Monday, September 17th 2018
NVIDIA Segregates Turing GPUs; Factory Overclocking Forbidden on the Cheaper Variant
While working on GPU-Z support for NVIDIA's RTX 20-series graphics cards, we noticed something curious. Each GPU model has not one, but two device IDs assigned to it. A device ID is a unique identification that tells Windows which specific device is installed, so it can select and load the relevant driver software. It also tells the driver, which commands to send to the chip, as they vary between generations. Last but not least, the device ID can be used to enable or lock certain features, for example in the professional space. Two device IDs per GPU is very unusual. For example, all GTX 1080 Ti cards, whether reference or custom design, are marked as 1B06. Titan Xp on the other hand, which uses the same physical GPU, is marked as 1B02. NVIDIA has always used just one ID per SKU, no matter if custom-design, reference or Founders Edition.
We reached out to industry sources and confirmed that for Turing, NVIDIA is creating two device IDs per GPU to correspond to two different ASIC codes per GPU model (for example, TU102-300 and TU102-300-A for the RTX 2080 Ti). The Turing -300 variant is designated to be used on cards targeting the MSRP price point, while the 300-A variant is for use on custom-design, overclocked cards. Both are the same physical chip, just separated by binning, and pricing, which means NVIDIA pretests all GPUs and sorts them by properties such as overclocking potential, power efficiency, etc.
When a board partner uses a -300 Turing GPU variant, factory overclocking is forbidden. Only the more expensive -30-A variants are meant for this scenario. Both can be overclocked manually though, by the user, but it's likely that the overclocking potential on the lower bin won't be as high as on the higher rated chips. Separate device IDs could also prevent consumers from buying the cheapest card, with reference clocks, and flashing it with the BIOS from a faster factory-overclocked variant of that card (think buying an MSI Gaming card and flashing it with the BIOS of Gaming X).
All Founders Edition and custom designs that we could look at so far, use the same -300-A GPU variant, which means the device ID is not used to separate Founders Edition from custom design cards.
We reached out to industry sources and confirmed that for Turing, NVIDIA is creating two device IDs per GPU to correspond to two different ASIC codes per GPU model (for example, TU102-300 and TU102-300-A for the RTX 2080 Ti). The Turing -300 variant is designated to be used on cards targeting the MSRP price point, while the 300-A variant is for use on custom-design, overclocked cards. Both are the same physical chip, just separated by binning, and pricing, which means NVIDIA pretests all GPUs and sorts them by properties such as overclocking potential, power efficiency, etc.
All Founders Edition and custom designs that we could look at so far, use the same -300-A GPU variant, which means the device ID is not used to separate Founders Edition from custom design cards.
90 Comments on NVIDIA Segregates Turing GPUs; Factory Overclocking Forbidden on the Cheaper Variant
It appears that purpose of dual device IDs is so AIBs can tell the difference. It also might be an attempt to prevent the changing of devices IDs. And/Or also to keep etailers from more price gouging and selling the lesser/weaker binned chips at the same price as OC variants and keep them closer to MRSP pricing.
2cp
No chip will be blocked for OC, but cards with better cooling and better power design, will receive the more binned chips to overclock even better. Which make perfect sense, since the buyer of an Zotac AMP! Extreme for example, pay more for that than who buys a basic Zotac card.
Just. Wow. How can a GPU scream 'AVOID ME' even louder?
LOL consider that - the 2080 already just barely seems to edge out the 1080 Ti, I bet vanilla versions lose to it.
The only real difference here is pricing, and that completely destroys the whole principle and incentive to upgrade this time, if they make such a performance jump.
gaming will soon be a service with nvidia is what came to mind
:roll::roll:
Like Apple nvidia can look forward to controlling all aspects of their hardware and marketing.
But what's going on right now is Nvidia is selling a TU106 die for $600 lol, and then calling it a xx70. xx106 is the same bracket used for the 550 Ti and the 650 Ti. It's pathetic.
-GV/TU100
-TU102
-TU104
-TU106 (2070)
-TU108
Low-end is being sold for $600...
NVidia certainly doesn't want to see reviews of pedestrian cards with poor performance tainting their image. I think this also ties to the since abandoned partnering thing. The poor overclock ability of the AMD version of the card, "taints" the brand. Also, in this manner, the consumer gains some protection. What's the use of that improved AIB PCB with 14 phases, beefier VRMs and better cooling if the silicon lottery pairs it with a weaker GPU ? Also, hopefully we'll see an end to product lines like EVGAs SC series which almost always come with a factory OC but only a reference PCB which is inevitably limiting.
If I get something like the FTW or Zotax AMP Extreme, I want to be sure that I am able to OC or boost higher than the reference card. Previously it was a lottery by chance. Now you pay a premium but will get a card that will OC higher. I think it is a fair trade-off for those that can afford it.
Pascal is actually pretty god awful at overclocking. There are very real frequency walls no one can get past, and the biggest boosts I have seen is about 9% on most models (Yes, there are exceptions). On the other hand Vega cards often see very real 20% boosts from tweaking, and the 7000 series was legendary in its overclocking abilities. Polaris and Fiji definitely were limited to about 10% gains though - that's true.
GK106 - 221m²
www.techpowerup.com/gpudb/1188/geforce-gtx-650-ti
I remember buying a GTX 660 at launch for 220 EUR. But there was also a 660ti on the same die... at about 300 EUR.
TU106 - 445m² - twice as big
Suddenly a 600 EUR price point isn't all that surreal in fact its a precise match if you compare die size of the 660ti + price... IF all of that die space was raw performance. Paying the premium just to drive RTX is ridiculous.
EDIT: this is also why I feel Nvidia is taking a huge risk with RTX. They are fast going the AMD route of chips with a lot die space reserved for tasks with questionable benefit for a use case - its the reason GCN can't really keep up.As for Pascal's OC capabilities; I see this as a matured architecture much like Intel's Core - they know it so well they can OC out of the box and not worry about stability. Overclocking these days is very much a marketing gimmick, really.
295mm² - still is significantly smaller though and it had disabled SMXes including a bus width cut to 192 bit.