AMD, Intel and Nvidia driver issues and last recommended version

[SamuriHL]

Imagine there is a comma there:


Thus, not using PC mode.Exactly. Bad grammar is all. Lol

Sent from my SM-G975U using Tapatalk

[NikosD]

A few more clarifications regarding the sneaky moves of nVidia regarding Ampere.

The cores are not exactly capable of 2x FLOPS versus Turing.
Each core is capable of INT + FP (like Turing) OR FP + FP (Ampere only) in every cycle.
So, if ALL instructions are FP then 2x FP is possible, but they are not.
And because there are also INT instructions, Turing (using INT + FP) is faster than Pascal.

So, the max TFLOPS number is wrong (in general for every case)

Also, 3080 has the same family GA102 chip like 3090.
It's the maximum configuration reserved only for Ti or Titan in the past.
For example, 2080 has the TU104 and not the larger TU102 chip that it was reserved for 2080 Ti
1080 has the GP104 chip and not the larger GP102 which was reserved for 1080 Ti.

What does this mean ?
It means that the performance difference of 3090 vs 3080 will be as I have already reported very low (~20%) almost half than 2080 Ti vs 2080 or 1080 Ti vs 1080 because they share the same huge family chip, like never before.

Why does nVidia scaled up for the first time the x080 card using the same family chip like the faster Ti/ Titan ?
Because nVidia is afraid of Big Navi performance and wants to be sure that 3080 is as fast as it could be, faster than ever before, pushing the performance to the edge, matching a normal 3080 Ti (if it could exist)

nVidia wants the official flagship 3080 to be faster than Big Navi by all means even using a bigger chip that used to be leveraged by Ti/ Titan and is selling it at the same price of 2080, which is insane.

nVidia definitely counts on the idiots buying 3090 for 20% more performance than 3080 with a 110% difference in price, in order to balance the reasonable price of 3080 and 3070.

I'll be back.

[huhn]

The cores are not exactly capable of 2x FLOPS versus Turing.
Each core is capable of INT + FP (like Turing) OR FP + FP (Ampere only) in every cycle.
So, if ALL instructions are FP then 2x FP is possible, but they are not.
And because there are also INT instructions, Turing (using INT + FP) is faster than Pascal.

great update isn't it? INT 32 is barely used at all they are just like AMD cores now. so that means the max Flops on both are wrong? and only turing is correct because they are dedicated cores sleeping all the time?

[nevcairiel]

FLOPS literally stands for "floating point operations per second", which is always a theoretical max value independent of any other constraints or workloads - so it cannot be wrong. For the context of Ampere, a single core is capable of twice the FLOPS as a Turing one.
FLOPS are literally calculated by something rather simple like "core count * calculations per tick * clock speed", it does not account for anything else. Its not measured. Its not based in reality. Its just math.

FLOPS are a very theoretical value, its not particularly meaningful for real-world performance. Never has been, on any vendor. But companies (all of them) also like to use it to boast, because its easy to make those high numbers.

Its sorta funny to see people scramble to make up reasons why everything is terrible.
Sure, the 3090 is too expensive, but it practically has the specs of a Titan RTX with all that memory and costs almost half of that. As everyone keeps saying, the performance gap is likely small enough to just get a 3080 instead, so might as well treat it like a Titan and ignore it. :)

The 3080 and 3070 are priced/spec'ed very competitively. Of course waiting for RDNA2 might be sensible if you have the patience, maybe it won't be too long anymore. But even without that, the 3070/80 are very worthy replacements of the 2070/80.

[huhn]

i agree with most points but Flops are "useful" getting a rough estimate of the performance between cards from the same generation and only the same generation.

[chros]

Thanks Nev for FLOPS info.

A few more clarifications regarding the sneaky moves of nVidia regarding Ampere.

The cores are not exactly capable of 2x FLOPS versus Turing.
Each core is capable of INT + FP (like Turing) OR FP + FP (Ampere only) in every cycle.
So, if ALL instructions are FP then 2x FP is possible, but they are not.
And because there are also INT instructions, Turing (using INT + FP) is faster than Pascal.
Yes, something is weird with such a huge cuda core count jump.
I understand that it doesn't scale linearly, but I'd wait for 1.5x return out of 2x.
In reality probably it will only be 1.25-1.30. So where the rest of 20% went?
Some think that by adding double of the micro-cores on a smaller node, IPC suffers about 30% decrease compared to Turing.

