Tuesday, August 31st 2021

AMD Reportedly Readying RX 6900 XTX, Bringing the Battle to NVIDIA RTX 3090

Graphics cards may be on their way to becoming unicorns that you can only pay for after finding the proverbial pot of gold from under a rainbow, but that doesn't mean AMD and NVIDIA will slow down their competition any time soon - especially in this market, there's a huge profit to be made. And AMD may just be finally readying their true halo product - a graphics card that aims to beat NVIDIA's RTX 3090 across the board. Twitter user CyberPunkCat shared an alleged AMD slide showcasing a new, overpowered RX 6900 XTX graphics card. AMD's naming scheme for their RX 6900 series may be slightly confusing nowadays: the original RX 6900 XT carries the Navi 21 XTX die, and AMD has recently released a higher-performance version of that Navi 21 chip in the form of the Navi 21 XTXH - which power the liquid-cooled versions of the RX 6900 XT, with higher overall clocks than the original GPU release. However, there hasn't been a change in the RX 6900 XT nomenclature - but this new slide suggests otherwise.

If the leaked slide is real (keep your NaCl ready, as always), it appears that the RX 6900 XTX might pair both the higher-performance Navi 21 XTXH chip with higher memory speeds. While both Navi 21 XT and Navi 21 XTXH both make use of 16 Gbps GDDR6 memory, the slide indicates that the RX 6900 XTX will feature 18 Gbps memory speeds, exploring another avenue for increased performance. This decision would bring an increase in maximum theoretical memory subsystem bandwidth from the 512 Gbps in the RX 6900 XT up to 576 Gbps - a 13% increase, which would not translate into a proportional increase in final performance. However, considering how our own reviews show that AMD's RX 6900 XT with the Navi 21 XTXH silicon is already between one and three percent faster than NVIDIA's RTX 3090, even a slight, 5% performance increase over that cards' performance means that AMD might be able to claim the performance crown for the mainstream market. It's been a while since that happened, hasn't it?
Sources: CyberPunkCat @ Twitter, via Tom's Hardware
Add your own comment

