RTX3080 signal pollution demonstration by the use of Oscilloscope

[kiriakost]

GPU architecture this is a complex topic.
Dedicated scientists among with electronics engineers both trying to keep thousands of signals separated from interfering to its other.
Its game frame this requiring thousands single signals them to establish an communication between GPU and with the on-board memory.

The ultimate amount of frames this is a sum of GPU processing capability and the max frequency that this is clocked at.
The faster the frequency this will aloud more single signals them to establish communication in the unit of time.
Its single signal this carrying information which eventually this will feed a single pixel in our monitor.
In a very simplistic way this is the base theory of operation.

What is signal pollution?
GPU this is an integrated circuit, this is one City full of transistors (buildings) and of roads ( highways).
The signals are the cars moving in it.
If you wish to speed up the circulation of the signals (cars) then you need to use powerful and faster transistors, then to invest more at making better quality of roads, and also to create enough distance between two roads of opposite direction.
In the world of electronics the barricade that we use to keep separate signals which has the potential to short-circuit at open air this called creepage distance.

Signal pollution within an GPU this is possible when GPU design this failed to asses correctly (predict or measure) of the isolation straight between two signals that the technology in use (silicon parts) this is capable to offer.
When signals collusion starts to occur there is no other away than to lower down the frequency to the point that the isolation straight (walls of barricade) they are capable to deliver signal immunity from foreign interference.

Some people at you-tube with low to basic understanding they did demonstrate of how a capacitor this is capable to operate as another barricade this able to reduce further the appearance of signal pollution.
The major fact here is that GPU it self by design, this is the one to fight back internally any signal pollution at 95% and that external capacitor influence this could assist at 5~10% at reducing signal pollution when the signals are traveling at the surface of the card due copper lines ( Printed circuit Board).

Demonstration by the use of Oscilloscope
This is a basic demonstration at 50.000Hz = 50 KHz
The oscilloscope this measures the signal with extreme clarity when the trigger point this is were it should be.
But when the trigger point this changes at the wrong setting, the Oscilloscope this demonstrate a wide amount of ghost signals and this phenomenon it is equal as visual presentation to what is happening at the core of RTX3080 when balance is lost at 2GHz or about there.

Written By Kiriakos Triantafillou - Maintenance industrial electrician (active freelancer) .
Founder of ITTSB.EU Industrial - Test Tools - Score Board
(Electrical test and measurement product reviews Blog since May 2012)
Volos - Greece

This is my offering to TPU an personal contribution and in recognition that this is a valuable source of information over the years, this helped and helping many generations of people to learn more about the market of VGA cards.

 

[mouacyk]

I get that lanes should be sufficiently wide enough for intended speeds, so cars can stay in lane at all times. Still need layman English please. Are the GPU's crashing because the analog signal fluctuates too much at the higher clocks? And this leads to register and memory corruption, right?

 

[mtcn77]

Don't be hurtful guys, otherwise you'll only have pantene Steve to untangle the strands of oscilloscope signals for you.

 

[sneekypeet]

Thread heavily cleansed. If you have nothing productive to say move along or earn points, your choice. Fair warning to all!

 

[mtcn77]

@kiriakost Do it for AMD, please.

 

[MikeSnow]

@kiriakost Do it for AMD, please.

The thread title is a bit confusing. It might initially lead people to believe this was a test / measurement done on an actual RTX 3080. But the rest of the post seems to indicate it is not. Instead, it's just a test trying to produce a visualization of what he thinks it happens on the RTX 3080 when the problems occur. He's just measuring un unrelated test signal, while playing with the triggering system of the measuring equipment to artificially introduce measurement artifacts.

In conclusion, there is nothing to "do for AMD", since nothing was actually done for NVidia, it's just a theoretical / hypothetical analysis regarding potential issues.

 

[Caring1]

The thread title is a bit confusing. It might initially lead people to believe this was a test / measurement done on an actual RTX 3080. But the rest of the post seems to indicate it is not. Instead, it's just a test trying to produce a visualization of what he thinks it happens on the RTX 3080 when the problems occur. He's just measuring un unrelated test signal, while playing with the triggering system of the measuring equipment to artificially introduce measurement artifacts.

In conclusion, there is nothing to "do for AMD", since nothing was actually done for NVidia, it's just a theoretical / hypothetical analysis regarding potential issues.

Theory is the same and carries across brands.
Well done @kiriakost this is well explained for the layperson even if some failed to understand.

 

[moproblems99]

Theory is the same and carries across brands.
Well done @kiriakost this is well explained for the layperson even if some failed to understand.

