards were binned?), everything seems rather logical/comparable, doesn't it?

[Massman]

Why is the 4th category not 96-127? For future-proofness should we also prepare 128-159, 160-191 and 192+ ?

 

Looks good, but I'd like to hear a more technical point of view on the segmentation per ROP count.

[K404]

Why not have, say, 5 categories?

• 0-31 ROPs: GTX750Ti (3363 on LN2), GTX760-192bit (2695 on air), R7-360 (4387 on air), RX460.
  
• 32-63 ROPs: GTX780Ti (8950 on LN2), GTX1060 (7151 on stock), RX480 (7698 on LN2).
  
• 64-95 ROPs: GTX980 (10459 on LN2), GTX1080 (13161 on LN2), FuryX (10556 on LN2), 2xRX480 (11863 on air).
  
• 96-128 ROPs: 2xGTX780Ti (15011 on LN2), 2xGTX980 (17198 on LN2), GTX980Ti (14432 on LN2), TitanXP (16604 on LN2), 2xFuryX (18184 on LN2)
  
• 129+ ROPs: 3/4-way stuff

I've put the top FSE score in brackets behind each card. Except for GTX780Ti monstering the 32-63 ROP rank (then again, how many of these cards were binned?), everything seems rather logical/comparable, doesn't it?

 

This kind of grouping only works in hindsight or until there is a change in GPU architecture.

 

Global is no longer global.

 

Why was nothing reallllly said during the reign of previous Titans? Why now, after a few weeks and results from a few benchers on air/water?

[elmor]

Why was nothing reallllly said during the reign of previous Titans? Why now, after a few weeks and results from a few benchers on air/water?

 

See my previous post, there are no cards from any vendors this time. Which means there's no point to market them and send out samples. They're only available from Nvidia.

[K404]

I don't see what difference that makes All previous Titan's were ref with AIB stickers and very similar/ the same MHz

[Casanova]

It's not impossible, but this will create a LOT of new global rankings. I just did a quick check with 3DMark03 and without verifying if all the graphics cards have their SP filled out correctly, there would be ~80 global rankings (instead of the 4 we have now). We'd also have to find a solution to group the older architectures that don't have the ROP/SP layout.

 

I already made a suggestion about this situation in another thread.

Solution: create a delay of about two months to give league points to new hardware.

[Strong Island]

Man this idea would be incredible.

 

Imagine being able to get globals with a 1070, 1060.

 

I was also thinking the other day, AMD has behind now for so long isnt it time we maybe split amd and nvidia globals. Or find a way to make amd cards relevant in globals again.

 

We should never adjust points based on a companies performance but this has been going on now for years.

 

AMD has some really nice priced gpus, getting them back involved with globals would help so much. It almost doesnt make sense for them to compete for the same global points because they are so far behind.

 

It would be so cool if we were all benching rx480's or something similar.

 

I wish I could afford a titan x but I just cant. I would love to bench one.

[K404]

 

Imagine being able to get globals with a 1070, 1060.

 

 

 

With the right CPU and a tactical choice of benchmark......