107 Comments on AMD Reportedly Readying RX 6900 XTX, Bringing the Battle to NVIDIA RTX 3090

#76
wolf
Performance Enthusiast
ratirtHow about a different approach.
Cyberpunk 2077 RT on DLSS on at 4k with 3090 high settings below 60 FPS.
ratirtNow my question again is. Is the hardware for RT sufficient for games now, considering CB2077 is not fully ray traced?
So your personal measure of success for hardware or game implementation is CP2077, a game that even in rast only cards perform below where you'd expect, or fully RT games? that was never my argument, and I've already agreed RT hardware capability needs to and will increase dramatically over coming generations.
ratirtNow tell me, when people see this happening on a card for $3k they would still hold your side?
current market conditions aside, not only is it a $1500 card, but a card with ~90% of it's performance has an MSRP of $700, so you don't need to spend 3k to get good RT perf, and again, I said RT performance generally wasn't a major factor in my choice, but it has sufficient grunt in the here and now for most of todays games with RT.
ratirtIf you expect me to find a link or a proof for 'MAJORITY OF PEOPLE' saying it, is as I've expressed it to be then you are mistaken. There is no link to majority of people saying this.
When you say it like it's a fact, I'd have assumed that notion comes from somewhere other than your opinion.
ratirtSo here is my question for you, since you don't agree with my statement 'majority of people say RT is good but the hardware lacks', is the hardware we currently have enough for fully ray traced games? Because the games will get more demanding for sure on the RT side and rasterization side.
Your measure of success, not mine. I'm saying a good hardware pairing can make for great experiences today.
ratirtAnd one more question, how do you want me to give you a link to 'majority of people say'? I think the conclusion is the one you seek not a link or a source to 'majority of people say'.
Cite a source? a TPU frontpage poll, another tech community poll or some sort, I dunno, something more than how you feel, packaged as how most people feel.
Vayra86I think its much more realistic to look at economic realities and actual content, both of which aren't rosy for RT.
Such a shame it's here to stay and a major part of the future of real time graphics eh. Both of those realities will catch up.
Posted on Reply
#77
ratirt
wolfSuch a shame it's here to stay and a major part of the future of real time graphics eh. Both of those realities will catch up.
Will catch up but not for current hardware.
wolfWhen you say it like it's a fact, I'd have assumed that notion comes from somewhere other than your opinion.
Isn't it a fact? What is your fact that you can support with source?
wolfSo your personal measure of success for hardware or game implementation is CP2077, a game that even in rast only cards perform below where you'd expect, or fully RT games? that was never my argument, and I've already agreed RT hardware capability needs to and will increase dramatically over coming generations.
No it isn't it is an example do you have better? for let's say 3060 or 3070? or are we going to focus on a $3k card? MSRP is meaningless at this point in time. I'm sure you know that right?
wolfcurrent market conditions aside, not only is it a $1500 card, but a card with ~90% of it's performance has an MSRP of $700, so you don't need to spend 3k to get good RT perf, and again, I said RT performance generally wasn't a major factor in my choice, but it has sufficient grunt in the here and now for most of todays games with RT.
Why current market conditions aside? You want me to drop the market conditions because it is not going up your alley but you bum me out due I chose CB2077 as an example for insufficient hardware for RT? what about other cards like 2070 or 3070 and below? do those not matter as well?
wolfYour measure of success, not mine. I'm saying a good hardware pairing can make for great experiences today.
yes and it doesnt have to be RT for a great experience. especially if it doesnt bring much to the game really.
wolfCite a source? a TPU frontpage poll, another tech community poll or some sort, I dunno, something more than how you feel, packaged as how most people feel.
Do you have a link or a source that contradicts what I said? Just curious.
Posted on Reply
#78
nguyen
ratirtDo you have a link or a source that contradicts what I said? Just curious.
HUB poll


So you would belong to the minority (28%) who don't care about RT, yet speak the loudest?
Posted on Reply
#79
Vayra86
wolfSuch a shame it's here to stay and a major part of the future of real time graphics eh. Both of those realities will catch up.
Not at all! See, I'm not opposed to the technology or the advancement... I'm opposed to the way it is being rolled out, the timing of it, and the effect on GPU pricing and quality.

It feels like one step forward, two steps back. I'll happily jump in on it when the market would deal with it in healthy way, but it never did and is not looking to anytime soon. And in that, RT(X) is different from previous big changes in GPU hardware and capabilities. Today its a tech for the happy few - handful even - because GPUs in general are unobtanium.
nguyenHUB poll


So you would belong to the minority (28%) who don't care about RT, yet speak the loudest?
Ahem, selective reading much? A VAST majority actually cares about raster BEFORE RT. And that is exactly the choice you have in the market right now, too. 51 + 28%. Not 28%.

I also believe that aligns well with what most topics on the subject contain. A few happy campers being all over RT, and a vast majority just waiting it out while using raster only, but looking forward to see where it goes next.

The clear minority is those who advocate RT progress before or even equal to raster progress: 13 + 5 + 3%. This poll simply contains five degrees of 'how much you can care' about one compared to the other, weighing them all.

You know what would kick RT adoption up in a major way? Nvidia making sure an RT capable GPU lands in the hands of every gamer. Now, look at the market again ;)
Posted on Reply
#80
nguyen
Vayra86Ahem, selective reading much? A VAST majority actually cares about raster BEFORE RT. And that is exactly the choice you have in the market right now, too. 51 + 28%. Not 28%.

I also believe that aligns well with what most topics on the subject contain. A few happy campers being all over RT, and a vast majority just waiting it out while using raster only, but looking forward to see where it goes next.

The clear minority is those who advocate RT progress before or even equal to raster progress: 13 + 5 + 3%. This poll simply contains five degrees of 'how much you can care' about one compared to the other, weighing them all.