[huhn]

a 2080 ti has 4352 cores. that means 4352 FP32 and 4352 INT32 cores so it technically has 8704 cores (50% are very small and mostly "useless/barely used" cores).
a 3080 has 8704 that means 8704 FPS32 and 4352 INT32 see the image below.

a cluster now has 2xFP32 not a core is doing 2xFP32. and we got this new not confirmed image: https://cdn.videocardz.com/1/2020/09/NVIDIA-GeForce-Ampere-SM-Layout_HardwareLuxx.jpg they just added the same FP32 cluster again on a different datapath where only the INT32 cores or the FP32 cores are usable so the INT 32 cores are not gone.

a 3070 has 5888 cores while they clearly don't scale linear to a 3080 with 8704 cores but it will be far more then 50 %... but this is comparable they are the same they physical there.

IPC. if nvidia claims are true and are generally applicable across the board (which you should never believe in).
a 3070 is supposed to be as fast as a 2080 TI.
assuming no INT32 was used when we compare 4352 cores at ~2000mhz vs 5888 at ~1700mhz then the IPC for FP32 is 13% lower on ampere and this is the worse case scenario treating the double FP32 as real cores as we should be.

[NikosD]

FLOPS are literally calculated by something rather simple like "core count * calculations per tick * clock speed", it does not account for anything else. Its not measured. Its not based in reality. Its just math.

FLOPS are a very theoretical value, its not particularly meaningful for real-world performance. Never has been, on any vendor. But companies (all of them) also like to use it to boast, because its easy to make those high numbers. OMG, I hope Top500.org is not reading this thread and this post!
The whole world of SuperComputers nowadays are based on real-world benchmarks of FLOPS, meaning that besides the theoretical value of FLOPS that someone can calculate of a processor, the whole world besides you probably, is interested in how many actual FLOPS the processor is capable of, given a specific workload or even a favor workload.
For example, at the last presentation of Radja showing Xe GPU, he used an OpenCL tool that is capable of trying different worksets in order to find maximum theoretical but actual values of FLOPS of a CPU/GPU, meaning that the figure is a best case but real world scenario.
So, we have a theoretical number of FLOPS, we have a theoretical best-case fully optimized number of FLOPS and we have a real-world actual sustained FLOPS leveraging common case workloads as those of benchmarks or real-world apps, like 3D games.
In conclusion:
FLOPS are not only theoretical but also real world and we are mostly interested in real-world measurement of FLOPS.
They are measured in a lot of cases and a lot of different workloads and scenarios.
They are actually the most used benchmark of SuperComputers since the beginning of their existence and of every actual processor (CPU/GPU) ever built, with hardware support of FP.
They are based on reality, because every actual program which uses FP numbers they are actually doing a FLOP (Floating Point OPerations) in hardware that supports FP numbers.
It's not just math, but are useful math in order to test/benchmark real-world performance of workloads/applications like 3D Games since we are talking about graphics cards.
That post that I just replied, could easily be one of the most misleading posts ever written in this forum!
I hope Top500.org measuring the Top500 Supercomputers of the world haven't committed suicide by laughing or crying, reading the post above.

[huhn]

you know what compare a vega 64 and a 3700XT performance and flops and then say that again...

GPU driver and game engines cheat at every corner. just buy a card based on FLOPS because they are the metric.

ridiculous...

[el Filou]

OMG, I hope Top500.org is not reading this thread and this post!
The whole world of SuperComputers nowadays are based on real-world benchmarks of FLOPS, meaning that besides the theoretical value of FLOPS that someone can calculate of a processor, the whole world besides you probably, is interested in how many actual FLOPS the processor is capable of, given a specific workload or even a favor workload. [...blahblah...] I hope Top500.org measuring the Top500 Supercomputers of the world haven't committed suicide by laughing or crying, reading the post above.That's why Nev is only talking about *theoretical* FLOPS, which are the *only* value ever advertised in consumer GPU marketing materials. When Microsoft says "Xbox Radeon GPU has 12 TFLOPS" or NVIDIA says "RTX 3080 has whatever FLOPS", it's not a benchmark, it's a simplistic theoretical formula like Nev explained. He never said you can't benchmark specific code and get a number out of them, but that's not what GPU vendors are advertising.Its sorta funny to see people scramble to make up reasons why everything is terrible.It's because those people are AMD fanboys who just want to rejoice in seeing their competitors not absolutely dominate for once.
The fact some are saying AMD is "destroying" Intel in CPUs while they only got competitive again is saying a lot.

[NikosD]

There are no performance numbers for 3090 yet.
I haven't seen a single game measured with this card and it's going to be released only one week later than 3080.
nVidia is clearly hiding 3090 because the performance difference from 3080 is so small to justify the 110% price difference.
The good news is that AMD is going to release Big Navi or 6000 series probably on October 7th, just two weeks later than 3090.
And Big Navi cards are not going to be paper launch cards like Ampere, because the superior TSMC 7nm process has already produced 1 billion chips of various kinds.
Very good yields for TSMC 7nm, with great performance and efficiency.

Boycott 3090 cards.

Friends don't let friends buy 3090 cards.

