New Nvidia Cards

[The ghost of raids past]

The 3070 looks like a bargin 500 quid for a card more powerful than a 1k plus 2080 TI.

The 2nd hand market for 20 series cards has collapsed. It looks from the launch demos like a very impressive beast. It will be interesting to see if Team Red can come up with anything to match it.

Dio knows but if he told us he would be Executed !

[venusian]

Slightly miffed as I bought a 2070S very recently on the assumption that all the new cards would be massively expensive and not much faster.

Saved myself have to buy a new PSU though.

[gilesgoat]

I would like to hear "something from Dio" about "what he thinks of such cards".

I am NOT convinced of that DLSS ..

Then I read about this "Tensor cores" :

Tensor Cores are able to multiply two fp16 matrices 4x4 and add the multiplication product fp32 matrix (size: 4x4) to accumulator (that is also fp32 4x4 matrix). In turn, being a mixed accumulation process because the accumulator takes fp16 and returns fp32.

So it can do : float Result[4][4] = float A[4][4]*float B[4][4] + float C[4][4] while keeping "intermediate results" of double temp[4][4] ?

Actually no, it does : double Result[4][4] = double A[4][4]*double B[4][4] + double C[4][4] while keeping "intermediate results" of float temp[4][4] ?

Presumably "in one something cycle" ?

Are they really "as good as they say" or "there's some trick somewhere" ?

Also .. what AMD has to offer "better" or similar to this ?

[KlaxonCow]

Grabbing a screenshot of this nice graphic that Linus Tech Tips made to illustrate the transistor counts across nVidia architectures:



Basically, they've switched from TSMC to Samsung for the fabrication, and dropped to an 8nm process, which packs in so many more transistors. So these things are packing more than twice as many transistors as the last gen ("Turing").

So, yeah, with 28 billion transistors going spare here, why not splurge on a dedicated single-clock "double Result[4][4] = (double)A[4][4] * (double)B[4][4] + (double)C[4][4];" instruction for your Tensor cores?

As well as just the fact that these cards now have double the Cuda cores of the previous gen. So this is like taking two Turing cards and squeezing them into one card, yet it takes up about the same die space as the previous architecture (and, presumably, Samsung's 8nm fabrication just has better yields, judging by the prices).

I also think that the Linus Tech Tips take on why this is happening is probably about right too.

Basically, AMD own the next gen of consoles. nVidia doesn't have a look in there at all. So their play here is to make PC graphics - and, thus, PC gaming - so compelling to tempt folks to stay with upgrading their PCs, rather than going to the next gen consoles. And to drop this first - before the next gen consoles come out - to try to grab that market before AMD / Sony / Microsoft have launched those next gen consoles.

(One thing that speaks to this is that these new cards also interestingly have an answer to that impressive "the GPU can load and decompress stuff off the SSD directly, without CPU assistance" thing that Sony were showing off for the PS5. Although Sony were probably over-hyping the notion of streaming assets directly off the SSD, as the player moves around - in the sense that someone can move their mouse to do a 180 turn to look at a big bustling cityscape behind them and one thinks you really can't stream that off SSD alone in practice, no matter how silly fast you make the GPU loading and decompression be, so there has to be practical limits to how much you can genuinely do in an actual real-world game - we now have both nVidia and AMD providing this capability in console and PC space. And though I can't imagine you could have it stream in an unlimited number of assets in real-time - because, you know, logic dictates - these guys could probably actually do it for real in a limited scene with limited assets - e.g. an indoor Quake-like corridor scene, where there's an upper limit on how many assets can be seen at one time.)

But, yeah, when the console guys are going "look at this amazing new ability!" and nVidia's response is "yeah, we've got one of those in our cards too", it's a direct salvo at AMD having the next gen consoles in the bag already.

And, in fairness to nVidia, it's the play to make. What else can they do? Tempt folks to stick with PCs, before the next gen consoles arrive, with a big generational leap to compel folks in and keep the prices as "console-equivalent" cheap as possible, so that it leaves people with the thought "I could buy a 3080 or a new console?", with the hope that they choose the 3080 instead of a PS5. Because, basically, if they let the next gen consoles do well, without a response, then they're locked out for a generation, as it's AMD that owns that space right now.

It's the move that, essentially, they have to make competition-wise. And, as always, heavy competition suits the consumer just fine - as be it a new PC graphics card or a next gen console, you're going to get something deeply impressive for your cash, either way. Yes, when it comes to stuff like this, capitalism can and does work.

Also, we should, as always, reserve our excitement until RDNA2 is also out. I mean, this salvo from nVidia is to meet what we've heard the next gen consoles can do with an AMD. If what's in those consoles comes back to PCs with Navi, then AMD does have an answer to this already, it seems. It's just not in the PC space right now. And we know AMD can fabricate on such tiny scales, because that's how they're making Intel look big and stupid at the moment in the CPU space.

In a sense, this is why nVidia are making these moves. AMD owns the CPU space and the console space. Like, totally own it. Intel is way behind and nVidia have been locked out of this gen of new consoles. So nVidia has to push hard on the dedicated PC graphics card here - to push their advantage - to stop AMD, with Navi, potentially stealing that space too. nVidia can't afford that, so they've had to make a very big play here and have timed it to grab the market now, before AMD launches their response.

