Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html

 

[Ajay]

20200193681 MECHANISM FOR SUPPORTING DISCARD FUNCTIONALITY IN A RAY TRACING CONTEXT
20200193682 MERGED DATA PATH FOR TRIANGLE AND BOX INTERSECTION TEST IN RAY TRACING
20200193683 ROBUST RAY-TRIANGLE INTERSECTION
20200193684 EFFICIENT DATA PATH FOR RAY TRIANGLE INTERSECTION
20200193685 WATER TIGHT RAY TRIANGLE INTERSECTION WITHOUT RESORTING TO DOUBLE PRECISION

How did you get those patents numbers? Google Patent finder is giving me different results. Oh, and thanks for all the heavy lifting on this!

 

[DisEnchantment]

How did you get those patents numbers? Google Patent finder is giving me different results. Oh, and thanks for all the heavy lifting on this!

These are patent applications. Google usually shows patent awards. It takes some time(years) to get awarded. You can take the numbers and search directly in USPTO.

 

[Ajay]

These are patent applications. Google usually shows patent awards. It takes some time(years) to get awarded. You can take the numbers and search directly in USPTO.

Thanks, just have to figure out how to show images. Maybe web pages are broken on Firefox

 

[uzzi38]

MLiD’s latest received leak - speculates reasons for fakeness Or not and why in places...

my take = fake

I know the guy that made it, it's fake.

EDIT: A bit of context, said person wanted to really check how much AdoredTV actually sanitise their sources given their latest few leaks. So he made this thing in an attempt to undersell RDNA2 to see how they would take it. He made an attempt to make it look bad too.

Surprisingly, the guys over at AdoredTV not only didn't believe it but also caught every (mostly intentional) mistake. So was a good showing.

Funnily enough MLID was given the slides more of as a "why not" thing. He wasn't the main target at all. Pretty funny outcome that came out of it.

 

[Det0x]

More predictions/guesswork from Coreteks
tldr: 3D stacked nand memory ontop of the Vega20 die

 

[DiogoDX]

Dedicated raytracing chip for nvidia and 3d stacking for amd. This guy is really on serious drugs.

 

[CakeMonster]

I love seeing optimists talking about "revolution". When was the last time we had an actual revolution in performance? I think the Radeon 9700 was possibly the only time in the last 20+ years that we've had more than 30% performance increase over what was available one year earlier.

 

[DeathReborn]

That would have to be Industrial NAND and the GPU temperature throttled to sub ~85c. One journalist said if the traversal coprocessor was real he'd eat his 2080Ti so I presume if this is real he would eat a Vega VII.

 

[DisEnchantment]

Traversal coprocessor (e.g. AMD's Ray Intersection Unit can be called a traversal coprocessor of sorts) is happening, just not offchip like said YouTuber claimed.
Die stacking on GPU/CPU is happening but whether it is on Sienna as claimed by said YouTuber remains to be seen. Same YouTuber said Navi10 has both GDDR6/HBM PHYs.
Die analysis indicate Navi12 doesn't have GDDR6 PHY and Navi10 does not have HBM PHY.

Erste Benchmarks und eine kurze Die-Analyse zu AMDs "Navi 12" Chip | 3DCenter.org

Erste seitens Videocardz sowie Twitterer Rogame ausgegrabene Benchmark-Werte zur Radeon Pro 5600M auf Navi-12-Basis zeigen Apples neue Mobile-Lösung beachtbar unterhalb der Performance einer Radeon RX 5600M auf Navi-10-Basis – welche ihrerseits mit 36

www.3dcenter.org

 

[RetroZombie]

Dedicated raytracing chip for nvidia and 3d stacking for amd. This guy is really on serious drugs.

Couldn't agree more.
This guy, adoredtv, moores law is dead, ... all make great videos.
But when they start doing videos about leaks, is just LOL, it seams they cant even use their own brain to analyze what they are saying.

For example in his first video he showed a dual fan design, according to him cools two chips instead of one., ok it's possible, but then:
Impossibility one
How can one fan be in one side of the card and the other reversed?

Impossibility two
To achieve one it would require the graphics card too have two PCBs.