Wait for Big Navi and decide with all cards on the table, but not 3090 cards.

Everything but 3090 for God's sake!

[SamuriHL]

I don't think even nVidia thinks anyone will really buy a 3090. It's more the "look at me" attention seeking card. The rumor today is that AMD was planning a 16gb card for 599 and they're now talking about bumping that down to 549. As I said before, this only means good things for consumers. I was originally looking at the 3090 because I genuinely didn't expect the 3080 to come out so strong but since the launch event I'm only looking at the 3080 at this point. Competition is what forced nVidia's hand here. They would have been quite happy to sell you a 10-20% bump over comparable 20x0 cards and raise prices while doing it, but, the fact that they absolutely killed it while keeping prices the same is a huge benefit to us. And the real world benchmarks for the 3080 are starting to come out. Not the 100% over 2080 claim that they're making, but, somewhere in the neighborhood of 75-92% in real world terms of AAA gaming titles. I mean, you gotta give that one to them. That's absolutely unreal and hasn't been seen since 2004. I'm not discounting AMD, but, I think they were likely expecting the 10-20% usual bump over the previous gen and thought they could sneak in comparable performance at a cheaper price. I doubt they have anything that's going to best a 3080 in terms of performance. Maybe I'm wrong. That'd be nice if I am.

[NikosD]

nVidia talked about almost 2x performance of 3080 vs 2080 not 2080 Ti, which is close but not 2x.
I expect around 70% on average for the 15 most popular games e.g techspot's suite.
The difference of 3080 vs 2080 Ti could be around 40% to 50%.
Big Navi could be close to 3080 even though nVidia named the 3080 Ti (GA102) released card as 3080 card.
They changed level but AMD can keep up, I think.
We'll see in a month...

[SamuriHL]

Yes I know, I edited my post as that was a typo on my part. It's still highly impressive.

I don't think AMD can shift on a dime. nVidia played their hand first. It's not like AMD was sitting on their hands this whole time. I don't know what they have planned, no one really does, but, given past history I just find it difficult to believe they're going to match a 3080. Maybe they will and that would be good. I absolutely believe they will beat the 2080ti.

[nevcairiel]

The technical embargo on Ampere expired this evening, although info was only given to the press after the event (ie. a few days ago), so expect the good tech publications to have articles out after analyzing the info, whenever they manage to get through all of that.
Benchmarks are likely still out until closer to launch, when cards are also being sampled to reviewers.

Highlights:
- The 10000+ CUDA cores are real, there is that number of FP32 capable ALUs that the scheduler can all individually use as the work load permits (128 per SM, Turing had 64 per SM, Pascal however also had 128)
- RTX IO works in the CUDA cores, and comes to Turing as well. The INT32 cores are likely a big factor here.
- RT and Tensor can now be used simultaneously in a core, which boosts tensor+RT performance, perfect for RT+DLSS rendering
- The 3080 has 96 ROPs, the 3090 has 112. For comparison, the 2080 Ti had 88. The ROPs are what pushes the final pixel to the screen and a key performance component next to the memory and cores.

I do sort of want a 3080 with more memory, but we'll see on reviews if that comes close to being a concern.

[SamuriHL]

Holy God, nVidia. I'm really impressed. Yes, the memory thing is a concern. I'd like to see 16 or 20 but oh well.

[nevcairiel]

My gaming system is on 1440p and I have no huge plans to change that, since at this screen size and viewing distance 4K does nothing but eat my FPS, so that saves me a bunch of memory. Will likely be fine. Would've just been nice.

Memory is a technical challenge at this point. They can't just offer a 3080 with 16GB or such, not without affecting everything about the card.
GDDR memory is all linked up to individual memory channel, each being 32-bit in width (technically 2x 16-bit on GDDR6, but same thing). And memory chips come in 1GB.

So 10GB? Fine. 10x 32-bit = 320-bit bus, 10 chips. All fine.
20GB? Sure. Do the same thing, connect two chips per channel.
16GB? Uhh. 8 channel with 2 chips each? But that reduces bandwidth. Full 16 channel? Way too expensive to build on the board.

You also cannot put more chips on some channel only, then you get the 970 situation where memory is off-balance and some parts are slower.

11 or 12 GB might've been possible by using the full memory controller loadout that the 3090 has (12 x 32 with 2GB per channel), but they likely needed to cut there to get the chip yield up.

20GB is probably too much, while 10GB might feel slightly constrained. But the nature of GDDR memory controllers dictates this, unfortunately.

Not sure if AMD is once again going to use HBM, which has similar constraints except that it comes in multiples of 4 or 8, so they can do 16GB. But HBM is also very expensive.
Or they could go for lower bandwidth and more capacity (ie. the 8 channel approach) with GDDR6.

