Some RTX 2060's have up to 47% better productivity performance in apps like Blender and Solid Edge

[Camofelix]

Gamers Nexus is reporting that they've been able to show that some versions of the RTX 2060 performs up to RTX 2080 performance in productivity applications. 

 

This is due to Nvidia selling their TU 104-410  gpu die (previously seen in the RTX 2080) that are speculated to not have met QC requirements. This new version of the die, codenamed TU 104-150, seem's to have many of its FPU's disabled, leading to floating point heavy applications such as gaming seeing similar performance to the TU 106 based RTX 2060's. What GN is speculating and hoping to gain confirmation on is that many of the CUDA core's on the cards did in fact pass validation, meaning that applications that can utilizes these cores being vastly accelerated. 

 

Initial performance analysis shows that the 2 difference versions of the RTX 2060 have similar if not identical gaming performance. The TU 104 based card see's all of it's productivity benchmarks either within margin of error at the worst, or up to 47% improved in certain edge cases compared to its TU 106 based counterpart. 

 

EVGA does not seem to have been aware that the different dies would perform differently. After reaching out to NVIDIA, GN was able to confirm that the performance uplift they measured is with NVIDIA observed before launch. 

 

The 2 dies are reported to cost nearly the same to add in board partners. 

 

GN speculated that they expect the TU 106 based version to perform better than the TU 104 based version when overclocked due to additional power headroom available to the former die when compared to the later. 

 

GN has yet to publish an article detailing their findings, but did publish a video on their YouTube channel detailing what they've discovered. 

 

The Video is embedded below. Youtube embedded privacy mode is enabled for the video. If you do not wish to have your data tracked add this extension to the YouTube URL to access the video directly: /watch?v=mUFRBnJdx3Y

https://www.youtube-nocookie.com/embed/mUFRBnJdx3Y

 

 

 

Edited January 23, 2020 by Camofelix
Forgot to embed video using forum embed tags

[Stu_Bear]

I saw that video this morning...pretty interesting...wish it had 2080 gaming performance..

[Camofelix]

Just now, Stu_Bear said:

I saw that video this morning...pretty interesting...wish it had 2080 gaming performance..

I'm personally really exited for it as a potential AI research card for myself. 

 

I'm a Computer engineering student and would have had to pay almost 50% more for this level of performance a week ago. I'm in Canada so that plays into it, but this could end up being a great budget card for people looking to maximize value for productivity. 

[leadeater]

Solid Edge? I think you mean SolidWork, nope Siemens just being annoying.

[leadeater]

52 minutes ago, Camofelix said:

EVGA does not seem to have been aware that the different dies would perform differently. After reaching out to NVIDIA, GN was able to confirm that the performance uplift they measured is with NVIDIA observed before launch. 

Most of Nvidia were also unaware, they confirmed it after GN contacted them. What made it so odd is that on the basic spec level there is no difference between the dies used, like for example the TU104-150 implemented on the EVGA RTX 2060 KO doesn't have a wider memory bus or something like that.

 

Edit:

The extra CUDA cores are indeed locked and not being used either, applications can't just access them because they are present in the die when they have been either microcode locked out or laser disabled. That would be a really simple explanation. Must be some deeper difference like caches being larger or the arrangement of the CUDA cores themselves being different from so much of it being disabled, or features that only work with the larger dies.

[williamcll]

The real Nvidia secret weapon against the 5600XT

[SenKa]

NOW what they do is sell this as the 2060 Super X or something so ebay scalpers can't make money off reselling these 2060.5's at obscene markups.

[GOTSpectrum]

34 minutes ago, leadeater said:

Most of Nvidia were also unaware, they confirmed it after GN contacted them. What made it so odd is that on the basic spec level there is no difference between the dies used, like for example the TU104-150 implemented on the EVGA RTX 2060 KO doesn't have a wider memory bus or something like that.

 

Edit:

The extra CUDA cores are indeed locked and not being used either, applications can't just access them because they are present in the die when they have been either microcode locked out or laser disabled. That would be a really simple explanation. Must be some deeper difference like caches being larger or the arrangement of the CUDA cores themselves being different from so much of it being disabled, or features that only work with the larger dies.

The TU-104 has a higher ratio of raster engines to CUDA Cores than the 106 due to the fact that the GPCs are laid out differently. each GPC has one raster engine, but on the 106 it has 12 SMs (384 CUDA Cores) per GPC, while on the 104 it has 8 (256 CUDA Cores). 

This may also mean that the pipelines on the 104 die are less congested than that of the 106, but this is just speculation as the level of my knowledge doesn't go this deep unfortunately. 

[Bombastinator]

The phrase “edge cases” is key here.  It’s not like they all do it.  Or even most of them.  It’s like that “up to X” phrase that means “while in certain theoretical situations you may see as much a X, in practice, you won’t”

[leadeater]

3 hours ago, GOTSpectrum said:

The TU-104 has a higher ratio of raster engines to CUDA Cores than the 106 due to the fact that the GPCs are laid out differently. each GPC has one raster engine, but on the 106 it has 12 SMs (384 CUDA Cores) per GPC, while on the 104 it has 8 (256 CUDA Cores). 