Impossibility three
In other to solve impossibility two, 4 slot designs are incoming, herrr no.

 

[TESKATLIPOKA]

I see significant space savings by using HBM2. They saved 42mm^2 which could translate to additional 16CU for a total of 56CU. HBM2 is great, the problem is It's price. I wonder how much It costs today.

 

[Glo.]

Traversal coprocessor (e.g. AMD's Ray Intersection Unit can be called a traversal coprocessor of sorts) is happening, just not offchip like said YouTuber claimed.
Die stacking on GPU/CPU is happening but whether it is on Sienna as claimed by said YouTuber remains to be seen. Same YouTuber said Navi10 has both GDDR6/HBM PHYs.
Die analysis indicate Navi12 doesn't have GDDR6 PHY and Navi10 does not have HBM PHY.

View attachment 23628

Erste Benchmarks und eine kurze Die-Analyse zu AMDs "Navi 12" Chip | 3DCenter.org

Erste seitens Videocardz sowie Twitterer Rogame ausgegrabene Benchmark-Werte zur Radeon Pro 5600M auf Navi-12-Basis zeigen Apples neue Mobile-Lösung beachtbar unterhalb der Performance einer Radeon RX 5600M auf Navi-10-Basis – welche ihrerseits mit 36

www.3dcenter.org

So it is actually 41 mm2 smaller than regular navi 10. That is genuinely surprising. Thanks.

 

[lobz]

I love seeing optimists talking about "revolution". When was the last time we had an actual revolution in performance? I think the Radeon 9700 was possibly the only time in the last 20+ years that we've had more than 30% performance increase over what was available one year earlier.

Not really. The 1070 was as fast as the 980ti and the 1080ti was A LOT faster than Maxwell. Not quite pascal 10x maxwell of course, but that was a huge generational leap, even though people expected even more from an almost double node jump.

 

[kurosaki]

I love seeing optimists talking about "revolution". When was the last time we had an actual revolution in performance? I think the Radeon 9700 was possibly the only time in the last 20+ years that we've had more than 30% performance increase over what was available one year earlier.

The 8800gt and gtx were huuuge bang for the buck, then I guess amd doubled amount of cu's every year from 1900xt -3870xt at least.

 

[dr1337]

I love seeing optimists talking about "revolution". When was the last time we had an actual revolution in performance? I think the Radeon 9700 was possibly the only time in the last 20+ years that we've had more than 30% performance increase over what was available one year earlier.

Actually RDNA 1 brought on average more than 50% increase in perf/watt over GCN. Both the 580 and 5700 have the same number of stream processors and compute units. But even in a game like GTA, the 5700 is almost exactly 2x the performance of the 580 and it consumes the same or less power depending on task.

All of that said the one place rdna1 wasn't a revolution is price. If the 5700 was released as a $250 gpu like the 580 was, we wouldn't even be having this conversation. Though I don't fault AMD for pricing navi the way they did, Im sure they didn't have the yields let alone volume to actually meet the markets demand with a hypothetical $250 36cu navi gpu.

Whats really interesting though is that if AMD really delivers on the 50% perf/watt gains again with RDNA 2 we will have actually seen a 125%* increase in performance in only two years. While AMD is still playing catch up architecturally compared to Nvidia, it seems like they are making some revolutionary strides in the right direction.

edit: maths

 

[H T C]

Whats really interesting though is that if AMD really delivers on the 50% perf/watt gains again with RDNA 2 we will have actually seen a 100% increase in performance in only two years. While AMD is still playing catch up architecturally compared to Nvidia, it seems like they are making some revolutionary strides in the right direction.

Won't that be 125%?

50% last year + 50% of that this year equals 125% VS two years ago, no?

 

[dr1337]

Won't that be 125%?

50% last year + 50% of that this year equals 125% VS two years ago, no?

You're absolutely right, I had a total brain fart. Even AMDs slide I posted specifically shows a 50% increase from rdna1 to rdna2. Which very much would be a 125% increase over GCN.

 

[Saylick]

I love seeing optimists talking about "revolution". When was the last time we had an actual revolution in performance? I think the Radeon 9700 was possibly the only time in the last 20+ years that we've had more than 30% performance increase over what was available one year earlier.

