AMD: RDNA 3 Speculation, Rumours and Discussion

[PSman1700]

They were so great back in the day. The 9700Pro was my favorite GPU I've ever owned.

I want RDNA3 to be another 9700 Pro.. just smoking the competition!

Still have one, the 9700pro. Played quite many DX9 games on that one, it was a very capable dx9 gpu compared to the competition at the time.

Same for me. Not only was it the top performer for a reasonable price, it aged great. it didn't start performing poorly in new titles 18 months in.

Seemed to also offer better image quality at the time, colors where better to me (texture filtering?). Also loved the red-board design which was fitting ATI's colors.
GCN (7950 etc) wasn't direectly faster than say 670 at the time however they aged much and much better due to the course of rendering we went (compute).

Anyway im still hopefull AMD will do another 'Zen3' with RDNA3 to disrupt the market abit. We also have Intel now so theres more competition than before.

 

[iroboto]

Still have one, the 9700pro. Played quite many DX9 games on that one, it was a very capable dx9 gpu compared to the competition at the time.



Seemed to also offer better image quality at the time, colors where better to me (texture filtering?). Also loved the red-board design which was fitting ATI's colors.
GCN (7950 etc) wasn't direectly faster than say 670 at the time however they aged much and much better due to the course of rendering we went (compute).

Anyway im still hopefull AMD will do another 'Zen3' with RDNA3 to disrupt the market abit. We also have Intel now so theres more competition than before.

It’s a great year to be a graphics enthusiast. People are overly critical, when they should be critical but appreciative of what each company is trying to do.

How rare that 2 companies split visions on how to approach the same problem. In a single year we get a massive innovation towards frame generation through AI, and now we will see massive innovation in silicon design in the form of chiplets in the GPU space.

Both may miss the mark of what customers may want from both of them, but it is progress. They will both get there.

 

[PSman1700]

It’s a great year to be a graphics enthusiast. People are overly critical, when they should be critical but appreciative of what each company is trying to do.

How rare that 2 companies split visions on how to approach the same problem. In a single year we get a massive innovation towards frame generation through AI, and now we will see massive innovation in silicon design in the form of chiplets in the GPU space.

Both may miss the mark of what customers may want from both of them, but it is progress. They will both get there.

Indeed, its one of the greatest times in GFX i think. AI/ML is intresting and so are new technologies like ray tracing but also normal raster tech that still sees nice improvements. Only hurdle so far is price, but that will adjust if competition enters. See what happened to intel and its stagnation untill Zen3 came along. Its also great to see Intel joining the club.

 

[techuse]

It’s a great year to be a graphics enthusiast. People are overly critical, when they should be critical but appreciative of what each company is trying to do.

How rare that 2 companies split visions on how to approach the same problem. In a single year we get a massive innovation towards frame generation through AI, and now we will see massive innovation in silicon design in the form of chiplets in the GPU space.

Both may miss the mark of what customers may want from both of them, but it is progress. They will both get there.

Great from a graphics enthusiast perspective, terrible from a consumer perspective. Never have you had to pay more for less.

 

[iroboto]

Great from a graphics enthusiast perspective, terrible from a consumer perspective. Never have you had to pay more for less.

Yea lol. So terrible

 

[PSman1700]

Great from a graphics enthusiast perspective, terrible from a consumer perspective. Never have you had to pay more for less.

Thats generally true for every tech market right now in special here in EU.

 

[swaaye]

I want RDNA3 to be another 9700 Pro.. just smoking the competition!

That happened because GeForce FX was hopeless for D3D 9. Probably because NVidia didn't predict that D3D 9 would matter much in 2002-2003. Ada doesn't seem to have anything like that going on.

 

[PSman1700]

That happened because GeForce FX was useless for D3D 9. Probably because NVidia didn't predict that D3D 9 would matter much in 2002-2003. Ada doesn't seem to have anything like that going on.

Somewhat similar to the Kepler architectured GPU's that werent directly worse to GCN gpus, just the wrong direction where GCN went compute and Kepler didnt. Maxwell adjusted that.

 

[TESKATLIPOKA]

Looks to me like they made compromises to save on the area and it will affect performance in higher resolutions. If you want the real deal, you'll want to get N31 with 3D stacked IC.

It's always about compromises, but RDNA3 SKUs should have enough bandwidth to not be significantly affected.
If It was affected significantly, then there is no reason to have so many WGPs inside, If they can't be kept fed.

N23 had enough bandwidth to not be affected at 1080p and even 1440p was not that bad.
N33 looks a lot similar to N23, but shaders and TFlops increased drastically.
N33 would need only 2.9GHz clockspeed to be on par with RX 6950xt in theoretical TFlops, realistically It will be clocked higher.
For gaming performance, let's say 1 RDNA3 WGP = 1.5 RDNA2 WGP, that would mean 10.8 Tflops * 1.5 * 1.27 (boost: 3.3/2.6 GHz) = 20.6 Tflops or ~RX 6800xt level of performance.

	Theoretical TFlops	GDDR6	Bus width	Bandwidth	Infinity cache
RX 6650xt	10.8 (100%)	17.5 GHz	128 bit	280 GB/s (100%)	32 MB (100%)
N33 at 3.3GHz	27 (250%)	24 GHz	128 bit	384 GB/s (137%)	32 MB (100%)
RX 6800xt	20.7 (192%)	16 GHz	256 bit	512 GB/s (183%)	128 MB (400%)
RX 6950xt	23.7 (219%)	18 GHz	256 bit	576 GB/s (206%)	128 MB (400%)

N33 has a serious deficit in bandwidth even with 24 GHz memory, and a small infinity cache doesn't help either.
Does It not affect lower resolution?

