2060 KO is actually a 2080 Chip?

[TheoneandonlyMrK]

Zotac has like 3 or 4 tu104 2060s and I want to add them but can't match them

Could be useful info for low-income devs and professionals, since 50% more performance is quite a leap, and money saved/earned.

 

[Cheeseball]

There's a thread about this story here, somewhere...

• A 3 year warranty for brand new purchases
• A transferable warranty if you bought a used card that was still under warranty (which it continues off of)
• I've seen many cases throughout the years where they send new (or unused) or upgraded models if repair RMAs can't be fulfilled
• 1 year warranty on B-stock units

More:

• Reference PCBs on most of their cards, just in case one wants to upgrade to a waterblock but couldn't get a Founder's Edition
• An actual upgrade program that just requires the difference between the newer model MSRP and the older model MSRP (not based off any sale prices)
• They pay for shipping to-and-from during the RMA process (this was when I was in the Philippines for an EVGA GeForce GTX 460 SuperClocked 1024MB, not sure about North America)

 

[londiste]

This Actually puts an interesting slant on the 2080's quick ditching for super's, perhaps their yields were not all that at the start leaving a lot of tu104's that didn't pass muster, Nvidia are very skilled at managing waste.

2070 Super, 2080 and 2080 Super are all TU104. 2080 Super has less disabled parts than 2080 as well. Ditching 2080 was not about chip yields but market situation and positioning.

Two things stick out when comparing TU104 and TU106:
1. TU104 has different GPC - 8 SM instead of 12 SM on TU106. Same amount of shaders means 2060 with TU104 has more GPC. The particular block that comes to mind when thinking GPC is Raster Engine.
2. Cache. TU104 has more cache. How much of it is left enabled on 2060 with TU104 is a good question. Is anyone able to check?

 

[TheoneandonlyMrK]

2070 Super, 2080 and 2080 Super are all TU104. 2080 Super has less disabled parts than 2080 as well. Ditching 2080 was not about chip yields but market situation and positioning.

Two things stick out when comparing TU104 and TU106:
1. TU104 has different GPC - 8 SM instead of 12 SM on TU106. Same amount of shaders means 2060 with TU104 has more GPC. The particular block that comes to mind when thinking GPC is Raster Engine.
2. Cache. TU104 has more cache. How much of it is left enabled on 2060 with TU104 is a good question. Is anyone able to check?

Is the Double precision performance equal between tu104 and others too?.

 

[londiste]

Is the Double precision performance equal between tu104 and others too?.

Yes, both have 1:32 for FP64.

 

[TheoneandonlyMrK]

Yes, both have 1:32 for FP64.

Fair enough, just seams strange that the performance only of professional applications is upto 50% better , especially just from more cache, without that also enhancing gaming performance.

 

[ppn]

4-6 Gpc instead of just 3. In Every GPC there is a raster engine. Apparently the raster engine is the bottleneck.

 

[Keullo-e]

Nothing new about this. GTX 1060 GDDR5X used GP104 GPU instead of the typical GP106.

 

[Darmok N Jalad]

The original 2070 was a TU106, but the 2070 Super went to TU104. It’s all just marketing names. In my mind, when they mess with the specs within a model, it’s a bigger issue. The RX 560 was a good example, where shaders and clocks were all over the place.

 

[Vayra86]

not every 104 is a 2080 chip

but....

tu104 has fewer cuda per sm,8 vs 12

therefore more sms are needed for same cuda count

2060 KO should have more sm than 2060 FE,therefore more RT cores

This makes me curious... is the 2060KO another one of those asymmetric VRAM bus cards again? *starts DDG*

 

[Keullo-e]

This makes me curious... is the 2060KO another one of those asymmetric VRAM bus cards again? *starts DDG*

6 memory chips = 6x 1GB.

 

[Vayra86]

6 memory chips = 6x 1GB.

View attachment 143081

What we need is how these are wired to the SMs

 

[Keullo-e]

What we need is how these are wired to the SMs

Ah, good point. I guess that pops out sooner or later.

 

[londiste]

This makes me curious... is the 2060KO another one of those asymmetric VRAM bus cards again? *starts DDG*

By asymmetric, do you mean something like GTX970? Nope. The problem with GTX970 was that it had 8 memory chips/channels populated but only 7 connections on the backend. I do not remember exactly whether that backend was memory controllers or the connection to the GPU parts beyond that (parts of that memory controller block are ROPs and L2 cache).