You know what would kick RT adoption up in a major way? Nvidia making sure an RT capable GPU lands in the hands of every gamer. Now, look at the market again ;)
LMAO, so between 2 GPU with equal Rasterization performance and cost per frame (rasterized), yet one is superior in RT than the other, which do you think people are buying? would that make 72% of people picking Ampere and the other 28% picking RDNA2? This is without factoring in DLSS, which skewer buyers towards Ampere much further.

You sure know the market so tightly that you would deny any market study made by professional so I wouldn't bother sourcing any :roll:.

Btw Intel is also dead set on Ray Tracing and XeSS, which will make Ray Tracing more accessible in the future
wccftech.com/intel-xess-interview-karthik-vaidyanathan/
Posted on Reply
#81
ratirt
nguyenBtw Intel is also dead set on Ray Tracing and XeSS, which will make Ray Tracing more accessible in the future
Intel is not dead set on RT but they can't afford to be behind even if majority doesn't really care because that would have been a bad marketing and no company will go that route.
nguyenYou sure know the market so tightly that you would deny any market study made by professional so I wouldn't bother sourcing any :roll:.
So what is the market study according professionals and who are they?
Posted on Reply
#82
nguyen
ratirtIntel is not dead set on RT but they can't afford to be behind even if majority doesn't really care because that would have been a bad marketing and no company will go that route.

So what is the market study according professionals and who are they?
Huh, why would Intel be focus on something the majority of people wouldn't care? (is 28% a majority? :roll: ) dedicated RT cores cost die size you know.

Jon Peddie research? Nvidia sure sold a hell lots of GPU in Q1/Q2 compare to AMD, and Steam Hardware Survey do counting them Ampere users nicely, but I guess someone here doesn't trust HWS all that much.
Posted on Reply
#83
londiste
nguyenIf I were to make an educated guess, RTX4000 will offer the same improvement to both Rasterization and RT compare to RTX3000, meaning the perf cost with RT ON will remain relatively constant between Turing, Ampere and Ada.
To be fair, the improvement in RT cores from Turing to Ampere was quite minimal in performance. They were mostly just reorganizing and moving things around.
Posted on Reply
#84
nguyen
londisteTo be fair, the improvement in RT cores from Turing to Ampere was quite minimal in performance. They were mostly just reorganizing and moving things around.
It's a trade off between Rasterization vs RT, if Nvidia add more RT cores it would take away available space for CUDA cores. The current DXR is hybrid Rasterization and RT, so reducing rasterization and boosting RT would just come to the same overall FPS number
Posted on Reply
#85
wolf
Performance Enthusiast
ratirtWill catch up but not for current hardware.
Sure, current hardware when compared to 1,2 ...3 generations forward will be far less performant at RT, I still agree, but you get to ride that wave, which has already started.
ratirtIsn't it a fact? What is your fact that you can support with source?
Dude I was curious, sources do also exist for these sort of things, one just got posted, more on that later. If you want to talk perceptions I'll gladly give you mine but I wont claim that my experiences of what people say and like is representative of a majority, by definition it's far too small a sample size.
ratirtNo it isn't it is an example do you have better? for let's say 3060 or 3070? or are we going to focus on a $3k card? MSRP is meaningless at this point in time. I'm sure you know that right.
Why current market conditions aside? You want me to drop the market conditions because it is not going up your alley but you bum me out due I chose CB2077 as an example for insufficient hardware for RT? what about other cards like 2070 or 3070 and below? do those not matter as well?
Most of those cards (amd and Nvidia) can be bought at or close to MSRP (caveat being some manufacturers MSRP's are what they are, you'll pay more for a ROG STRIX than you will a TUF for example) if you are determined enough, but sure, the market is what it is. Again, I'm not claiming a majority of people are buying RTX or even RT capable cards primarily for RT games, but when it's a strong secondary or tertiary factor, that's nice too, as can be other features or talking points, like potentially more VRAM, power efficiency, software suite etc.
ratirtyes and it doesnt have to be RT for a great experience. especially if it doesnt bring much to the game really.
That's cool too, It's not essential but it's fun and fascinating.
ratirtDo you have a link or a source that contradicts what I said? Just curious.
The one that got posted by @nguyen is interesting, in the way that we can already see in these comments there are multiple interpretations of the data.

