Greed has no bounds. Makes complete sense. They are literally pulling an Intel.
78% nut
2080 is TU104 :
2070 is largely speculated to be an even smaller TU106 :
I'd expect a different die for 2070, given the complete lack of Nvlink, cutdown features, TDP drop and lack of detailing. There would be no sound reason besides absolutely horrid yields to use a flawed TU104 for 2070.
I long speculated that Nvidia was going to stupid, greedy, and artificially segment the hell out of this new product line and it seems I am correct. They're pulling a classic Intel move through and through. Every step of the RTX lane has a pain point that compels you further up the stack. You see the insane prices of 2080ti and its a total writeoff, so you step down to 2080 but then you're hit with price shock of $800, so you pray 2070 while still quite expensive is just a linear step down..
Nope. No nvlink sucker. Pay up
My reply : Cool. See you in 2019. I hope AMD roasts you like they did Intel. I wont be touching your products for a while.
For reference : Nvidia is going to be detailing the Turing micro-architecture and other details about RTX on Sept. 14th. I think a lot of the speculation about details will be covered at this point.
So if true, instead of the 2070 being a ~33% price increase (comparing launch prices of 10xx series cards), it will be a %100 price increase. Makes the 2080Ti's 71% increase seem tame.
Well, color me shocked. This is really unexpected, TU106 doesn't appear to be a cut down x104, it's more similar to being half a TU102 if those slides are correct.https://videocardz.com/77895/the-new-features-of-nvidia-turing-architecture
Note that TU106 is bigger than GP104 (445 mm2) and is only 100 mm2 smaller than TU104. Almost as big as GP102!
It's TU106. It's confirmed and was largely speculated to be this for some time for the reasons posted. There were already dead give away namely that Ngreedia sought to artificially segment the hell out of the product line and push pricing into the quadro level while creating a new higher premium for quadro and beyond. They've essentially pushed all of their cards into a ridiculously higher price bracket and purposely cut features to force you into the higher margin brackets. They're done for.I wouldn't say 2070 is largely speculated to be TU106. The TDP drop is in line with the specs (81% of RTX 2080 TDP with only 78% of the shaders at 95% of the boost clock). Outside the very small cut in shaders though, the 2070 has the same bus width as 2080. The last time x06 wasn't 1/2 the shaders of x04 was Kepler where it was 5/8th the shaders, but it's been a decade since nvidia released a smaller chip with the same bus width as the big one.
I'd honestly be shocked if TU106 had the same memory bus width, same L2 cache, same number of ROPs, same tensor:RTX:shader ratio as TU104 but just had 75% of the shaders and no nvlink.
https://videocardz.com/77895/the-new-features-of-nvidia-turing-architecture
Note that TU106 is bigger than GP104 (445 mm2) and is only 100 mm2 smaller than TU104. Almost as big as GP102!
Outside of SLI which is sadly a dying feature anyway, what features are they cutting out of the RTX2070 to push people to the higher bracket?It's TU106. It's confirmed and was largely speculated to be this for some time for the reasons posted. There were already dead give away namely that Ngreedia sought to artificially segment the hell out of the product line and push pricing into the quadro level while creating a new higher premium for quadro and beyond. They've essentially pushed all of their cards into a ridiculously higher price bracket and purposely cut features to force you into the higher margin brackets. They're done for.
Artificial Segmentation. They wanted to ensure Nvlink wasn't available below a certain price point and wanted to save the die space and money on cutting it out. The even further down the line cards will be cut down even further. They essentially know the 2080ti wont sell nor will the 2080. The 2070 is essentially the ceiling card for lots of people so they wanted to maximize their profit therein and cut it down to nothing. The logic is clear. They're being greedy just like intel was. They're created a schizophrenic number of products w/ artificial cut outs just like intel. They're also forgoing MCM just like Intel and going to pay dearly for it.Well, color me shocked. This is really unexpected, TU106 doesn't appear to be a cut down x104, it's more similar to being half a TU102 if those slides are correct.
At first glance I don't understand the rational for making both a 545 mm² and 445 mm² chip. The difference in SMs per GPC is really interesting though, and it'll be interesting to see the architechture deeps dives to understand why nvidia went this route.
At first glance I don't understand the rational for making both a 545 mm² and 445 mm² chip. The difference in SMs per GPC is really interesting though, and it'll be interesting to see the architechture deeps dives to understand why nvidia went this route.
A higher speed bus link with lower latency is a standard for any new generation GPU.Outside of SLI which is sadly a dying feature anyway, what features are they cutting out of the RTX2070 to push people to the higher bracket?
I paid £260 for my 8800gts.Prices have gone up just like they did with the 8x00 series, the equivalent to the 2070 was the 8800GTS (based on being 3rd SKU from the top before anyone gets snippy) at $400-450 in 2007.
That's too bad for Ngreedia then because the level of discontent across the internet about these cards is quite significant.This time, with such huge dies it makes more sense to use new dies at each segment, rather than cut down dies. As much as people like to think cut down dies are scavenging bad parts, they are mostly just product segmentation, and selling such huge dies at lower prices and higher volumes doesn't make that much sense.
That 545 mm2 vs 445 mm2 is more than than 22% savings that size alone would suggest considering the way yield drops with increasing size. Even 22% saved over a million cards starts to add up. This also implies to me, that these are not short duration stop gap cards to be followed by 7nm 9 months later as some have suggested. It can cost something like 80 million dollars to do the masking for this class of chips. You need to sell a LOT of cards to recoup that. If this were a short duration cost, you would just use bigger die cut down for the smaller volume of lower end parts. Instead we have a specific lower end part, that will need a lot of sales to recoup the up front costs.
