AMD Radeon RDNA2 Navi (RX 6500, 6600, 6700, 6800, 6900 XT)

Didn't David Wang promise yearly cadence? Also wth is N31 entry in MacOS? It's obviously not something RDNA3 based.

Did he? For Zen they've mentioned they're on a 12-18 month cadence. I can't recall anything specific for graphics. Why can't it be RDNA3?
process drop navi 2 next year , navi 3 in 2022 makes sense to me

or RDNA3 is the process drop

So it will be late Q1'21 at best by the time they launch the entire RDNA 2 stack, maybe even Q2 for N24. And Q3 by the time mobile variants will be out. I don't see them launching the next gen in just the next quarter (i.e. Q4'21). Large GPUs on new processes take a lot of effort, and I don't think AMD is still in such a position that they have the engineering resources to do so much in such short periods of time. Even Nvidia took ~2 years between Maxwell, Pascal, Turing and Ampere. I expect AMD to launch minor refreshes of the desktop chips late next year, with RDNA 3 following in 2022.
What about rumors about the existence of other versions of Navi21 with HBM2? The memory bandwidth level for tandem from HBM2 and Infinity Cache is very interesting. Maybe this option will appear in Apple Mac Pro systems.

Infinity cache further obviates the need for HMB2 in consumer GPUs, though it could perhaps be used for premium mobile solutions for extreme power efficiency. HBM2 + Infinity cache should be a good fit for CDNA though, and I wonder when that's going to launch. Certainly should be soon if its meant to go into supercomputers by mid 2021.
 
Did he? For Zen they've mentioned they're on a 12-18 month cadence. I can't recall anything specific for graphics. Why can't it be RDNA3?
The roadmap shows RDNA 3 in 2022 ("before the end of 2022"), but Navi 31 is expected in 2021, if I'm not mistaken. Maybe Navi 31 is Navi 21 on N7+ or N5(P).
 
The roadmap shows RDNA 3 in 2022 ("before the end of 2022"), but Navi 31 is expected in 2021, if I'm not mistaken. Maybe Navi 31 is Navi 21 on N7+ or N5(P).
To repeat myself: while they have told AnandTech that it's inclusive, every single roadmap they've had with that format hasn't been, meaning that roadmaps ending in 2020 for example has had everything launching before the start of 2020 and so on.

So it will be late Q1'21 at best by the time they launch the entire RDNA 2 stack, maybe even Q2 for N24. And Q3 by the time mobile variants will be out. I don't see them launching the next gen in just the next quarter (i.e. Q4'21). Large GPUs on new processes take a lot of effort, and I don't think AMD is still in such a position that they have the engineering resources to do so much in such short periods of time. Even Nvidia took ~2 years between Maxwell, Pascal, Turing and Ampere. I expect AMD to launch minor refreshes of the desktop chips late next year, with RDNA 3 following in 2022.
If RDNA3 is the process drop with possible minor tweaks there's no reason why it shouldn't come in 2021.

Here's David Wang saying they'll try to get back to "new GPU every year" schedule
https://www.pcworld.com/article/327...early-gpu-releases-to-make-pcs-fun-again.html

Navi31 being shrunk and possibly slightly tweaked Navi21 coming within a year would fit that perfectly
 
Last edited:
The roadmap shows RDNA 3 in 2022 ("before the end of 2022"), but Navi 31 is expected in 2021, if I'm not mistaken. Maybe Navi 31 is Navi 21 on N7+ or N5(P).

I haven't seen any rumors about the N31 launch date (or perhaps I've missed them), only that the configuration is the same as N21. Quite possible that N31 is some sort of a shrink, and as I've mentioned, the deliberate use of "advanced node" for RDNA3 on the roadmap instead of an explicit mention of 5nm leads me to believe it could be on 6nm instead.
To repeat myself: while they have told AnandTech that it's inclusive, every single roadmap they've had with that format hasn't been, meaning that roadmaps ending in 2020 for example has had everything launching before the start of 2020 and so on.

For Zen 4 which has a similar position on the roadmap as RDNA 3, if I'm not mistaken they clarified that the time-frame implied that all products based on Zen 4 would launch before the end of 2022 (desktop, workstation, server, mobile)
If RDNA2 is the process drop with possible minor tweaks there's no reason why it shouldn't come in 2021.

Here's David Wang saying they'll try to get back to "new GPU every year" schedule
https://www.pcworld.com/article/327...early-gpu-releases-to-make-pcs-fun-again.html

Navi31 being shrunk and possibly slightly tweaked Navi21 coming within a year would fit that perfectly

Thanks for the link, I'm quoting the relevant part from it:-

"Nonetheless, in a small roundtable following AMD’s presentation, Wang confirmed that AMD would be bringing out a new graphics product every year, via a new architecture, process changes, or “maybe incremental architecture changes.”