Both TU104 and TU106 have more than enough backend resources to work with 6 memory chips and both have some of the controllers disabled in the first place.

This is also the reason I would suspect L2 cache is not what is behind the Blender performance boost of KO - it should be tied to the memory controllers and ROPs that are enabled and with 6 memory chips, it is the same 6 of those for RTX2060 with both TU104 and TU106.

 

[Vayra86]

By asymmetric, do you mean something like GTX970? Nope. The problem with GTX970 was that it had 8 memory chips/channels populated but only 7 connections on the backend. I do not remember exactly whether that backend was memory controllers or the connection to the GPU parts beyond that (parts of that memory controller block are ROPs and L2 cache).

Both TU104 and TU106 have more than enough backend resources to work with 6 memory chips and both have some of the controllers disabled in the first place.

This is also the reason I would suspect L2 cache is not what is behind the Blender performance boost of KO - it should be tied to the memory controllers and ROPs that are enabled and with 6 memory chips, it is the same 6 of those for RTX2060 with both TU104 and TU106.

Yeah its pretty elegant how these two SKUs match their SM / shader counts for the 2060KO (3x8 / 2x12), too. And thanks, I could also not find anything like it as I believe they changed the way that backend works as well. Other than 970 I was thinking more of something like the 660 with its 0.5 GB allocated to specific SMs and 1.5 to others. But it no longer seems to work that way.

 

[newtekie1]

I know that this happens but is this not dishonest?

I never understand this logic at all. How is this dishonest at all? You aren't paying for the core, you are paying for the performance and specs, which aren't changing. They could use a 2080Ti core in it if they wanted to as long as they cut it down enough to match RTX2060 specs and it performs like a RTX2060.

 

[kapone32]

I never understand this logic at all. How is this dishonest at all? You aren't paying for the core, you are paying for the performance and specs, which aren't changing. They could use a 2080Ti core in it if they wanted to as long as they cut it down enough to match RTX2060 specs and it performs like a RTX2060.

The reason I say that there are 2 different chips for the exact same card in name. Was this not done before with the 1060 3GB and 6GB? There is factual data that the card with the 2080 chip is much faster in productivity than the "regular`2060. (Try to think of yourself as a noob) How would you feel if you bought a 2060 months ago and your friend buys one now and both of you are doing video production. Would you not be miffed if he tells you he is finished his run and you still have 20 minutes left for the exact same project?

 

[TheoneandonlyMrK]

I never understand this logic at all. How is this dishonest at all? You aren't paying for the core, you are paying for the performance and specs, which aren't changing. They could use a 2080Ti core in it if they wanted to as long as they cut it down enough to match RTX2060 specs and it performs like a RTX2060.

While I do agree, you are buying the spec of hardware.
And I personally don't think it dishonest just borderline so.
It's worth noting these are performing upto 50% better in professional apps, bench proven by Gamers Nexus, so they aren't matched specs.

And this IS exactly the kind of information I and my like would want when buying a card, yes so i can dodge and avoid being ripped off but also so that i can buy via an informed decision.

And some would gain quite a fair bit of performance in the apps they use by just choosing wisely when buying a 2060, that's clear to see.

 

[londiste]

I get a feeling that this result was a surprise even to Nvidia.
Cutting the GPU resources here appear to be perfectly matched at the first glance. The number/size of all major resources is identical.

 

[newtekie1]

Was this not done before with the 1060 3GB and 6GB?

It's been done countless times.

But the 1060 3GB and 6GB used the same chip, but different varients. Just like the RX 5700 and RX 5700 XT.

There is factual data that the card with the 2080 chip is much faster in productivity than the "regular`2060. (Try to think of yourself as a noob)

Under very specific workloads, that the card was not intended for, sure. But people really interested in those workloads weren't likely buying RTX2060's anyway. Even GN mentions that really these new cards will likely get gobbled up now that it is known they have this massive performance advantage at this price point. But before that, people weren't seriously buying RTX2060's for workstation work.

How would you feel if you bought a 2060 months ago and your friend buys one now and both of you are doing video production. Would you not be miffed if he tells you he is finished his run and you still have 20 minutes left for the exact same project?