My take would be ~20% put a fair bit of stock in it for a current purchase, another 50% of people are still regarding it as a meaningful feature. Being that it is the future, this notion is only going to increase.
Posted on Reply
#86
Prima.Vera
Miss the days when the top dogs were actually affordable and easy to get. This is now just like watching a battle between a Ferrari and a Lambo.
Interesting, but that's about it.
Posted on Reply
#87
ratirt
wolfThe one that got posted by @nguyen is interesting, in the way that we can already see in these comments there are multiple interpretations of the data.

My take would be ~20% put a fair bit of stock in it for a current purchase, another 50% of people are still regarding it as a meaningful feature. Being that it is the future, this notion is only going to increase.
To be fair, what our colleague here posted is not what this is all about. It is not about the opinion of people who think RT is currently ok or would rather stay with rasterization.
He's completely disregarded the fact that the hardware is barely capable of RT and you need a piles of money to make it happen in a respectable way. The opinion for people willing to invest in RT now or not is clearly moving leaning towards rasterization anyway. RT is great and it will be the future but it wont happen tomorrow considering hardware advancement keeping in mind what Ray Tracing can really do.
wolfThat's cool too, It's not essential but it's fun and fascinating.
Sure it may be fascinating but it still is long down the road for me.
wolfMost of those cards (amd and Nvidia) can be bought at or close to MSRP (caveat being some manufacturers MSRP's are what they are, you'll pay more for a ROG STRIX than you will a TUF for example) if you are determined enough, but sure, the market is what it is. Again, I'm not claiming a majority of people are buying RTX or even RT capable cards primarily for RT games, but when it's a strong secondary or tertiary factor, that's nice too, as can be other features or talking points, like potentially more VRAM, power efficiency, software suite etc.
Is it possible really to get a card for MSRP? I know the 2000 series cards cost more than they did during launch.
wolfSure, current hardware when compared to 1,2 ...3 generations forward will be far less performant at RT, I still agree, but you get to ride that wave, which has already started.
Actually you dont have to. If it doesn't bring much to the game and causes a huge impact in performance you can simply wait because the money you have to put in for it and the performance you get is not that good.
wolfDude I was curious, sources do also exist for these sort of things, one just got posted, more on that later. If you want to talk perceptions I'll gladly give you mine but I wont claim that my experiences of what people say and like is representative of a majority, by definition it's far too small a sample size.
Which source to support your point of view have you provided yourself? Our colleagues post is pointless in our argument.
Posted on Reply
#88
Viilutaja
nguyenHUB poll


So you would belong to the minority (28%) who don't care about RT, yet speak the loudest?
This poll shows that you and other RT diehards are 8%!!! No compromises gaming- like here as been said.
Then there are 13% of who think that 50-50. If we even add that to RT is a MUST company- then you get 21% of the poll.
79% do not CARE about RT!

This what I can call majority. I have seen similar percent's give or take on other techtuber polls.