So all options are possible then.

 
Some major swings back and forth depending on the title. It’ll be interesting to see if and how numbers change based on different settings choices by independent reviewers. E.g. AMD chose to disable AA on SotTR.
Isn't that a huge red flag? Isn't it normally tested with TAA?

One thing I have to say. If independent reviewer benchmarks bear out AMD's claims WRT to RDNA 1, then RDNA 2 is a phenomenally impressive accomplishment for AMD regardless of how it compares to NV.

NV's accomplishment with Ampere compared to Turing is pretty impressive, but this goes beyond that WRT being a IHV centric generation to generation improvement if RDNA 2 ends up being well over twice as fast as RDNA 1.

For AMD (formerly ATI). as a generation to generation uplift? I don't know if it rivals R300 versus R200, but it's pretty damned impressive none-the-less.
I can't say I'm impressed. 5700XT clearly has far too much bandwidth and fillrate. It's a seriously unbalanced architecture. There's also the lingering rumour of serious problems that made AMD abandon "5800" and why the Vega 7 emergency pisstake was launched. A lot of the improvements in RDNA 2 appear to be "we couldn't afford these on RDNA 1", e.g. the super-fine-grained clock gating which is claimed to bring 30% extra performance per watt in RDNA 2's CUs.

Infinity Cache is deeply impressive to me. This is AMD "catching up" to:

RDNA2 is certainly looking like AMD's Maxwell
but brute force cache doesn't seem elegant in comparison with Maxwell's extremely robust tiled rendering.

On the other hand, as pixel shading becomes less of a bottleneck (i.e. the percentage of frame time spent running pixel shaders) a rambo cache comes into its own.

Let's say average clock of the 3090 is 1900 MHz. It has 112 rops so that's 213 GPixels/s
Let's say the average game clock of a RX 6800 is 1815. It has 96 rops so that's 174 GPixels/s

Pretty interesting. Means it's not a fillrate advantage or a a compute advantage. Shouldn't be texture sampling either. Definitely not bandwidth. 3090 should even have a geometry advantage. I wonder if the difference is a straight up power limit in some way. Wonder if a 3090 undervolted with roughly the same clocks would actually perform better. I am very curious. Does SAM actually a dress a significant bottleneck that gone under the radar? Or are some of these games very cache friendly in some way, where the infinity cache leads to big gains.
Hmm...

Reviews should be good. I wonder if default clocks were used on the latest benchmarks or if there were any overclocks.
It would appear that AIB cards are going to be a solid 10% faster than reference. So Rage Mode is just a hint of what Ampere will be fighting.
 
but brute force cache doesn't seem elegant in comparison with Maxwell's extremely robust tiled rendering.

True, but it's much more elegant than pushing GDDR6X to ridiculous speeds and sacrificing Perf/W with it.
The big question is, why wasn't this possible before? What's the secret sauce in the Infinity cache, that it only became feasible with RDNA2.
 
It would appear that AIB cards are going to be a solid 10% faster than reference. So Rage Mode is just a hint of what Ampere will be fighting.
Interesting, can't wait to see the AIB OC scenarios tested. Rage Mode is similar to Evga's Precision X1 or MSI's Afterburner, so wonder if AMD's EULA will have include Rage Mode as part of warranty coverage.
October 29, 2020
True automatic overclocking can be found in EVGA’s Precision X1 overclocking software, as well as in MSI’s Afterburner. These programs alter clock speeds past the points designated by Nvidia, and attempt to find maximum stable settings. With Radeon Software, enthusiasts will be able to perform the same function. AMD said there will be “automatic overclocking with one-click options to undervolt the GPU and overclock the GPU or VRAM.” In addition to that, there will be “manual tuning, which offers precise control over a range of GPU operating parameters.”

The distinction is that these options will void warranties, as is traditionally expected. Users will have to agree to the EULA before using the features, and AMD is not liable for whatever you choose to do.
https://www.pcinvasion.com/amd-rage-mode-overclocking-warranty/

Nvidia supposedly has a new automatic OCing feature in GFE but have not seen it used in reviews. Reviewers will likely draw comparisons between the two once the Rage Mode or Nvidia's OC feature are enabled.
 
Last edited by a moderator:
but brute force cache doesn't seem elegant in comparison with Maxwell's extremely robust tiled rendering.

On the other hand, as pixel shading becomes less of a bottleneck (i.e. the percentage of frame time spent running pixel shaders) a rambo cache comes into its own.