[SamuriHL]

I game at 4K but I still don't think it'll be a problem.

[huhn]

- RTX IO works in the CUDA cores, and comes to Turing as well. The INT32 cores are likely a big factor here.

good point but doesn't nvidia have fixed function for decompression and compression similar to the lossless memory compression?

[NikosD]

The first game ever that used CUDA cores do decompress textures, was RAGE on 2011.

With RTX IO, nVidia looks cheap again by excluding Turing GTX cards.

[ryrynz]

Turing will support RTX IO (https://www.nvidia.com/en-us/geforce/news/rtx-io-gpu-accelerated-storage-technology/).

[huhn]

let's see if AMD will look cheap with navi.

[NikosD]

nVidia at the first comment:
RTX IO is supported on all GeForce RTX Turing and NVIDIA Ampere-architecture GPUs

[chros]

Boycott 3090 cards.
Friends don't let friends buy 3090 cards.
Everything but 3090 for God's sake!
:D

I game at 4K but I still don't think it'll be a problem.
I think 10GB is just on the limit :(. 1080Ti had 11 and Titan X(p) 12.

Not the 100% over 2080 claim that they're making
:D I just watched again Jim's (AdoredTV) 2 years old nVidia turing marketing survival guide (https://www.youtube.com/watch?v=b_x1JGG4JC8) video in 25 mins when he only talks about 1 (!) slide :D (he talks about memory bandwidth) It's hilarious to see how they operate.

They would have been quite happy to sell you a 10-20% bump over comparable 20x0 cards and raise prices while doing it, but, the fact that they absolutely killed it while keeping prices the same is a huge benefit to us.
...
I mean, you gotta give that one to them. That's absolutely unreal and hasn't been seen since 2004.
Yes, some already said that it's the first generation for a long time (since Fermi?) when a xx80 class GPU gets the 102 die! (with Turing/Pascal it had 104 die). That means almost the full die in a generation! It's basically a Ti class card in the previous generations. :)
That's why the claimed huge performance uplift compared to the previous generations!
I'm pretty sure they were/are scared of something :) (consoles/amd/etc ?)

And that's one of the reasons why the power usage is in the clouds with the 3080!
Along with Samsung 8nm, that Jim state in his new Ampere analysis (https://www.youtube.com/watch?v=1milKOPa9zA&t=2s) that it's only around a half-node shrink compared to TSMC's 12nm!

One more thing to note: it can happen (due to how nvidia treats/treated their partners in the present/past) that they will be stuck with Samsung for a long time! And this can be a real issue in the long run (2-4 years later).
They won't be able to do this very same trick 2 generations later, because they already gave the full die ... :)

- The 10000+ CUDA cores are real, there is that number of FP32 capable ALUs that the scheduler can all individually use as the work load permits (128 per SM, Turing had 64 per SM, Pascal however also had 128)
Interesting about Pascal, thanks.

- RTX IO works in the CUDA cores, and comes to Turing as well. The INT32 cores are likely a big factor here.
What do you think, can this be interesting for us, playing back (high bandwidth) videos or not?


let's see if AMD will look cheap with navi.
:) It will be interesting to see what will happen, there are controversial "insider" information, some state it's nowhere near, some state it'll be competitive. I hope the latter, it's good for the industry and us in the long run.

[nevcairiel]

What do you think, can this be interesting for us, playing back (high bandwidth) videos or not?


Its doubtful that its useful for video. Afterall video decoding is already on the GPU, with near to no CPU load, and the bandwidth requirements arent going to massively increase. If anything video bitrates are reducing, or at best stagnating.

[chros]

Thanks.

Yes, some already said that it's the first generation for a long time (since Fermi?) when a xx80 class GPU gets the 102 die! (with Turing/Pascal it had 104 die). That means almost the full die in a generation! It's basically a Ti class card in the previous generations. :)
That's why the claimed huge performance uplift compared to the previous generations!
And if I look at it this way, that means the 3070 is a xx80 class card and hopefully the 3060 will be a xx70 class card for (hopefully) $300-350! The latter will be a real deal for lot of users with hdmi 2.1 and av1 decoder ...

One more thing to note: it can happen (due to how nvidia treats/treated their partners in the present/past) that they will be stuck with Samsung for a long time! And this can be a real issue in the long run (2-4 years later).
They won't be able to do this very same trick 2 generations later, because they already gave the full die ... :)
And if I think of this more, I'm pretty sure the next gen nvidia GPUs will "only" bring the Turing type of increase tops, nothing like we see now.
So, unless AMD's RDNA3 (not 2) will bring massive change, these 3000 cards will be a good value for a long-long time.
Maybe I should also buy a 3060 (although I said I won't :) ), but let's wait for the specs and real benchmarks first, and hopefully some madvr user reports as well ... :)

[NikosD]

NVLink SLI is reserved for 3090 card only.