[gravy]

I would like to hear "something from Dio" about "what he thinks of such cards".

I am NOT convinced of that DLSS ..

DLSS 2.0 is pretty amazing

[Dio]

I'm not convinced that DLSS is better than good algorithmic solutions.

Can't talk much about the rest. They do appear to be being very aggressive on pricing - the 3090 price looks about as expected, but there doesn't look much room for profit in the 3070/80.

[KlaxonCow]

I'm not convinced that DLSS is better than good algorithmic solutions.

Agreed.

I am also impressed with what DLSS 2.0 can do, but I'm mindful that it is, somewhat by definition, a "stop gap" technology.

That is, current tech can't render all that at 4K? Never mind, render it at 1080p and then let the AI fill in the information gaps. Great, but the inherent thing about that is, well, in the intervening 5 years, current tech can now actually render all that at 4K now. So why fill in the gaps with an AI "best guess", when you can just have no gaps in the first place?

Okay, at that point, you might say "ah, but current tech is still not good enough to do this at 8K". But that is, of course, just kicking the can further down the road and, eventually, we will catch up to where we kicked the can.

Also, surely we're headed to real-time ray-tracing. Indeed, particularly when these new cards - and AMD's offerings - seem to make that even more realistic a proposition to eventually become the norm. The 2000 series were adding a bit of ray-tracing here and there - testing out the concept with baby steps - but the 3000 series (and AMD equivalent) seems to have the capacity for a game or two perhaps deciding to dump rasterisation completely.

And ray-tracing works fundamentally differently from rasterisation that I foresee completely different algorithms and ideas being thrown up by that to deal with these issues by alternative methods.

Such as, for example, cone tracing. See Farbrausch's 64KB demo at 8:40 here:



(Whilst Conspiracy's demo - at 29:30 - is artistically much more impressive, the Farbrausch demo is technically more significant, but doesn't look so good. Ah, the dangers of being an "early adopter" of technology.)

Instead of throwing rays, you're throwing cones. Volumes and areas, instead of lines and points. This inherently has scope for implicitly handling anti-aliasing and upscaling algorithmically, as part of how the tracing itself happens.

But, yeah, as this stuff is not properly here yet, then DLSS is a good "stop gap" - but, with one eye on the future, I can't get so enthused by something that, one day soon, will almost certainly be superseded by way better stuff, as we all move into completely different rendering methods.

The hardware / software will get good enough to do 100% real-time ray-tracing of the whole thing. And then it's "game over" for rasterisation, really.

[gilesgoat]

I'm not convinced that DLSS is better than good algorithmic solutions.

Can't talk much about the rest. They do appear to be being very aggressive on pricing - the 3090 price looks about as expected, but there doesn't look much room for profit in the 3070/80.

Do they think they get at least a 10% margin ? Which means whoever sell those GPUs gets .. 1% maybe ?

I suppose all the "medium/low GPU range" is sold at near loss then ?

Turn on into another question then, if I should buy a "performant GPU that works well for many things" ( graphics and possibly some HPC ) what should I get ? AMD or NVIDIA and what AMD ?

Also not that cares me but what my bro does, I've seen some "graphs" ( that could mean all or nothing at all ) but why performance wise it would appear that using OpenCL vs CUDA make a SIGNIFICANT HIT in peformance ? It would appear that CUDA goes lot faster ? It's the implementation of OpenCL "so crap" ?

[gilesgoat]

*SNIP*
Instead of throwing rays, you're throwing cones. Volumes and areas, instead of lines and points. This inherently has scope for implicitly handling anti-aliasing and upscaling algorithmically, as part of how the tracing itself happens.

If you would have read certain ACM Transactions on Graphics or IEEE articles of MANY years ago stuff like "distributed ray tracing" "cones" and such like that been discovered quite a while ago, also I think POV Ray for AMIGA ( https://en.wikipedia.org/wiki/POV-Ray ) was one of the first capable to implement such techniques.

I think also there were things like "adaptive" ray tracing and more complex issues about "space subdivision" , octrees, quadtrees and stuff like that.

We are talking 199X if not 198X at least.

What I still never seen implemented anywhere was "Radiosity" ( https://en.wikipedia.org/wiki/Radiosity ... _graphics) ) I am not sure if PRT ( that thing with SPH compression - https://en.wikipedia.org/wiki/Precomput ... e_Transfer ) helps to get "similar results" or if it but as far as I know I don't know any SW implementing it.

[edit] so WHEN this in REAL TIME ( 60 fps at 4K ? )



.. actually 2X that because we need it in VR

[KlaxonCow]

They do appear to be being very aggressive on pricing - the 3090 price looks about as expected, but there doesn't look much room for profit in the 3070/80.

Yeah, that's the other reason for presuming the interpretation that this is nVidia trying to "take ownership" of PC graphics.

And they're compromising on profits to do so - because they feel they have to, as they've seen what AMD has done to Intel in the CPU space, and can't let something like that happen in the GPU space too. It would spell disaster, so they're willing to take a knock to profits to try their best to make sure it doesn't happen.