It's not like all of the perf/W improvement is from the cache though. From the slide AMD showed, it appears that roughly a third of the per/W improvement is due to the cache. And I suspect the the infinity cache design is forward looking, more so for APUs and future chiplet architectures as well perhaps. And It seems to be something that should scale better than analog with smaller nodes.

I haven't read much into rambo cache. Is this for Ponte Vecchio? What are the differences?
It would appear that AIB cards are going to be a solid 10% faster than reference. So Rage Mode is just a hint of what Ampere will be fighting.

Apparently rage mode is just a few percent increase and is not meant to challenge AIB OC cards directly. Those should still offer the best performance, with even more OC potential.
True, but it's much more elegant than pushing GDDR6X to ridiculous speeds and sacrificing Perf/W with it.
The big question is, why wasn't this possible before? What's the secret sauce in the Infinity cache, that it only became feasible with RDNA2.

GDDR6X is actually more power efficient than GDDR6 so you're not sacrificing per/W but as the transfer speeds are higher you are expending more absolute power.

I don't think there's any secret sauce, there are simply trade-offs involved. The cache design is probably a combination of AMD being able to invest more in R&D, having a process node where the PPA (power, performance, area) implications are good enough, CPU design methodologies being integrated and all of this coming together in time to implement in RDNA2.
 
Interesting, can't wait to see the AIB OC scenarios tested. Rage Mode is similar to Evga's Precision X1 or MSI's Afterburner, so wonder if AMD's EULA will have include Rage Mode as part of warranty coverage.

https://www.pcinvasion.com/amd-rage-mode-overclocking-warranty/

Nvidia supposedly has a new automatic OCing feature in GFE but have not seen it used in reviews. Reviewers will likely draw comparisons between the two once the Rage Mode or Nvidia's OC feature are enabled.
https://www.anandtech.com/show/1620...-starts-at-the-highend-coming-november-18th/2

Meanwhile it should be noted that AMD has confirmed that Rage Mode is not warranty voiding. So users can still claim warranty coverage on their cards if they’ve used AMD’s official, user-friendly methods of overclocking.
 
True, but it's much more elegant than pushing GDDR6X to ridiculous speeds and sacrificing Perf/W with it.
The big question is, why wasn't this possible before? What's the secret sauce in the Infinity cache, that it only became feasible with RDNA2.
The driver may be the secret sauce?

The re-worked L1 and L2 themselves are required to make this practical? Those improvements require the driver?

The "CGI die shots" appear to show that the "Infinity Fabric" is quite large in area and some parts of the Infinity Cache, which bulge (the "squares"):


might be the memory controllers? Or, put another way, the memory controllers are not labelled, so somewhere in that combined "grey and blue" we should find the memory controllers.

So how much die area would 256MB of infinity cache be? Substantially less than double the area seen for the combined "grey and blue", would be my guess.

Is 128MB the "sweet spot" or is it merely an experiment? Why didn't AMD go for about 600mm²?

Overall we're looking at something like 150mm² for the grey and blue areas - that's a guess though. Also, some of that would be present even without the infinity cache. Regardless, that's a lot of die space, though redundancy makes it cheap space.
 
...so is Raja finally vindicated with Navi?

The man took a lot of flak for Hawaii, Polaris, and Vega.
He was a part of AMD between April 2013 and September 2017. He would barely affect the Oct 2013 launch of Hawaii. However, he was at full force at the "overclockers' dream" 4GB Fiji, clownish hyping of Polaris and was the face of the "Poor Volta" hype surrounding Vega.

Just for illustration, this was Raja's work:
* Mar 23rd 2016 - GDC 2016 - next-gen highend Vega announcement
* Dec 12th 2016 - AMD Tech Summit - Vega ES running Doom@4k
* Jan 1st 2017 - "After the Uprising" aka "Poor Volta" video & web http://ve.ga - promo featuring Vega buzzwords
* Jan 5th - CES 2017 - "Vega architecture preview" slides
* Feb 28th - GDC 2017 - "RX Vega brand reveal"
* May 17th - AMD FAD 2017 - "Vega FE reveal"
* Jun 27th - "Vega FE launch" with almost no availability
* Jul 31st - Siggraph 2017 - "RX Vega NDA end"
* Aug 14th - "RX Vega launch"

Sure, he was there during major parts of Navi development and apparently also oversaw initial stages of Navi2's too. But still, he and his stunts destroyed the RTG's public image.
 
Last edited:
Cause for hope but if this were accurate why wouldn't AMD have shown it themselves at the launch?

I’m curious as well. Tomb Raider is shadows only right? I wonder if their RT performance scales differently across other games. Or maybe it’s just drivers and they didn’t have good numbers when they readied the presentation. Who knows. It’ll be interesting if they put out their own benchmarks.
 
Back
Top