This may also mean that the pipelines on the 104 die are less congested than that of the 106, but this is just speculation as the level of my knowledge doesn't go this deep unfortunately. 

Still could be more to it since TU102 is 12SMs.

[leadeater]

3 hours ago, Bombastinator said:

The phrase “edge cases” is key here.  It’s not like they all do it.  Or even most of them.  It’s like that “up to X” phrase that means “while in certain theoretical situations you may see as much a X, in practice, you won’t”

No the results are repeatable, edge cases here means not gaming. Not all test scenarios in Specviewperf see a performance increase but there are many that do and multiple by a large amount, it's not theoretical as the tests use real software profiles so you will see those gains but only in those applications and in those specific performance areas of the applications.

[Bombastinator]

5 minutes ago, leadeater said:

No the results are repeatable, edge cases here means not gaming. Not all test scenarios in Specviewperf see a performance increase but there are many that do and multiple by a large amount, it's not theoretical as the tests use real software profiles so you will see those gains but only in those applications and in those specific performance areas of the applications.

I got the impression it was both.  It not only is specific to certain tasks, but within those certain tasks it depends on the exact way a particular cpu was crippled and apparently that varies a lot.  So some will do it and some won’t.  For chips that were crippled in a specific way, sure.  But apparently that isn’t all of them.

[leadeater]

12 minutes ago, Bombastinator said:

I got the impression it was both.  It not only is specific to certain tasks, but within those certain tasks it depends on the exact way a particular cpu was crippled and apparently that varies a lot.  So some will do it and some won’t.  For chips that were crippled in a specific way, sure.  But apparently that isn’t all of them.

It's for anything that uses the TU104-150 die which is only confirmed to be in the EVGA RTX 2060 KO cards, it may now be found on others if AIBs buy the dies but at least for EVGA that model definitely does have it.

 

Blender will always see the performance uplift, it's not as high at 47% but it's still way higher so pretty much better choice than the RTX 2070 Super. Specviewperf has 89 tests

Spoiler

Yea lol unreadable

 

The ones above 10% faster:

 

It's something around 25% of the test suite is faster than TU106 based RTX 2060.

[AkatsukiKun]

Are there any performance changes to the Tensor cores? It could be a cheaper way to get better ML parts on the cheap.

[Beskamir]

This is why I really like rtx even though I hate that it's still nvidia only as amd's still working on their version of it.

[Bombastinator]

I don’t know what the real total cost of CPUs is. Simple physical Production cost is probably pretty low.  There was a chinese company selling what are apparently old intel designed CPUs optically shrunk and packed in multiples for around $8 at one point.  They wouldn’t share total cost, because design was stolen,  but physical production would be the same.  There a bunch of other costs that have to be added though.  Design, research, etc..  that simple physical production is a heckuva lot lower than $300 though that I am willing to bet

[mr.squishy]

Is there any reliable way to obtain one of these cards? I'm building a blender/solidworks/fusion360 rig right now.

[Letgomyleghoe]

2060 ti?

[Bombastinator]

According to the WAN show It’s apparently currently just EVGA 2060KOs that have 04 chips instead of 06 chips.  They apparently get 1080 level performance in some specific tasks.

[Mira Yurizaki]

I'm going to take a stab at this, but it probably isn't enough to explain what's going on.

 

The TU-106 and TU-104 dies do have a difference, the TU-106 has 12 SMs per GPC (3 total), whereas the TU-104 has 8 SMs per GPC (6 total). Each GPC also has its own raster unit. Given that TU-106-200 used in the RTX 2060 and the TU-104-150 used in the RTX 2060 KO have the same number of SMs, 30 in this case, this means that the RTX 2060 KO has to have another GPC, which means it has to have another raster unit.

 

I don't know if this helps with these workloads, but that's the biggest difference I'm seeing between the TU-106-200 and the TU-104-150

 

EDIT: Image references:

Spoiler

 

TU-104

 

TU-106

 

 

EDIT 2: Also another musing, but I believe the 2060 KO only seemed to show noticeable improvements in professional/workstation software. IIRC, those types of software heavily favor geometry processing over pixel shading. And since raster engines deal with processing triangles into pixels, the higher performing card having more of them would to me make sense.

Edited January 26, 2020 by Mira Yurizaki

[Colonel_Gerdauf]

On 1/23/2020 at 9:37 AM, leadeater said:

Solid Edge? I think you mean SolidWork, nope Siemens just being annoying.

That.... name...

 

I understand that there are some really old companies based in non-English-primary countries, but how did nobody in Germany notice the innuendo in the name? Did Linus Sebastian travel back in time to give some random-ass company the unwitting idea of basing the name off a late-night dad joke?