What will be really interesting is to see what happens for 2060? It seems likely that the rumors are true about it being a GTX card, since the RT overhead seems to high for smaller chips, and what is Turing if you strip out the RT/TPUs?
Artificial segmentation would be using the same die and just disabling nvlink. Removing the feature entirely and cutting it out segments the cards, but not artificially. I honestly don't get the outrage, SLI is hit or miss on game support anyway, and really only makes sense anymore at the top end where there is no where else to go. 1070 SLI doesn't perform better than a 1080 Ti on average and is way more variable in effectiveness in addition to costing more than a single 1080 Ti. Especially with the 2080 Ti launching at the same time as the 2070, going 2070 SLI over 2080 Ti wouldn't make sense anyway from a gaming perspective.Artificial Segmentation. They wanted to ensure Nvlink wasn't available below a certain price point and wanted to save the die space and money on cutting it out. The even further down the line cards will be cut down even further. They essentially know the 2080ti wont sell nor will the 2080. The 2070 is essentially the ceiling card for lots of people so they wanted to maximize their profit therein and cut it down to nothing. The logic is clear. They're being greedy just like intel was. They're created a schizophrenic number of products w/ artificial cut outs just like intel. They're also forgoing MCM just like Intel and going to pay dearly for it.
This is why top runners always stumble... They get the most greedy at exactly the wrong moment.
These are gaming cards. The only thing that nvlink will be integrated into and is intended to be integrated into for a Geforce RTX is another Geforce RTX. If you have a specific compute application that would benefit from having nvlink present that's one thing, but that is not a gaming issue.A higher speed bus link with lower latency is a standard for any new generation GPU.
SLI is dying feature. A coherent fully featured bus interconnect is now a standard feature.
With new features like ray tracing, utilization of statistical processing to enhance performance, etc... The future is MCM with high speed/low latency bus interconnects.
RTX2070 is an abomination for this reason especially at a ridiculous $600.
If 2080ti is already having ray tracing performance issues, you might as well not even include it on a 2070 in its cut down form. The 2080 is $800 and is stupid expensive. The whole launch is a failure. Their intent as I stated about a year ago was to create new artificial segmentations and push people into quadro prices. They will fail simply because no one has any interest in paying such stupid prices for GPUs.
If GPU companies want to pretend like the GPUs are now a core compute component of a computer they better start acting like it and stop pretending like these new features are gifts from God. A CPU has evolved over the years to include a ton of new feature supports. No one has been stupid enough to try and clammer for insane amounts of money when virtualization support was added to CPUs. A more sound business surveyor would instead seek to sell GPUs in higher volumes alongside these new use cases not get even more ridiculous with margins thus choking off volume.
With the ridiculous supply of Pascal cards and the breaking of the meme chain paradigm, there is enough supply on the market to serve as an overhang for years. It's outrageous to think you can launch such a ridiculously priced set of cards and be of success. This could even stoke regression where people turn away from even the current standard pascal set and revert even further back into a retro state. The whole gaming industry is facing headwinds due to their political stunts. The PC gamer wave has already matured and peaked IMO.
The question is simple. I have a $300 Pascal 1070, what in the world are these goofballs offering me such that i'd pay double? ($600 2070)
Artificial Segmentation involves going out of your way to segment a product along arbritary lines that aren't sensible. Nvlink is not SLI. I don't think you understand exactly what a coherent low latency high bandwidth bus is. If a 2070 costs $500 and had Nvlink, there would be no reason to buy a 2080ti. People would buy two 2070s and Nvlink them for far better performance. Again, Nvlink is not SLI. I suggest you research this. Nvidia thus decided Nvlink will not be made available below a certain price point, cut it out and artificially segmented the 2070 out of the picture. Given the ridiculously Ray/sec throughput and performance of the 2070, its an absolute joke of a product thus further highlighting that it was an artificially segmented out product. Then you completely fall off the wagon below that whereby you have non RTX cards that are essentially 12nm Pascal albeit with higher prices.Artificial segmentation would be using the same die and just disabling nvlink. Removing the feature entirely and cutting it out segments the cards, but not artificially.
Again, Nvlink is not SLI. Nobody is talking about SLI here. Nvlink (nvidia) and xGMI(AMD) are new standard features of GPUs. If a GPU doesn't have it, its worthless.I honestly don't get the outrage, SLI is hit or miss on game support anyway, and really only makes sense anymore at the top end where there is no where else to go. 1070 SLI doesn't perform better than a 1080 Ti on average and is way more variable in effectiveness in addition to costing more than a single 1080 Ti. Especially with the 2080 Ti launching at the same time as the 2070, going 2070 SLI over 2080 Ti wouldn't make sense anyway from a gaming perspective.
Wrong. The line was blurred between GPUs and compute some time ago.These are gaming cards. The only thing that nvlink will be integrated into and is intended to be integrated into for a Geforce RTX is another Geforce RTX. If you have a specific compute application that would benefit from having nvlink present that's one thing, but that is not a gaming issue.
What kind of convoluted "logic" do you use to come up with that?
The TU106 model number is a poor indicator though. This die is barely smaller than TU104, it could easily be called TU105. GM106 to GP106 gave a 25% increase in shaders, GP106 to TU106 is an 80% increase. It might have the same number at the end, but it's not really a traditional x06 chip.Let's see how convoluted the logic is:
GM102 > GP102 > TU102
$650 > $700 > $1200
GM104 > GP104 > TU104
$550 > $550 > $800
GM106 > GP106 > TU106
$200 > $300 > $600
Phew, it was a tough logical road, but I think we got there /s.