If we can play RT games without any performance loss (60fps min with all maxed out) and without the need of FSR, DLSS, XeSS etc. Then I can say it is not a gimmick.
Right now everytime I play and try to enjoy there is weird shimmering when using DLSS. I noticed it in Metro EE and CP2077 and in Call Of Duty Warzone. No bueno.
Unless the DLSS improves this, i will not be using it and that automatically means I cannot use RT on higher resolutions than 1440p and not use the DLSS.
1080P and RT turned ON - it's fine with most RT capabale cards.
Posted on Reply
#89
ratirt
nguyenIt's a trade off between Rasterization vs RT, if Nvidia add more RT cores it would take away available space for CUDA cores. The current DXR is hybrid Rasterization and RT, so reducing rasterization and boosting RT would just come to the same overall FPS number
And that my friend is clearly showing Ray Tracing will have hardware problems to make it run good enough on all hardware in all given scenarios. Not to mention the advancement in Ray Tracing in games is not helping either if you have to make trade-offs, you bet on a horse and will must support it with the marketing. That is how I see it today.
Posted on Reply
#90
londiste
This is getting really off topic.
ratirtAnd that my friend is clearly showing Ray Tracing will have hardware problems to make it run good enough on all hardware in all given scenarios. Not to mention the advancement in Ray Tracing in games is not helping either if you have to make trade-offs, you bet on a horse and will must support it with the marketing. That is how I see it today.
This is a classic chicken and egg problem. Developing RT effects does not make sense if there is no hardware to support it. Adding hardware to support RT does not make sense if there is no RT effects in software. RT Cores take up ~3% on Turing dies. Relatively small investment for a new capability giving time for the industry to adjust and use/showoff it in a still meaningful way. Scaling them up should not be difficult but there is the same question of cost vs benefit.

In addition to optimizing the very specifically ray tracing side of things, RT advancements in games and software are very much around all the supporting areas that are not hardware accelerated. Notably, building and updating the BVH and other data structures.

Game development is a multi-year process and largely only this year have the major game engines implemented RT effects in a real (and non-beta) way. APIs have somewhat matured in their first incarnation.
Similarly, with AMD having hardware support for RT and Intel also getting some hardware support for RT, this is already getting industry-wide.

Clamoring for full path-tracing is one possible viewpoint but dismissing entire RT until that is very short-sighted IMO. RT has obvious benefits for use cases it is already being used - basically everything to do with lighting but in practice mostly GI, shadows, some aspects of transparency. Reflections too but that probably has more to do with us not having found better ways to do proper reflections than current strengths of RT.
Posted on Reply
#91
nguyen
ViilutajaThis poll shows that you and other RT diehards are 8%!!! No compromises gaming- like here as been said.
Then there are 13% of who think that 50-50. If we even add that to RT is a MUST company- then you get 21% of the poll.
79% do not CARE about RT!

This what I can call majority. I have seen similar percent's give or take on other techtuber polls.

If we can play RT games without any performance loss (60fps min with all maxed out) and without the need of FSR, DLSS, XeSS etc. Then I can say it is not a gimmick.
Right now everytime I play and try to enjoy there is weird shimmering when using DLSS. I noticed it in Metro EE and CP2077 and in Call Of Duty Warzone. No bueno.
Unless the DLSS improves this, i will not be using it and that automatically means I cannot use RT on higher resolutions than 1440p and not use the DLSS.
1080P and RT turned ON - it's fine with most RT capabale cards.
Dude, saying I don't care about RT is very different to saying I care about Raster but RT is also important, it's not that nuanced.
Idk about you but I would turn on RT Reflection in CP2077 and tweak other settings for ~60FPS (including DLSS)

Also I'm seeing less shimmering in CP2077 with DLSS as opposed to Native rendering, DLSS is just black magic really :D

As for RT, it can make the game looks like a different, more advanced version of the same game, The Ascent for example
Posted on Reply
#92
Vayra86
nguyenLMAO, so between 2 GPU with equal Rasterization performance and cost per frame (rasterized), yet one is superior in RT than the other, which do you think people are buying? would that make 72% of people picking Ampere and the other 28% picking RDNA2? This is without factoring in DLSS, which skewer buyers towards Ampere much further.

You sure know the market so tightly that you would deny any market study made by professional so I wouldn't bother sourcing any :roll:.

Btw Intel is also dead set on Ray Tracing and XeSS, which will make Ray Tracing more accessible in the future
wccftech.com/intel-xess-interview-karthik-vaidyanathan/
Don't move the goalposts if you stand corrected now, mate. That's not fair is it. You completely misread that poll, end of story, and I assume its because of tinted glasses - not necessarily bad, but just be honest about it. You are entitled to your opinion as much as we all are.