 

[BRiT]

Let's try to keep this thread about RNDA 3 and not about all the things it isn't, like Intel CPUs or Nvidia GPUs.

 

[LordEC911]

Rumored Infinity Cache amounts still makes very little sense to me looking from a performance/hit-rate perspective, ignoring cost.
192mb makes sense for high performance +4k
128mb makes sense for high performance 1440p/4k-ish, though 96mb would be closer to the sweetspot for 1440p.
64mb appears very close to being the sweetspot for 1440p/1080p.
32mb really only makes sense for a pure 1080p GPU with hitrate at ~55%, which if Navi33 is close to Navi21 performance levels shouldn't be a 1080p GPU.

Writing out a rather long explanation about the rumored die size of the MCD and cache, I went back and re-read Navi21 info.
Part of the explanation of the large cache was power efficiency through locality and keeping data on-die.... but MCDs are technically off-die.
Maybe I'm late to the game and missed where this was discussed earlier- but what if there is an increased cache on the GCD before data gets sent out to the MCD?
It wouldn't have to be as large as the MCD cache but it could explain why the MCD cache isn't as large as I assumed it should be....

I'm still assuming that Navi33 should have 64mb of Infinity Cache, though.
Edit- Another sidepoint I forgot to note, is how much FSR2 might have impacted AMD's design/architectural choices this generation.

 

[eastmen]

Rumored Infinity Cache amounts still makes very little sense to me looking from a performance/hit-rate perspective, ignoring cost.
192mb makes sense for high performance +4k
128mb makes sense for high performance 1440p/4k-ish, though 96mb would be closer to the sweetspot for 1440p.
64mb appears very close to being the sweetspot for 1440p/1080p.
32mb really only makes sense for a pure 1080p GPU with hitrate at ~55%, which if Navi33 is close to Navi21 performance levels shouldn't be a 1080p GPU.

Writing out a rather long explanation about the rumored die size of the MCD and cache, I went back and re-read Navi21 info.
Part of the explanation of the large cache was power efficiency through locality and keeping data on-die.... but MCDs are technically off-die.
Maybe I'm late to the game and missed where this was discussed earlier- but what if there is an increased cache on the GCD before data gets sent out to the MCD?
It wouldn't have to be as large as the MCD cache but it could explain why the MCD cache isn't as large as I assumed it should be....

I'm still assuming that Navi33 should have 64mb of Infinity Cache, though.
Edit- Another sidepoint I forgot to note, is how much FSR2 might have impacted AMD's design/architectural choices this generation.

It seems like you'd want more than 192 for 4k

 

[iroboto]

View attachment 7058

It seems like you'd want more than 192 for 4k

It'll have to scale with the number of CUs as opposed to a straight number otherwise CU utilization will be low.
Curious to see how they handle IC with chiplets

 

[Frenetic Pony]

View attachment 7058

It seems like you'd want more than 192 for 4k

192mb seems like a compromise between continually adding more cache for performance gains in ever fewer titles and being big enough. Cache is still expensive, now because of the packaging rather than making it. Maybe when packaging gets cheaper we'll see a 600w liquid cooled Navi 31 with 288mb, "7900LC".

 

[LordEC911]

View attachment 7058

It seems like you'd want more than 192 for 4k

I could be wrong but scaling-wise it looks like the 4k line would start to flatten out pretty quickly. 192mb should be hitting that 75-80% hit rate at 4k that Navi21 was doing at 1440p/1080p.

 

[Bondrewd]

Rumored Infinity Cache amounts still makes very little sense to me looking from a performance/hit-rate perspective, ignoring cost.

Only N33 has the ~same amount of offchip bytes as its predecessor.
Remember that N31 is a 384b device.

 

[Kaotik]

Semi-ot, but please don't use GHz when talking about memory. The correct term is Gbps and nothing is working at even remotely that kind of clocks.
I hate it how even "serious tech sites" get it constantly wrong, no matter are we talking about GDDR or DDR or whatever memory.

 

[TESKATLIPOKA]

Semi-ot, but please don't use GHz when talking about memory. The correct term is Gbps and nothing is working at even remotely that kind of clocks.
I hate it how even "serious tech sites" get it constantly wrong, no matter are we talking about GDDR or DDR or whatever memory.

Ok, no problem, I can use Gbps, but at wiki they use MT/s(transfers per second) for example. So which one is the correct one?

 

[Kaotik]

Ok, no problem, I can use Gbps, but at wiki they use MT/s(transfers per second) for example. So which one is the correct one?

Both. MT/s is better, as it doesn't care about encoding scheme used it's just raw transfers, but Gbps is acceptable too (you just need to take encoding into equation too)
Or rather Gbps is one step ahead in the equation of raw transfers to actual data transferred
(Just came home from night shift and might not be sober anymore, hope it's understandable)

 

[Phantom88]

Same for me. Not only was it the top performer for a reasonable price, it aged great. it didn't start performing poorly in new titles 18 months in.

It seems that all ATI cards from that era, from the 8500 to 9000 series and beyond perform better than they did at release, probably due to more mature drivers. The 9700 PRO is doing exceptionally well in high end games like FEAR which were 3 years later.