But why ?

Because people shouldn't have the chance to combine 2x3080 cards with the price of a single 3090 and get ~80% more performance than 3090.

I think a heavily overclocked custom 3080 card could overcome 3090's performance.

[QBhd]

...

Why? Because SLI/CF is totally dead from a gamer perspective... More problems and headaches and games don't really leverage the ability. For compute workloads, yes. Hence, the new "Titan" (3090) has it.

QB

[ryrynz]

get ~80% more performance than 3090.


Two 2080s in SLI couldn't even beat a 2080Ti on average and most of the time managed only somewhere in between a normal 2080 and a Ti.
3080s in SLI would be a serious waste of cash and the above reply from QBhd factors in as well. I think from a logical standpoint this decision makes sense, SLI is basically dead in it's current form (with gamers)

[NikosD]

3080 has the same chip like 3090.

2x3080 combined would be a killer combination for the use of NVLink SLI

20GB of 2x3080 VRAM vs 24GB of 3090 VRAM and 80% more performance of 2x3080 vs 3090 for the same price.

For the use of NVLink SLI.

[butterw2]

hopefully the 3060 will be a xx70 class card for (hopefully) $300-350! The latter will be a real deal for lot of users with hdmi 2.1 and av1 decoder ...


It seems these are the 2 only new features of rtx 3000 (vs rtx 2000). Maybe a used rtx 2060 Super 8GB can be had at a more reasonable price once the 3060 and 3050 and navi2 are out ?

I bought my current PC new, monitor/SSD-HDD included, for less than the starting price of the "cheap" RTX3070. It also uses less power than the RTX3080. So yes it has less TFlops, but it is also silent and has no driver issues (don't update).

I can't help noticing Nvidia does a form of planned obsolescence with the GDDR6 quantity on their cards. Rtx 2060 only had 6GB VRAM. While 8/10GB is enough right now (for 4K gaming), what will the situation be in 2 years time, given that the next-gen consoles come with 16GB ?

More relevantly how much VRAM do you actually need on the gpu to be able to make real use of AI features ? Will MS DirectStorage (RTX-IO) be applicable to Machine learning to overcome VRAM limitations ?

PCIe4, M.2 NVMe SSD, zen3 compatibility look like useful features for future upgrades.

[NikosD]

For those who think with an eye on the next-gen games at 4K resolution, a fast enough Big Navi with 16GB VRAM will be a better choice than 3080 at 10 GB VRAM.

If it manages to be competitive though.
Which I think it will.

[chros]

For those who think with an eye on the next-gen games at 4K resolution, a fast enough Big Navi with 16GB VRAM will be a better choice than 3080 at 10 GB VRAM.
Like last year, right?! :D We have to wait for it to see ...

It seems these are the 2 only new features of rtx 3000 (vs rtx 2000). Maybe a used rtx 2060 Super 8GB can be had at a more reasonable price once the 3060 and 3050 and navi2 are out ?
Well, it depends for what. If you want to fool around (coding) with the RT cores, then probably yes, because you can't play games with it :)
But even then if you can get (let's say) with $100 plus a truly useable card with the aforementioned features then I'd go for it instead. (If you don't need the RT cores at all then just go for a used Pascal, they will be dirty cheap :) )
I hope nvidia won't dare to come out with a 6GB 3060 :D

I can't help noticing Nvidia does a form of planned obsolescence with the GDDR6 quantity on their cards.
I'm also pretty sure that's the case! :) As others said, there will higher versions as well but for higher price (maybe called Ti/Super/etc.) :)
And they have to think of the next gen GPUs as well, otherwise they won't have any card left in their hands :)

[NikosD]

No Founders Edition RTX 3000 series card will be sold in Australia.

Aussie buyers have to wait for custom cards from OEMs

Don't ask me why.

Ask the leather-jacket-man instead.

Happy days of new nVidia products!

[butterw2]

hopefully the 3060 will be a xx70 class card for (hopefully) $300-350! The latter will be a real deal for lot of users with hdmi 2.1 and av1 decoder ...

A 300$ card from Nvidia ? Maybe a custom rtx 3050, with 6GB GDDR6 ? Can do raytracing at 1080p with DLSS enabled in supported titles such as fortnite. :D

[chros]

:) Yep, $300 3060 8GB is a bit unlikely, but who knows? If 3080 $700 and 3070 $500 ...
I think it will all depend on AMD (this time again).