I generally agree with the poster you quoted because I don't believe he actually has access to a 3080 which would make this a representation. It is beneficial in illustrating what the issue is/could be but there is no actual 'proof' this is what is happening. Therefore, it is theory only but beneficial theory only.

 

[Nike_486DX]

That is actually a universal statement that applies to any cpu / gpu / etc.
It is the reason why overclocks are less and less significant nowadays, and is why u can still do some crazy oc with LN (because at lower temps the conductivity increases).

RTX 3080 / 3090 (Ampere in general) is a new product, these issues are normal (because maybe there is a bug in tdp control, or gddr6x, or the drivers, etc) and eventually nvidia will sort them out.

ps
Really dont think that this is due to the use of cheaper filtering caps (like the video that Jayz2cents made a couple of days ago)

, because we have seen a lot of high end aftermarket cards that were really good at oc and that used these very same "cheapo" caps ... Unless Ampere is too sensitive in that aspect

 

[wolf]

I generally agree with the poster you quoted because I don't believe he actually has access to a 3080 which would make this a representation. It is beneficial in illustrating what the issue is/could be but there is no actual 'proof' this is what is happening. Therefore, it is theory only but beneficial theory only.

I tend to agree, this is an interesting subject without a doubt, and I don't mean to misrepresent that I enjoyed the content and find it informative, but the title is quite misleading.

rtx3080 signal pollution demonstration... I mean really, people are coming here thinking you have directly demonstrated signal pollution on an RTX3080.

I'd have thought something like "GPU signal pollution demonstration by the use of Oscilloscope, and how this phenomenon is believed to affect RTX3080's"

 

[kiriakost]

I get that lanes should be sufficiently wide enough for intended speeds, so cars can stay in lane at all times. Still need layman English please. Are the GPU's crashing because the analog signal fluctuates too much at the higher clocks? And this leads to register and memory corruption, right?

In electronics they are several fundamental laws them explaining of what a signal is, how it behaves under specific conditions, and the limitations.
The same issues or problems arise in to single cable with may pairs of wires and up to the most complex IC chip.
At higher clocks, we are not able to maintain signals as free of interference or crosstalk as they should be.
Such challenges they find a solution by the use of more proper design of the chip it self, but any workaround it is also dependent to lithography in use.

At 1GHz and above it gets harder to keep signals secure from them interfering to its other.
INTEL , NVIDIA and or AMD, what they do this is inventing new science and not regular electronic applications, all their designs they are pioneering steps in to a direction which its hard to predict the conciseness.

I am aware such answers are not simple to digest if you missing the necessary education.
Therefore I will simply deliver few hints, so anyone capable to explore by him self.
Link explaining crosstalk issues ......

Additionally the Oscilloscope which I did use for this demonstration, this is one of the the very few them capable to capture and measure and demonstrate Ghost signals on screen.
This were developed in this decade, this is a sample of latest developed tool which can explore pulsing power supply issues and has a price tag of 2000 Euros.
INTEL , NVIDIA and or AMD, for their development of technology they use also Oscilloscopes to capture and measure and demonstrate Ghost signals on screen up to 100GHz at the price of above 100.000 Euros.

In conclusion dear friend, the answer in your question it requires not plain words, but a special tool this worth more than the quantity of 150 RTX3090 cards in retail value.

The thread title is a bit confusing. It might initially lead people to believe this was a test / measurement done on an actual RTX 3080. But the rest of the post seems to indicate it is not. Instead, it's just a test trying to produce a visualization of what he thinks it happens on the RTX 3080 when the problems occur. He's just measuring un unrelated test signal, while playing with the triggering system of the measuring equipment to artificially introduce measurement artifacts.

There is no guess work involved, when I was in school at 1988 we did not had oscilloscope in the class of electricians.
But feel free to offer your own better analysis by the demonstration of hard evidence.
If you succeed you might get hired by NVIDIA !

 

[ratirt]

RTX 3080 / 3090 (Ampere in general) is a new product, these issues are normal (because maybe there is a bug in tdp control, or gddr6x, or the drivers, etc) and eventually nvidia will sort them out.

These issues are normal to some extent. Can the interference happen yes. Maybe Ampere is pushed to its limits and it can't go any higher. Maybe with better capacitors it could but AIB's and NV will not use much more expensive capacitors just to go 50Mhz above of what the 3080 or 3090 can do now. Not worth the effort and money. The chips are massive and it is understandable that these will not OC extremely. There's always a limit to this to what the chip's frequency will be. The more complex the circuit (chip) then the signal pollution will happen faster and clocks wont reach as high as we would want.

 

[MikeSnow]

There is no guess work involved, when I was in school at 1988 we did not had oscilloscope in the class of electricians.
But feel free to offer your own better analysis by the demonstration of hard evidence.
If you succeed you might get hired by NVIDIA !

