NVIDIA mid range for top price

This is a VERY interesting video.

New volta or whatever NVIDIA wants to call em are basicially pascal cards as I already expected.

 

Nothing new, been this way since Kepler. You can blame GCN's initial lack of competitiveness since Nvidia could beat their top dog a the time (7970) with midrange GK104 so they were in no way threatened to bring out the big gun (GK110) first.

 

I think its a mixture of lack of AMD competition and the fact that games are not really demanding anymore thanks to consoles. The only games that struggles are games that are unoptimized. Lets face it, you can play pretty much all games on near ultra settings 1080p on an old kepler GTX 780.

 

That's a bit of an exaggeration. It doesn't have enough VRAM, and Kepler really hasn't aged well in general. It gets slaughtered by Maxwell/Pascal (780 equaling 960 and slower than 1050 Ti) and GCN in many games from the last few years and anything using DX12 or Vulkan. There were a lot of conspiracy theories floating around a few years ago about Nvidia purposely gimping Kepler performance, but the reality is they simply stopped optimizing their drivers for Kepler and the Kepler architecture is less efficient in modern engines when handling large amounts of tessellation and compute and also has higher CPU usage than Maxwell/Pascal at the same frame rate.

 

It really isn't, when playing 1080p 4GB of vram is enough for most games and will run games no problem on very high settings. You are right tho, maxwell kicked off a new good generation with so much potential, it was the best architecture thus far. Sure pascal did a lot in terms of efficiency, but in terms of customizability NVIDIA had brainfarts.

You gotta think about it, kepler released in 2013 5 years later still playing games no problem.



(CPU bottlecnnking Asassins cree origins)

But you get my point. I'm not saying there didn't happen much in terms of performance, I'm just saying that the performance gained over time isn't as big anymore as it used to be.

Also a GTX 780 is much faster than a 960, it's slightly slower than a 970, and 960 is noticably faster than a 1050 ti.

 

780 has 3GB VRAM. But that's what I'm saying. 780 is supposed to be much faster than 960, but in most games of recent years it's essentially tied or even beaten by the 960 because Kepler has aged horribly, and it looks atrocious compared to its 2013 direct competitor, the 290.

Just look at this

 

I'm aware of that, some games run way better on AMD cards than NVIDIA and the other way round. In DX 12 and Vulkan the RX580 beats the GTX 1060 getting close to GTX 1070 performance. It's very weird lately.

depends on model. There are 6GB versions as well, such as the ASUS Strix GTX 780 6GB or the EVGA GTX 780 SC 6GB

 

That too, but it's sobering how bad Kepler looks compared to Maxwell and Pascal these days.

 

It's NVIDIAs way of giving their own cards the apple treatment while AMD makes the old cards run like ferarris.

It's only a matter of time when Maxwell will get worse for no other reason than NVIDIA wants to make you buy their new products.

 

It's because AMD has been on GCN this while time, while Kepler to Maxwell/Pascal is a much bigger architectural change.

 

Yeah that's what AMD does. Same with ryzen, they stick with an architecture for quite some time, by doing that they increase support and performance for users even if they bought something years ago. This makes AMD cards more reliable than NVIDIA cards.

 

Have a look at this video where they retested the RX 580 vs GTX 1060 in Doom but with 2018 drivers:


It appears Nvidia have improved the Vulkan performance of their drivers since Doom was released in 2016.

I don't think that Nvidia will stick to using the Pascal architecture for the 11 series GPUs. Pascal is basically the same as Maxwell, but with better delta colour compression and on 16nm FinFET. Nvidia seems to have a tick-tock roadmap:

GTX 400 - architecture (fermi), gtx 500 - refresh. GTX 600 - architecture (kepler), gtx 700 - refresh. GTX 900 - architecture (maxwell), GTX 1000 - refresh.

They will be using GDDR6 on the new cards meaning they will have to redesign the memory controller. With Volta (specifically GV100), they completely reworked the cache hierarchy and changed the CUDA core count per SM from 128 down to 64 (better support of async compute). They also tweaked the scheduler as well. The Titan V performs better in DX12 per core per clock than Pascal because of this.