[ryrynz]

I can't help noticing Nvidia does a form of planned obsolescence with the GDDR6 quantity on their cards

The way you talk about it is like the performance absolutely tanks or the card just doesn't render certain elements or even stops working if the required texture memory is higher than the card has onboard, none of these situations is the case.
At most you'd lose maybe ~20% or less FPS if you had a game needing say 16GB or more texture memory, it may actually ending up being less on these new cards with RTX IO.
Anyway, you can always just lower the texture quality down a notch of two and you're golden. It's up the the devs how much memory they want to allow a program to use and up to you to set the value appropriately.
Games are made for today's hardware not tomorrows, purchase what suits your needs, Less out of the box performance but with more RAM? Looks like it's AMD for you :P
Remember the market is designed to make you want to buy new hardware and companies are rewarded for selling it to you, you can't expect a capitalist market to operate any differently than that.

I do feel that 10GB on the 3080 seems a tad on the low side for a card of this calibre, however Tensor Memory Compression is going to reduce texture memory requirements upwards of 20%
so there's that too. There will be plenty of reports of how this holds up to the most demanding games later on this or early next year, it'll will probably do better than you think.

[NikosD]

After the announcement of a real rocket from Sabrent using Phison E18 controller for the PCIe v4.0 NVMe SSD, which eats Samsung 980 Pro for breakfast, the last obstacle for this year's upgrade has been lifted.

I'm ready to move to PCIe v4.0 platform for both Graphics Card (less impact) and SSD (could be more useful with suitable software)

I was hoping for a battle in the CPU arena between Rocket Lake and Zen 3, but unfortunately Rocket Lake moved once again to Q1 2021.

The battle on GPU arena will be intensive between Big Navi and Ampere cards and we expect it around October.

All reports say that Navi21 will be faster than 3070 and probably competitive with 3080 with more VRAM and better price as always.

For my new rig and my budget, I have to wait for 3050 Ampere card (if ever is going to be released) and the RDNA2 card of the same budget around 250$

I'm very pleased with 1080p gaming and I need nothing more.

The only case that I would wait for Rocket Lake or buy an Ampere card would be, if RDNA2 doesn't eventually support AV1.

Then probably one of my two SoCs (CPU or GPU) should support AV1 so it's either iGPU Xe inside Rocket Lake or Ampere 3050.

If RDNA2 with Big Navi supports AV1, I will probably go ALL AMD after many many years since AthlonXP and ATi Radeon 9500 -which wasn't AMD(!) actually.

Around December close to the holidays season, a new rig could be a nice gift!

[NikosD]

We had also another release this week, which is actually a non release but rather an announcement.

But I will leave Charlie speak with the voice of reasoning:
More to the point, Intel can’t actually supply Tiger Lake CPUs. SemiAccurate’s checks with OEMs all say they Intel is shorting them on supply by 7-digit numbers in several cases.

If you are a glass half full person, you could say demand is high.
If you are realistic, you can say that yields are still not viable on 10nm, and wafer starts are… the number is shockingly small.

Because of this, we'll talk about DJs, how often people re-encode and run AI over their entire photo collection, and more ‘realistic’ use cases!

Just don’t ask about ship dates, prices, yields, or the important bits, and really don’t ask about the specs on the ‘tests’ they presented.

[Magik Mark]

Thank You, guys, CRU 1.4.1 with DisplayID also solved my above mentioned issue. That's what I did:
- wrote down the madvr timings, removed the madvr created custom resolution, then restart
- in CRU 1.4.1
-- add DisplayID to Extension blocks and set it as the 1st entry
-- set madvr timings and set Pixel clock and Native as well

It's a big help to get rid of a major annoyance, now I can set 23p as the default refresh rate of the TV :)

Couple of notes:

The difference between CRU and nvidia created (by madvr) custom timing is:
- nvidia adds a new one to the already existing ones: that can confuse applications which one to use (e.g. madvr)
- CRU replaces an existing one with the custom one on the OS level, so there won't be any confusion

When you create custom timings with madvr (http://madvr.com/crt/CustomResTutorial.html):
- always start from the default one (EDID) and pay special attention to the "sync" attributes (e.g. + , +):
-- don't select a mode that is different!
- don't deviate a lot from the default EDID values
-- remember, this approach is modifying "back porch" (horizontal and vertical values) only!
- the above 2 wrong settings can easily modify the response of the panel! (e.g. modifying gamma curve, causing black crush, etc.)
- if you get >= 10 hours after the first optimisation then it's already really good, you don't need to do more runs

Examples, original EDID timing:

EDID/CTA:
front sywi back pixels sync
hor 1276 88 296 1660 3480 5500 +
ver 8 10 72 90 2160 2250 +
picl 296.70 mhz
23.9757575757576
OSD result: 23.97791Hz (~4.41 minutes)


Result after the first optimisation run:

Custom:
front sywi back pixels sync
hor 1276 88 356 1720 3480 5560 +
ver 8 10 72 90 2160 2250 +
picl 299.94 mhz
23.9760191846523
OSD result: 23.97537Hz (~22 hours - 1 day)



About nvidia driver versions:
- some of them can block custom timings completely (doesn't matter how they were created)
- so always check in madvr OSD whether your timing is applied if you install a new driver

Btw, I still use the Display Changer II (https://forum.doom9.org/showthread.php?p=1863306#post1863306) util to easily manage 2 displays.