I'll be the first to admit that if RT gains traction, it gains traction. Really. But it hasn't and doesn't - and no, the games don't look like different games either. Most of them look like somebody was too busy cleaning everything meticulously. Reflections are overdone, lighting passes are generally lacking balance, but yes, they all move all the time, fully dynamic - that is true. Is it a different experience in terms of gaming? Absolutely not, it just looks like an extra post effect. In Metro Exodus, some scenes even detract from being a good experience. You linked the Ascent, the first thing that I noticed was the stuttering of the RT half of that video.

What we have now is 16-20% of GPU floor plan reserved to boost RT performance (extra cache also serves it). Those cores are working hard to produce somewhat flashier reflective surfaces and somewhat more dynamic lighting and shadow. That is all she wrote so far, and often just picking one of them - never all at once. Imagine how much more you need hardware wise to make it all tick.

Also... its good Intel is stepping in this race too. But all these generations are, are attempts at finding a good equilibrium between marketable RT and not too shitty raster. The balancing act is ongoing and every company is going to try and carve out a piece of Unique Selling Point. AMD really does it too, even by not offering hardware for it - if they find ways to accelerate RT performance over their regular GPU floorplan, they have a winner that Nvidia can never get close to again. Alternatively, if Nvidia proves they can do fantastic RT (they haven't, yet) at good performance levels on marketable dies without per-title support being required, they also have a winner. So far though all Nvidia has shown us is price hikes for a minimal RT perf advantage, and exploding TDPs.

So far, not one single color/camp has the best solution and it is anyone's guess where this might go. And the economy around GPUs plays a major role in it too, and its not looking very good currently. That's a huge part of what colours that poll you linked, too. If its affordable, people will happily adopt.
Posted on Reply
#93
londiste
Vayra86What we have now is 16-20% of GPU floor plan reserved to boost RT performance (extra cache also serves it).
RDNA2 Infinity Cache is ~15% of the die and it is huge. Pretty sure it has more benefits than just aiding AMD's RT performance.
Posted on Reply
#94
ratirt
londisteThis is a classic chicken and egg problem. Developing RT effects does not make sense if there is no hardware to support it. Adding hardware to support RT does not make sense if there is no RT effects in software. RT Cores take up ~3% on Turing dies. Relatively small investment for a new capability giving time for the industry to adjust and use/showoff it in a still meaningful way. Scaling them up should not be difficult but there is the same question of cost vs benefit.

In addition to optimizing the very specifically ray tracing side of things, RT advancements in games and software are very much around all the supporting areas that are not hardware accelerated. Notably, building and updating the BVH and other data structures.

Game development is a multi-year process and largely only this year have the major game engines implemented RT effects in a real (and non-beta) way. APIs have somewhat matured in their first incarnation.
Similarly, with AMD having hardware support for RT and Intel also getting some hardware support for RT, this is already getting industry-wide.

Clamoring for full path-tracing is one possible viewpoint but dismissing entire RT until that is very short-sighted IMO. RT has obvious benefits for use cases it is already being used - basically everything to do with lighting but in practice mostly GI, shadows, some aspects of transparency. Reflections too but that probably has more to do with us not having found better ways to do proper reflections than current strengths of RT.
I'm not dismissing anything. ray tracing is cool really but the hardware cant support this in the way it should. Every company is getting their customers teased about the new features and capabilities but that doesnt change the fact, RT has a huge miss in hardware requirements which is hard to compensate for.
With the better ways as you probably know, there are demos that dont use RT. Reflections, lightning, shadows are there and you would not tell the difference. RT is great no doubt but for now it is mostly a marketing point.
Posted on Reply
#95
Tomorrow
Stay on topic guys. This is about possible XTX variant not about RT.
Posted on Reply
#96
Vayra86
londisteRDNA2 Infinity Cache is ~15% of the die and it is huge. Pretty sure it has more benefits than just aiding AMD's RT performance.
I believe part of it is compensating for the choice of GDDR6, isn't it?