My point is we don't know exactly what the problem is. We just have a bit of evidence that the capacitor choice has some influence on the behavior, but signal pollution, either inside or outside the GPU chip might not be the most significant factor in this issue.

Even with a simple digital circuit, consisting of a few logic gates and registers and synchronized by a clock signal, you can experience instability as the clock frequency increases. And that can happen even if the input signals are almost perfect, and there is almost no interference between the signals. Some logic gates or chains of such gates can simply be unable to change the state of their outputs fast enough, before the next clock tick, if you push the clock frequency too far.

When that happens, the clock tick can catch some gate outputs in an intermediate state, that is not clearly distinguishable as a 1 or a 0. So, registers that need to rely on that output as their input to change their state, will start to accumulate more and more random errors as the clock increases, resulting in data corruption.

Of course, any signal pollution or power noise can exacerbate such problems, pushing a circuit that is already close to its limits over the edge, by creating more randomness and degradation in the signals. In extreme cases it might even be the main cause of the errors, and in such cases lowering the frequency might not even help. But it's not mandatory to have signal pollution or noisy power to experience such problems, pushing the circuit to a frequency that is too high can be enough.

After a certain point, if you keep increasing the frequency, you'll not just get a few random errors, instead some gate outputs may become completely unreliable, and after a certain point the circuit will not be able to do any useful work.

Increasing voltage can make the gates switch faster, which allows faster clock speeds. But that comes with its own set of problems, after a certain point. Like the current jumping over gaps between components and traces and causing errors or, in extreme cases, even deteriorating or destroying the components. And, of course, even when things are stable, increased voltage means increased power usage and temperature. And if you happen to have a power supply that can't handle all that, you have an additional potential source of instability.

Anyway, we can keep discussing this for a long time, but my point is that you didn't demonstrate signal pollution on the RTX 3080, at best you demonstrated signal pollution in general. Starting the name of the thread with RTX3080 feels a bit like "clickbait", which made me feel "clicker's remorse".

 

[StaticVapour]

Am i the only one who likes the look of that ghost signal

 

[mtcn77]

Am i the only one who likes the look of that ghost signal

Winamp.

I still don't get it. @kiriakost is there a temporal component involved which entails which part of the hardware is wrong? Like, if events <1ms, or >1ms are causing the issue therefore the suspected problem would lead us to a component, or are all units affected? It seems they are doing quite well if only overclocking margin is affected. I thought transient response were affected to the contrary.

 

[wolf]

Anyway, we can keep discussing this for a long time, but my point is that you didn't demonstrate signal pollution on the RTX 3080, at best you demonstrated signal pollution in general. Starting the name of the thread with RTX3080 feels a bit like "clickbait", which made me feel "clicker's remorse".

Again absolutely correct, I came here because the title directly led me to believe this was tested on an RTX3080 and I too have clickers remorse. Ahh well, at least he feels good that he has taught people something...

 

[kiriakost]

I still don't get it. @kiriakost is there a temporal component involved which entails which part of the hardware is wrong? Like, if events <1ms, or >1ms are causing the issue therefore the suspected problem would lead us to a component, or are all units affected? It seems they are doing quite well if only overclocking margin is affected. I thought transient response were affected to the contrary.

I made clear two times so far, that any well defined answers they can come due the screen of one latest developed Oscilloscope this cost thousands of Euro dollars and only NVIDIA and large corporations they are able to afford it.

Now due the fact that NVIDIA R&D department this is VAULT , and strangers are not aloud to visit or see it.
I thought to give you a glimpse of what NVIDIA engineers will also see in their more expensive also latest developed Oscilloscope.
The ability of one Oscilloscope to review Ghost signals (that they are the first step of appearing interference) this is a new one as tool, since the year 2012 and later.

NVIDIA's technology this is patented and a legally protected property.
My presentation this is generic example and a safe one regarding any legal actions against my self and or TPU in which this content it is hosted at.

Thank you.

My point is we don't know exactly what the problem is. We just have a bit of evidence that the capacitor choice has some influence on the behavior, but signal pollution, either inside or outside the GPU chip might not be the most significant factor in this issue.

Even with a simple digital circuit, consisting of a few logic gates and registers and synchronized by a clock signal, you can experience instability as the clock frequency increases. And that can happen even if the input signals are almost perfect, and there is almost no interference between the signals. Some logic gates or chains of such gates can simply be unable to change the state of their outputs fast enough, before the next clock tick, if you push the clock frequency too far.

There is no simple digital circuit here.
The wavelength above 2GHz this is challenging the materials in use over that card.
The factory is the Chip.
The less important helpers are the capacitors.