Have you tried using 47.952 instead of the usual 23.976. It's basically doubling the latter.

So far no success for me

[chros]

No success with which? :) I haven't tried 48p because LG TV has 23p but not 48p.

[NikosD]

A quick recap of Ampere RTX 30 series launch before the first real benchmarks on September 17th:

As I have told you from the beginning, nVidia decided to go nuke, giving the largest die of GA102 at both 3080/3090 chips.
It also went nuke by releasing both cards at >300 Watts for desktop gaming which is a record high for nVidia previously reserved only for two chips combined (like GTX 690)

The leather-jacket-man with a Duke Nukem EGO, didn't manage to get TSMC 7nm and decided to go with Samsung at 8nm.
The comparison of TSMC 7nm vs Samsung 8nm for Ampere using GA100 vs GA102 gives a 47% higher density for TSMC and better efficiency.

ASUS announced a RTX 30 series card with LEDs, flashing after a system crash indicating that you have to change PSU in order to feed properly the power hungry card.

nVidia fan boys used to blame ATi/AMD when it raised power consumption in order to be competitive with nVidia.

Now we see the opposite, but noone complains.

As I have already mentioned, Ampere's TFLOPS are "fake" meaning they advertise a maximum theoretical value not reachable in real-world.
So the first calculations give 1TF of Ampere to be around 2/3 TFLOPS of Turing aka 3080 performance of 30TFLOPS is approximately around 20TFLOPS of Turing.

That's why even nVidia says that 3080 card with 30TFLOPS is up to 2x faster than 2080 with just 10TFLOPS and not 3x.

In real-world games, in average of at least ten modern games, I expect 3080 to be around 70-80 percent faster than 2080.

Judging by MS Flight Simulator 2020 and recently released Marvel game, the 8GB VRAM of 3070 and even 10GB VRAM of 3080 could be a problem for 4K max settings even from now, not in the future.

The good thing about this release is that I didn't read anywhere the embarrassing and disgusting "Just buy it!" from Tom's Hardware two years ago with the release of Turing RTX 20 series.

Digital Foundry was paid by nVidia for a biased presentation of 3080 vs 2080 benchmarks, but this is minor.

And what about the competition ?

The Big Navi21 with 80CUs could be around 20+ TFLOPS of RDNA2 architecture which is closer and probably a little faster than Turing's TFLOPS so it could compete with 3080.

All AMD's Big Navi cards could potentially have lower power consumption than nVidia's Ampere cards for all tiers.

Also, it seems that Big Navi21 could have 20GB of VRAM competing with 3080 and Navi22 16GB of VRAM competing with 3070 card.

So, lower TDP and double VRAM for AMD with close to target performance.

But a very bad thing appears more and more in rumors.

Due to the large die of Big Navi21, they may want to reserve the quantities of TSMC's 7nm for more profitable and smaller console and CPU (Zen 2/ Zen 3) chips than GPUs that have little margins.

VERY BAD THING for GPU competition in the high segment.

We'll see...

[Magik Mark]

No success with which? :) I haven't tried 48p because LG TV has 23p but not 48p.Using CRU to make custom refresh rates in LG Oled. Is it possible for you to make experiment on this? 48fps

Sent from my SM-G988B using Tapatalk

[chros]

Using CRU to make custom refresh rates in LG Oled. Is it possible for you to make experiment on this? 48fps
The problem is, as the guide suggests, that we should start from the EDID of a given mode and only change the "back porch", but we don't have any initial values for 47p.
But why don't you use 23p on the C9?

As I have told you from the beginning, nVidia decided to go nuke, giving the largest die of GA102 at both 3080/3090 chips.
It also went nuke by releasing both cards at >300 Watts for desktop gaming which is a record high for nVidia previously reserved only for two chips combined (like GTX 690)
Good summary.

nVidia fan boys used to blame ATi/AMD when it raised power consumption in order to be competitive with nVidia.
True.

Now we see the opposite, but noone complains.
I do, not that it matters much :)

So, lower TDP and double VRAM for AMD with close to target performance.
That's what I'm interested in, but let's wait to see it happen :)

VERY BAD THING for GPU competition in the high segment.
I don't think it's bad, for me power efficiency is way more important. The only question is whether they can achieve it.

And what about DLSS (marketing BS) regarding to gaming? Am I the only one who feels like cheating? :)
I mean the need of DLSS clearly demonstrates that even these GPUs can't deliver native 4k content with reasonable fps. And they started to talk about 8k :D

Lastly, pricing: lot of folks are happy that nvidia haven't raised prices but a 3070 for $500 is already really high. They basically kept the Turing prices (previously a xx70 class GPU did cost around $400).