I wasn't referring to AMD's 16-20% - I was referring to Nvidia's, where the shader efficiency hasn't improved since Pascal, instead it was chopped up in smaller bits and the cache is likely a byproduct of that too. And they needed that change to cater for their additional RT/Tensor patchwork on top. Not to improve raster.

Meanwhile:

6900XT = 520 mm²
3090 = 628 mm²

That's about 20% bigger right there. Despite the changes since Turing, they still have a net die size increase of 20% for similar rasterized perf. Albeit on a slightly bigger node.
TomorrowStay on topic guys. This is about possible XTX variant not about RT.
True actually.
Posted on Reply
#97
ratirt
Vayra86Meanwhile:

6900XT = 520 mm²
3090 = 628 mm²

That's about 20% bigger right there. Despite the changes since Turing, they still have a net die size increase of 20% for similar rasterized perf. Albeit on a slightly bigger node.
I wonder how much bigger the Samsung's 8nm is vs TSMC 7nm. Something tells me if you had shrunk the 3090 die to TSMC 7nm it would have still been bigger than 6900XT's die is.
almost a 100mm2 is a lot of a differnece
Posted on Reply
#98
londiste
Vayra86I believe part of it is compensating for the choice of GDDR6, isn't it?

I wasn't referring to AMD's 16-20% - I was referring to Nvidia's, where the shader efficiency hasn't improved since Pascal, instead it was chopped up in smaller bits and the cache is likely a byproduct of that too. And they needed that change to cater for their additional RT/Tensor patchwork on top. Not to improve raster.

Meanwhile:
6900XT = 520 mm²
3090 = 628 mm²

That's about 20% bigger right there. Despite the changes since Turing, they still have a net die size increase of 20% for similar rasterized perf. Albeit on a slightly bigger node.
As you well understood, my post was deliberately ignoring Nvidia there :)

There is a reddit post where detailed Turing die shots were analyzed. What he came up seems to be correct enough, Tensor cores and FP16 capabilities may be more nuanced but RT Cores are distinguishable and straightforward. RT Cores make up about 6% of TPC and about 3% of total die size. The increase for Tensor cores and/or FP16 capability concurrent to FP32 has more/most uses outside RT, same for cache. Implementation for AMD and Intel should not be too much different in terms of transistors and area cost, possibly less.

I wish there were good/readable enough die shots for RDNA2 and Ampere but apparently not so far. Would also need comparisons without RT and in case of RDNA where RT capability is part of some other block (TMU?) it is probably impossible to read.

3090 is on Samsung 8N, 6900XT is on TSMC N7:
- 3090 die is 28.3B transistors on 628 mm² - 45 MTr/mm²
- 6900XT die is 26.8B transistors on 520 mm² - 51 MTr/mm²
This highlights the differences in manufacturing processes more than anything.
In terms of transistors/area cost of latest improvements RDNA2 has huge amount of transistors (at least 6.4B plus some control logic which is 24% of total transistors) in Infinity cache, Ampere no doubt has a lot of transistors in the doubled ALUs in shaders.

More cache has been the go-to improvement for a few generations before RDNA and Turing. More likely than not adding more and more cache (at different levels) would happen with or without RT.
ratirtI wonder how much bigger the Samsung's 8nm is vs TSMC 7nm. Something tells me if you had shrunk the 3090 die to TSMC 7nm it would have still been bigger than 6900XT's die is.
almost a 100mm2 is a lot of a differnece
Assuming similar transistor density as 6900XT, 3090 on N7 would be 5.5% larger, about 30 mm².
That assumption is obviously suspect though. Without Infinity Cache 6900XT die would be noticeably less dense. On the other hand, there is A100 on TSMC's N7 with 54.2B transistors and 826mm² making the density out to 65,6 MTr/mm².
Posted on Reply
#99
Richards
londisteAs you well understood, my post was deliberately ignoring Nvidia there :)

There is a reddit post where detailed Turing die shots were analyzed. What he came up seems to be correct enough, Tensor cores and FP16 capabilities may be more nuanced but RT Cores are distinguishable and straightforward. RT Cores make up about 6% of TPC and about 3% of total die size. The increase for Tensor cores and/or FP16 capability concurrent to FP32 has more/most uses outside RT, same for cache. Implementation for AMD and Intel should not be too much different in terms of transistors and area cost, possibly less.