In a sense, it's an odd kind of compliment to AMD. You've got them worried enough that they feel they've got to bring their "A game" to the table - and price it super-cheap, so adoption goes through the roof, they hope - so that, basically, AMD doesn't hand them their arses on a silver platter, like Intel are currently facing on the CPU side of things (Intel are paying a big price for failing to get their fabrication processes down from 14nm).

They've read the room. They have to make a big play here - and even accept shaving off profits to do so - in order to keep themselves in the game.

And, as I say, that actually implicitly says good things about AMD, in that you've scared them. Indeed, as you suggest, the 3090 price seems reasonable, but are the 3070 / 3080 prices essentially foregoing any useful profit, in order to try to "take ownership" of PC graphics?

Kind of like how consoles are often sold "at cost" (or even potentially less than cost) because you need to shift as many units as possible - and then you get your profits from the games and stuff instead.

That maybe the 3090 - aimed at "the professional" - is subsidising the gamer's 3070 / 3080, just because nVidia need to make a play on that space, before AMD enters the room. Because AMD have CPUs right now, and they also have consoles. nVidia can't afford to let AMD also take the GPU space, or you guys would have the full house here.

[venusian]

nVidia is doing a pretty good job of being the default choice for autonomous robots (in the broadest sense, including AVs). That's a HUGE market for specialised compute where AMD are absent and Intel are weak. It's the most rational explanation behind their otherwise unfathomable rumoured purchase of Arm - they don't need or want Arm, but it will destabilise their competitors in edge compute.

The margins there aren't down at MCU levels either - just a Jetson modules is hundreds of dollars.

[KlaxonCow]

If you would have read certain ACM Transactions on Graphics or IEEE articles of MANY years ago stuff like "distributed ray tracing" "cones" and such like that been discovered quite a while ago, also I think POV Ray for AMIGA ( https://en.wikipedia.org/wiki/POV-Ray ) was one of the first capable to implement such techniques.

Oh, yeah, totally. None of this is new.

Ray tracing is "ancient technology" in the theoretical / offline rendering space. But, of course, the difference here is talking about hardware that does it all in real-time.

That's the bit that's new - having hardware that's got the grunt (and the acceleration structures) to actually pull it off at interactive frame rates.

[gilesgoat]

I would like to know MORE DETAILS about GPUs in general, there's really "TOO MUCH HYPE" on the products and really I HAVE NO FUCKING IDEA WHAT I AM GOING TO BUY !

Is that really "the best choice I could have made given a budget" ? Is that "the good thing to buy" or there maybe was something better ?

Should I buy NVIDIA *OR* AMD ? WHAT ARE the SIGNIFICANT differences between one and the other ? What do I get from one side I don't get from the other ?

Well in terms of PCs "by NOSE and what I can see", for 20 PCs "grab it and bring it home" 19 mounts NVIDIA, one has maybe AMD, in that sense I think NVIDIA already dominates at least the "cheap to medium" PC market.

I don't know much about "super high end PC gaming" where "by nose and by hear" I hear more frequently "Ryzen" and "AMD" said that "Ryzen" is NOT "THAT CHEAP" as it was initially "promoted" ? ( i.e. last time I checked Ryzen is quite expensive ) yeah it may be given me "something more per pound" ( I don't know really ) but it's still NOT cheap.

So assuming you want to remain in the domain of "A thousand and something but not too much more" pounds PC .. where do you get ?

Or assuming you want to sacrifice a bit of something else "going all for the graphics" .. what do you buy ?

That "the world is still stuck at 1920x1080" .. I HOPE that is NOT TRUE !

4K "at least" should be the thing to go because there's this thing called VR that "at very MINIMUM" requires 2X 1920x1080 and 90HZ, minimum.

Anyway as a BUYER "with some money" and as a devloper too I think "there is NOT enough information around" or if there's you have to get literally MAD to collect it all, compare it and try to understand something to make "a really informed choice".

It gets the point "the best" you can say is "this is the max money I can spend, I throw it there I HOPE I hit it right".

I hope at least AMD/NVIDIA will put up "documentation and code samples enough" to show "how to make those features work" because they put up "whatver new shit markets more" in the video card and then "gamers" EXPECT NEXT DAY as the GPU hit the shelves "oh that game MUST SUPPORT ALL THOSE NEW FEATURES" like if it'd do "with a click of a button" and we know IT DOES NOT !

And then they moan "I just bought this new 7800 Special Edition I paid 12,000 pounds for it and the game does not support the ultra features" and they are partially right.

Mah .. anyway "as a buyer" I'd love "more details" about all this new hype.

[Misfit]

And, as I say, that actually implicitly says good things about AMD, in that you've scared them. Indeed, as you suggest, the 3090 price seems reasonable, but are the 3070 / 3080 prices essentially foregoing any useful profit, in order to try to "take ownership" of PC graphics?

I think there was also some suggestion form Linus that they expect yields to be increased this time around, so whilst they will be more expensive to produce they won't be wasting as many and thus could keep costs similar.

[Deckard]

I have never bought a graphics card in my life - Is that odd ?