[huhn]

xx70 is just a number.
the problem with turing and the reason for the low sales number was you got the same performance for your money making them a terrible upgrade.

the ampare GPU are suppose to bring much more performance on the table for the same money if it is called 3070 for a 1000€ or 500€ doesn't matter what matters is that there is more performance for your money smaller cards will follow soon enough.

the problem with amd is not just the power number it's the performance per watt. that doesn't mean there is a problem if they have a card with 300w+ in there hands as long as there are smaller one with a reasonable power usage are following.

maybe ignorance reaches a new high here but NAVI is already 7nm and it can't beat turing in performance per watt. it is impossible to fix that clearly not but it doesn't change current facts.

[SamuriHL]

Couple that with the fact that we don't yet know (at least I don't anyway) what they plan on putting in for encoder/decoder support. And that matters. A lot. At least to this crowd. Traditionally they've fallen short on the encoder/decoder side of the equation. Couple that with Huhn's accurate portrayal of their performance per watt, uh, challenges....and it doesn't usually spell good things. I still maintain they're going to fall short of the 3080's performance. It might be "close enough" to make the case for it. Especially if they crank up the memory, but, will it be GDDR6X or the equivalent? At the price points they're going to have to compete at, I have a hard time believing they're using a 7nm process AND fast memory. I bet they sacrifice memory speed to give you gobs of it. 16gb at near 3080 performance, right? Let's hope that's not closer to 3070 performance, although it could be. If they could match or beat the 3080 performance with 16gb of ram they'd not be talking about cutting prices to 550....just above the 3070's price. What's that say to you about where we can expect the card to fall? All conjecture at this point. Nothing concrete until we see it from the electric bike's mouth, as it were. ;) :D

[ryrynz]

Calling it, Nvidia has the crown for a long time yet. 16GB at near 3080 performance? DOUBT. Competitive prices? Yes @ 3070 levels and below. Nvidia is not worried in the slightest about Big Navi.
This gen is a big deal for various reasons, will be a long time till you see anything like this again, new memory standard required.


VERY BAD THING for GPU competition in the high segment.


Overall prices are great, nobody is really complaining about these. Sure 3090 price is high but we're talking top of the line fastest GPU in the world, they can milk it.

Nvidia has competition, they've priced this aggressively to beat the competition.. This is a great release all round and still people complain about various minor things as always, so tiring.

[SamuriHL]

I mean I tend to agree with the assertions that they'll be slightly faster than a 3070 but not near the 3080. Which is why, IMO, the leaked price is probably accurate. But we have nothing concrete so it IS just conjecture.

[NikosD]

A few things to correct, because ignorance reached the sky and is still going up.

So, before ignorance reaches the moon I have to say:

1) GDDR6X is a custom-made memory produced by Micron only and while we don't know what will happen in the future, the only customer for now is nVidia.
So Big Navi is going to have GDDR6 or if AMD surprises once more will have HBM2.

2) There is memory faster than GDDR6X and it's called HBM2 and HBM2e from Samsung, SK Hynix and AMD

3) NAVI was based on RDNA and according to Techspot the NAVI10 (5700 XT) chip of 10.3 billion transistors with a TDP of 225 Watts has the same performance with the TU104 (2070 Super) of 13.6 billion transistors with a TDP of 215 Watts

So, extremely close I could say essentially the same performance per watt for RDNA and Turing.

4) AMD has promised 50% better performance per watt of RDNA2 versus RDNA1

5) Big Navi is based on RDNA2 architecture.

6) Do the math.

The completely ingnorant and insignificant troll who is comparimg TFLOPS with game engine cheating (Oh Jesus!) should stop trolling below my comments and read a little technical stuff before exposing himself so badly.

Just my two cents...

[NikosD]

Nvidia is not worried in the slightest about Big Navi. and Nvidia has competition, they've priced this aggressively to beat the competition This is the definition of contradictory opinions.
You can't say these two phrases in the same post.
One of them is wrong and obviously it's the first one.
Sure 3090 price is high but we're talking top of the line fastest GPU in the world, they can milk it. This was and still is one of the biggest problems regarding customer behavior against Intel and nVidia compared to AMD.

Just look at the price of the fastest CPU in the world of servers - the EPYC with 64 cores - the fastest CPU in the world of High End Desktop - the Thread ripper with 64 cores and zero competition - and the fastest CPU in the world of desktop - the Ryzen 3950X with 16 cores.

Who is milking and what is the reaction when AMD tries to put the price a little higher ?
The war of trolls.
But for Intel and nVidia is just Wednesday.

Be more responsible customers not fan boys and support the competition.