I wish there were good/readable enough die shots for RDNA2 and Ampere but apparently not so far. Would also need comparisons without RT and in case of RDNA where RT capability is part of some other block (TMU?) it is probably impossible to read.

3090 is on Samsung 8N, 6900XT is on TSMC N7:
- 3090 die is 28.3B transistors on 628 mm² - 45 MTr/mm²
- 6900XT die is 26.8B transistors on 520 mm² - 51 MTr/mm²
This highlights the differences in manufacturing processes more than anything.
In terms of transistors/area cost of latest improvements RDNA2 has huge amount of transistors (at least 6.4B plus some control logic which is 24% of total transistors) in Infinity cache, Ampere no doubt has a lot of transistors in the doubled ALUs in shaders.

More cache has been the go-to improvement for a few generations before RDNA and Turing. More likely than not adding more and more cache (at different levels) would happen with or without RT.

Assuming similar transistor density as 6900XT, 3090 on N7 would be 5.5% larger, about 30 mm².
That assumption is obviously suspect though. Without Infinity Cache 6900XT die would be noticeably less dense. On the other hand, there is A100 on TSMC's N7 with 54.2B transistors and 826mm² making the density out to 65,6 MTr/mm².
So tsmc's 7nm is 91mm but got 51mm on rx 6900xt. So tsmc's scalling of sram and logic is lesser than advertised ?
Posted on Reply
#100
Vayra86
RichardsSo tsmc's 7nm is 91mm but got 51mm on rx 6900xt. So tsmc's scalling of sram and logic is lesser than advertised ?
The density of different blocks on a chip varies a lot.
londisteAs you well understood, my post was deliberately ignoring Nvidia there :)

There is a reddit post where detailed Turing die shots were analyzed. What he came up seems to be correct enough, Tensor cores and FP16 capabilities may be more nuanced but RT Cores are distinguishable and straightforward. RT Cores make up about 6% of TPC and about 3% of total die size. The increase for Tensor cores and/or FP16 capability concurrent to FP32 has more/most uses outside RT, same for cache. Implementation for AMD and Intel should not be too much different in terms of transistors and area cost, possibly less.

I wish there were good/readable enough die shots for RDNA2 and Ampere but apparently not so far. Would also need comparisons without RT and in case of RDNA where RT capability is part of some other block (TMU?) it is probably impossible to read.

3090 is on Samsung 8N, 6900XT is on TSMC N7:
- 3090 die is 28.3B transistors on 628 mm² - 45 MTr/mm²
- 6900XT die is 26.8B transistors on 520 mm² - 51 MTr/mm²
This highlights the differences in manufacturing processes more than anything.
In terms of transistors/area cost of latest improvements RDNA2 has huge amount of transistors (at least 6.4B plus some control logic which is 24% of total transistors) in Infinity cache, Ampere no doubt has a lot of transistors in the doubled ALUs in shaders.

More cache has been the go-to improvement for a few generations before RDNA and Turing. More likely than not adding more and more cache (at different levels) would happen with or without RT.

Assuming similar transistor density as 6900XT, 3090 on N7 would be 5.5% larger, about 30 mm².
That assumption is obviously suspect though. Without Infinity Cache 6900XT die would be noticeably less dense. On the other hand, there is A100 on TSMC's N7 with 54.2B transistors and 826mm² making the density out to 65,6 MTr/mm².
Great insights, thank you.
Posted on Reply
Add your own comment
Apr 26th, 2024 12:39 EDT change timezone

New Forum Posts

Popular Reviews

Controversial News Posts