No, because I understand how the world works. Time moves on, things improve. I don't get pissed because someone bought something newer and got something better than what I bought in the past.

I bought a 2019 Camry, I'm not pissed that people can now buy a 2020 Camry with better features for the same price or even cheaper. That's just how the world works.

While I do agree, you are buying the spec of hardware.
And I personally don't think it dishonest just borderline so.
It's worth noting these are performing upto 50% better in professional apps, bench proven by Gamers Nexus, so they aren't matched specs.

And this IS exactly the kind of information I and my like would want when buying a card, yes so i can dodge and avoid being ripped off but also so that i can buy via an informed decision.

And some would gain quite a fair bit of performance in the apps they use by just choosing wisely when buying a 2060, that's clear to see.

So what you're saying is, assuming an uninformed consumer that just thinks they are buying a RTX2060, they are at best getting the performance they paid for an at worst are getting more than what they paid for and somehow this is "borderline dishonest" and a bad thing? Sorry, I don't see that logic.

 

[TheoneandonlyMrK]

It's been done countless times.

But the 1060 3GB and 6GB used the same chip, but different varients. Just like the RX 5700 and RX 5700 XT.



Under very specific workloads, that the card was not intended for, sure. But people really interested in those workloads weren't likely buying RTX2060's anyway. Even GN mentions that really these new cards will likely get gobbled up now that it is known they have this massive performance advantage at this price point. But before that, people weren't seriously buying RTX2060's for workstation work.



No, because I understand how the world works. Time moves on, things improve. I don't get pissed because someone bought something newer and got something better than what I bought in the past.

I bought a 2019 Camry, I'm not pissed that people can now buy a 2020 Camry with better features for the same price or even cheaper. That's just how the world works.




So what you're saying is, assuming an uninformed consumer that just thinks they are buying a RTX2060, they are at best getting the performance they paid for an at worst are getting more than what they paid for and somehow this is "borderline dishonest" and a bad thing? Sorry, I don't see that logic.

Borderline because if I were that hypothetical customer and I used it for modeling ,cad etc then this came out i would be pissd, it's not dishonest but I would feel disgruntled, anyway as i said if people were informed ,all good.
We both know the package says nothing about what chips on it soo.

 

[Darmok N Jalad]

But we are talking about the KO. That is enough distinction, as sure as something like Nitro, Nitro+, XT, or OC all mean a card is running at a different spec than the original model name that proceeds such suffixes. If you bought an RTX 2060, and your friend buys an RTX 2060 KO, then you didn't have the same card to begin with to make such a comparison. There is no "spec creep" going on, where later models are faster--this is a different model. Perhaps its a rehash or it could be better named, but regardless, it's a different model than the original 2060.

 

[kapone32]

But we are talking about the KO. That is enough distinction, as sure as something like Nitro, Nitro+, XT, or OC all mean a card is running at a different spec than the original model name that proceeds such suffixes. If you bought an RTX 2060, and your friend buys an RTX 2060 KO, then you didn't have the same card to begin with to make such a comparison. There is no "spec creep" going on, where later models are faster--this is a different model. Perhaps its a rehash or it could be better named, but regardless, it's a different model than the original 2060.

I get what you are saying. All of the cards you mentioned have the same chip and for me the main difference is in the cooler. There is also the level of factory OC that I see in different models of the same GPU, not the chip that resides on the PCB.

 

[TheoneandonlyMrK]

Pp

But we are talking about the KO. That is enough distinction, as sure as something like Nitro, Nitro+, XT, or OC all mean a card is running at a different spec than the original model name that proceeds such suffixes. If you bought an RTX 2060, and your friend buys an RTX 2060 KO, then you didn't have the same card to begin with to make such a comparison. There is no "spec creep" going on, where later models are faster--this is a different model. Perhaps its a rehash or it could be better named, but regardless, it's a different model than the original 2060.

See Tc fantasy's post, it's going to be some 2060's too ,he would know.

 

[newtekie1]

Borderline because if I were that hypothetical customer and I used it for modeling ,cad etc then this came out i would be pissd, it's not dishonest but I would feel disgruntled, anyway as i said if people were informed ,all good.
We both know the package says nothing about what chips on it soo.

So you're saying that they should always announce that there might be a new better version of what you are buying coming out at some point in the future whenever you buy something. Yeah, the world doesn't work that way and not doing so is in no way dishonest, not even borderline.