The 9700 was a beast that I was too young to experience, but I'd argue the last time we got >30% improvements year-over-year was when AMD and Nvidia transitioned to 28nm. Up to and including that point, it was common to see 50% or greater gains every generation, and each generation came out more or less within a year of the prior one. The transition to smaller nodes after 28nm has been slower than usual, with 20nm being a bust, 16nm being 20nm with FinFETs, and the relatively slow transition to 7nm. 7nm, as you know, will be another multi-year node like how 16nm/14nm was. The same will be true of 5nm, etc.

 

[TESKATLIPOKA]

Actually RDNA 1 brought on average more than 50% increase in perf/watt over GCN. Both the 580 and 5700 have the same number of stream processors and compute units. But even in a game like GTA, the 5700 is almost exactly 2x the performance of the 580 and it consumes the same or less power depending on task.

All of that said the one place rdna1 wasn't a revolution is price. If the 5700 was released as a $250 gpu like the 580 was, we wouldn't even be having this conversation. Though I don't fault AMD for pricing navi the way they did, Im sure they didn't have the yields let alone volume to actually meet the markets demand with a hypothetical $250 36cu navi gpu.

Whats really interesting though is that if AMD really delivers on the 50% perf/watt gains again with RDNA 2 we will have actually seen a 125%* increase in performance in only two years. While AMD is still playing catch up architecturally compared to Nvidia, it seems like they are making some revolutionary strides in the right direction.

edit: maths

1. The performance or perf/W increase wasn't free. Navi10 (RDNA1) has 80% more transistors than Polaris (10.3 ×10^9 vs 5.7 ×10^9 ) and by using 7nm process they could achieve higher clocks while keeping power consumption within an acceptable range.
2. RDNA1 is more expensive to make than polaris thanks to using 7nm process and gddr6, but It certainly doesn't cost $100 more to produce. AMD just wanted to make more more money from a possibly limited supply.
3. Polaris was already released 4 years ago and Vega almost 3 years ago and RDNA2 is yet to be released. BTW 125% is performance per W.

 

[Headfoot]

I love seeing optimists talking about "revolution". When was the last time we had an actual revolution in performance? I think the Radeon 9700 was possibly the only time in the last 20+ years that we've had more than 30% performance increase over what was available one year earlier.

8800 GTX too. Big shakeups like 9700 and 8800 dont happen very often so it would be really cool if it did again

 

[Mopetar]

It's getting a lot harder to get big shakeups. Node advances are slowing down which makes it more likely to get multiple generations on the same process node or on one with minor tweaks.

The chips themselves have been pushed to the reticle limit so there's no way to just make a monster chip that's almost twice as large as anything before it. Similarly the power limits have been hit as well and there's no major headroom to play with there either.

We're also seeing companies use transistor budget for features like compute or ray tracing that don't provide as much performance uplift in a general sense.

Unless someone makes an exciting discovery that allows them to do things in a fundamentally different way that's vastly more efficient then we're unlikely to see anything quite on the level of the 8800 or 9700 unless we get a lot of stagnation by I can't see NVidia or AMD slowing down.

 

[Headfoot]

if they manage to make chiplets for GPU work with the same success they have for Ryzen that could be a major shakeup moment but I agree, I dont see it happening following the same monolithic chip trends of today

 

[pandemonium]

if they manage to make chiplets for GPU work with the same success they have for Ryzen that could be a major shakeup moment but I agree, I dont see it happening following the same monolithic chip trends of today

Did....did you just justify Intel's Xe architecture?

 

[GodisanAtheist]

Did....did you just justify Intel's Xe architecture?

"Having chiplets work with the same success as Ryzen" was a key condition of his statement, so no.

I don't think anyone expects Xe to come even close to current monolithic designs for gaming purposes.

 

[Ajay]

if they manage to make chiplets for GPU work with the same success they have for Ryzen that could be a major shakeup moment but I agree, I dont see it happening following the same monolithic chip trends of today

Multi-chip implementations are usually straighforward for compute. Filtering and then coalescing data structures for VFX rendering is a bigger hurdle. One doesn't want code in core1 needing data in core3's L2 cache.