AMD: Navi Speculation, Rumours and Discussion [2019-2020]

[Bondrewd]

What am I supposed to see in that firmware header code commit?

GDDR6 inits in their usual BIOS <> driver interface added for N21.

 

[CarstenS]

Wouldn't a 384-bit bus limit AMD to using 12 GB or 24 GB?

12, 24 or 48 GByte.

Techtubers and their sources. And they always have a few of them. To cover all the bases.

It would indeed be very sad, if AMD had to pick the xx70 as competition.

GDDR6 inits in their usual BIOS <> driver interface added for N21.

If I were AMD, I would write some really weird shit in there, just for the LULz.

 

[DegustatoR]

Why not throw 96 and 192 GB there as well while we're at 48?

This is getting a bit ridiculous. 10-12 may be there at lower acceptable boundary for 4K+RT but anything higher that 16 won't be used in gaming till PS6 or so.

 

[CarstenS]

Why not throw 96 and 192 GB there as well while we're at 48?

This is getting a bit ridiculous. 10-12 may be there at lower acceptable boundary for 4K+RT but anything higher that 16 won't be used in gaming till PS6 or so.

Because the question was, what kind of limit a 384 bit bus imposes. And 48 GByte is what's possible and has been done already with todays tech, not some fancy yet-to-be-produced memory dies.

 

[Bondrewd]

This is getting a bit ridiculous

That's the point of silly VRAM configs.

 

[pTmdfx]

Why 384-bit? Why not throwing 512-bit into the party, since it lines up with rumours of 8GB/16GB configurations and having 16 memory channels (the “HBM confirmed” patch)?

 

[Bondrewd]

Why 384-bit?

Because.

Why not throwing 512-bit

Because not possible with G6, also just stack 3 HBMs side to side if you're this desperate (and AMD's squarely mobile first uArch is obviously not).

since it lines up with rumours of 8GB/16GB configurations

That's N22.
(also how the fuck do I make 8gig setup on 512b with G6s? No 4Gb G6 IC exists.)

 

[DDH]

Because.

Because not possible with G6, also just stack 3 HBMs side to side if you're this desperate (and AMD's squarely mobile first uArch is obviously not).

That's N22.
(also how the fuck do I make 8gig setup on 512b with G6s? No 4Gb G6 IC exists.)

How is AMD going to feed a supposedly 80cu monster with a 384 bit bus? What speed gddr6 do you imagine them pairing with that?

 

[Bondrewd]

How is AMD going to feed a supposedly 80cu monster with a 384 bit bus?

Miracles and magic.
And cache subsystem improvements.

They gotta push APU and lower end GPU performance, aka segments where bandwidth won't ever rain from the skies.

What speed gddr6 do you imagine them pairing with that?

14/16/17Gbps depending on segment (or AIC vendor in question).

 

[Kaotik]

Why 384-bit? Why not throwing 512-bit into the party, since it lines up with rumours of 8GB/16GB configurations and having 16 memory channels (the “HBM confirmed” patch)?

I'm pretty sure there's no 4 Gbit GDDR6 so 512-bit would mean 16GB at minimum

 

[Bondrewd]

I'm pretty sure there's no 4 Gbit GDDR6 so 512-bit would mean 16GB at minimum

At this point I feel like we need 12Gb GDDRs for oddball but needed 6/9/12/15/18GB configs.

 

[Jawed]

RTX 2080 has the same bandwidth as RX 5700 XT, yet is from 10-30% faster (excluding 1080p), usually well over 20% faster. e.g. Doom Eternal from:

https://www.techspot.com/article/1999-doom-eternal-benchmarks/

In my opinion that tells us that RX 5700 XT was severely unbalanced. If nothing else these new Navi GPUs should fix that ridiculous waste of bandwidth.

 

[Bondrewd]

In my opinion that tells us that RX 5700 XT was severely unbalanced

They had to hit 8 gigs VRAM and 128b isn't really an option.

If nothing else these new Navi GPUs should fix that ridiculous waste of bandwidth.

No, we really-really need 12Gb ICs.

 

[Frenetic Pony]

RTX 2080 has the same bandwidth as RX 5700 XT, yet is from 10-30% faster (excluding 1080p), usually well over 20% faster. e.g. Doom Eternal from:

https://www.techspot.com/article/1999-doom-eternal-benchmarks/

In my opinion that tells us that RX 5700 XT was severely unbalanced. If nothing else these new Navi GPUs should fix that ridiculous waste of bandwidth.

It had room for improvement, and there's been some sort of improvement as per the Series X SOC talk; specifically one SE to one 64bit bus no longer need be the case and more instructions can issued per clock. But that it, they didn't really go into anything informative for the GPU otherwise. The presenter even went out of their way to avoid talking about TDP during the Q&A, like they signed an NDA for it.

AMD playing shit close real close to the vest.

 

[Bondrewd]

The presenter even went out of their way to avoid talking about TDP during the Q&A, like they signed an NDA for it.

Yeah and the actual SoC overview was much shorter vs comparable Renoir session aka neither MS nor AMD wanted to really talk the details.

 

[Jawed]

AMD playing shit close real close to the vest.

3070 appears to have the same bandwidth as 2080 and RX 5700 XT. And it performs like a 2080 Ti. So Navi now has to catch up "2x". It has to get 50%+ more efficient with its bandwidth.

That seems extremely unlikely to me.

 

[CarstenS]

3070's got an ace up its sleeve.

 

[Bondrewd]

That seems extremely unlikely to me.

Everything about AMD's new GPU IP is in extremely unlikely tier yet it is real.
Being fair, not as extremely unlikely as was AMD sticking the boot in mobile yet it also happened.

 

[Erinyes]

To be the fair I believe the belief is that the PS5 has had so many revisions / respins on its silicon to reach that clock speed it didnt start like that.

Except respins are typically required to fix bugs and not for clock speed reasons (a notable exception being the ATI R520 GPU, though that was also caused due to a bug). Clock speed is more a function of the design, process and physical implementation. And I highly doubt the PS5 APU has a significantly different one than XSX. MS chose a fixed clock and Sony seemingly went higher up the curve and with some boost. Why the disbelief about 7nm clockspeeds? Look at Renoir desktop clocks.

3070 appears to have the same bandwidth as 2080 and RX 5700 XT. And it performs like a 2080 Ti. So Navi now has to catch up "2x". It has to get 50%+ more efficient with its bandwidth.

That seems extremely unlikely to me.

Has anyone ever tried downclocking the memory of either 2070/RX5700 and seeing how much the performance drops? I'd wager its not linear. Conversely, memory overclocking never gets you a linear performance increase. While it may be a bit short on bandwidth with a 384 bit GDDR6 memory bus, I dont expect it to have a significant impact on the performance of Big Navi.

 

[Bondrewd]

RX5700 and seeing how much the performance drops?

5600XT is the living example.
6% less perf for 25% memory b/w chopped off.
Literally made 5700 obsolete (unless you need extra 2GB VRAM).