Nvidia have spent billions of dollars developing Volta. It would make little financial sense to release a Pascal refresh and leave untapped performance on the table when you already have a new architecture. Increasing the CUDA core count and moving to 12FFN on a Pascal refresh will not increase performance by a large enough margin (+30% max?) that people who have a gtx 10XX card (or even a gtx 980) would bother upgrading to new cards. This is not even considering the fact that the increased die size on the Pascal refresh would hurt yields and therefore push up the price.

 

Nvidia Vulkan drivers have definitely improved. It was only within the last 6 months that Vulkan finally started to perform better than OpenGL in Doom on my previous i7-4720HQ and 980M (Optimus) system. On the drivers before that, Vulkan would either run worse than OpenGL or would stutter or would crash and not work at all.

GPUs don't follow tick-tock, the new process and new architecture happen together. But Moore's Law has slowed down even for GPUs.

 

It depends what you mean by a new architecture. Pascal changes relatively little when compared to Maxwell, but Maxwell changed pretty much everything when compared to Kepler. Therefore, I'd call Pascal a "refresh" of Maxwell.

 

Pascal isn't a refresh. 500 and 700 series were refreshes.

 

Pascal is built mostly off the maxwell arch, not new.

 

What is upcoming is the 2nd dieshrink and tweak of a 4 year old microarchitecture and it will wipe the floor with the opposition's 2nd dieshrunk/4th tweaking of their 7 year old GCN architecture - so there's just no push for them to do anything revolutionary... but you can very well argue that they haven't needed to do anything else but rely on dieshrinks and new memory tech to get the 30-40% leap from maxwell-pascal and that is also the rumours about the pascal-volta delta; you could say Nvidia's approach has been completely vindicated.

Why fire your best artillery when your opponent is still miles away from the battlefield?


The exploitation of the market in the absence of a competitor is behind their pricing bracket creep. They are a publicly listed corporation with a shareholder duty to make as much money as they can, they are not a charity and they make a product that is a completely discretionary luxury item, not one where access is deemed an essential human right, so no government gives a firetruck about crying GPU buyers (except for highly public isolated instances of brainlessly obvious anticompetitive or price fixing behaviour over the decades; there's a lot of chicanery that goes uninvestigated, unremedied, and unreported)

MXM board prices have been ridonkulous for decades due to exclusivity (the only ones making the boards don't want you to buy them separately, they want to sell an entire unit to replace a slightly broken repairable one) and now desktop GPU owners are feeling a bit of the karma from all the "cheaper to build a desktop bekoz gaming laptops are stupid" comments over the years.

 

Yes, however, AMD managed to keep up with Nvidia even with their GCN architecture which was refined.
The 1080ti for example is nearly matched by Vega 64 LC (but falls short by 10% - personally, I don't see it as a bit issue).
Make note of the fact that both RX 4 and 5 series (Polaris) and Vega cards when undervolted match or exceed Nvidia's equivalents in terms of efficiency, and in terms of performance, AMD cards at lower clocks match Nvidia's at higher clocks... and we know AMD overvolts their GPU's to increase yields and packs a lot more compute units which are power hungry, plus Nvidia had an advantage of using a better manuf. process better suited for higher frequencies and lower power draw (basically, Nvidia can clock higher before reaching a point where power draw starts jumping like mad, whereas for AMD, the threshold is lower and the process produced lower yields - so if AMD was able to use TSMC 16nm like Nvidia did, I think we'd see much lower stock power draw and higher performance on existing AMD cards - but AMD will finally be able to use TSMC 7nm process like Nvidia, and possibly achieve parity... we'll see - depends on how much has changed between Pascal and Volta, but thus far, Volta seems like it will be used mainly in AI, not for consumer cards - and AMD will likely just use Infinity Fabric from now on to further decrease costs and increase yields).

 

Pascal is effectively a die-shrunk Maxwell with higher clocks.
Same underlying architecture though.

 

You can drop 30-40W while getting better performance than stock by undervolting the 1080, as I have shown in the past.

There are still hardware feature set differences though. It's not a refresh of the same architecture.