AMD RDNA 2 graphics card clues found in ROCm firmware update

[HEXUS]

Files indicate RDNA 2 Sienna Cichlid will sport 80 CUs, and Navy Flounder 40 CUs.

Read more.

[3dcandy]

Rumours of much less bandwidth are true then... but 80 cu's

[lumireleon]

Gddr6 or gddr6x ??

[Gentle Viking]

Good news,,, well sort off, at my age a month pass by like ( snaps fingers )
And worst of all, this too will probably trigger a really bad case of shopaholic even if i am not going to buy a GFX anytime soon.

Or that is, if AMD do a trade in, then i might have to say BB to my Red devil 5700 XT and somehow sell the unused alphacool cooler for it too. :-)

[3dcandy]

Gddr6 or gddr6x ??

GDDR6 non X - NV grabbed all of that after AMD bagged TSMC 7nm

Also it seems that the GDDR6X is one of the main reasons for higher power useage and heat

[Tabbykatze]

Rumours of much less bandwidth are true then... but 80 cu's

It depends if they've managed to do something like what Nvidia has done their new memory controllers by being able to do 4 signals per clock instead of just 2.

Else there is a big issue with memory starvation as Vega was super memory hungry, that's another reason why it really needed HBM.

To have a huge big "game changer" card then starve its memory will be nuts...

[3dcandy]

It depends if they've managed to do something like what Nvidia has done their new memory controllers by being able to do 4 signals per clock instead of just 2.

Else there is a big issue with memory starvation as Vega was super memory hungry, that's another reason why it really needed HBM.

To have a huge big "game changer" card then starve its memory will be nuts...

Supposed to have a massive cache on board to mitigate it as well though

[ik9000]

To have a huge big "game changer" card then starve its memory will be nuts... just like AMD (sadly)

fixed that for you.

[kalniel]

It depends if they've managed to do something like what Nvidia has done their new memory controllers by being able to do 4 signals per clock instead of just 2.

Not just memory controllers. New variant of memory. Expensive and seemingly not great on power.

Else there is a big issue with memory starvation as Vega was super memory hungry, that's another reason why it really needed HBM.

To have a huge big "game changer" card then starve its memory will be nuts...

We'll see. The Xbox chip is at 320bit bus. Big Navi could be on 512bit. Or there could be a caching scheme, or compression.

[badass]

Supposed to have a massive cache on board to mitigate it as well though

Where did you see that?

[kalniel]

Where did you see that?

Some of the very sketchy rumour places had a block diagram that mentions HBM alongside GDDR6, which no-one really understands, along with a couple of supposed code references to HBM. Throw in a bit of fear that they're going with a 256bit bus and it becomes a bit hopeful that they have something to mitigate, leading to guesses that the HBM might be a cache.

[3dcandy]

Where did you see that?

Couple of those American tech sites with "leaks" 128mb cache been mentioned on a few

[badass]

Have either of you got links? I've probably got better ways to spend my time but I am curious what the odds are that it's not hobbled by memory bandwidth. One thought I've had is that perhaps the firmware parameters can be interpreted another way and actually the cards are 512 Bit/384 bit. AMD have released an (underwhelming) 512 bit card in the past. IIRC it was a similar die size so enough room for the pads for memory connection.

For some reason, 28th October seems a long way away. Even though I'm not buying anything this year :-S

[kalniel]

I wouldn't do them the SEO of linking them from Hexus to be honest, far too unsubstantiated. Rumour aggregators like WCCF and Videocardz have most of them, also threads on beyond3d (which is deserving of a link: https://forum.beyond3d.com/threads/a...on-2019.61042/) have posters who pass on various twitter posts, chiphell articles etc.

[Xlucine]

Ooh, so no rebranded 5700XT? AMD offering a next-gen GPU in the space under the 3070 could be a good move, there's little point in the 3XXX cards shown so far unless you're running a 4k display so there's a big unserved market

It depends if they've managed to do something like what Nvidia has done their new memory controllers by being able to do 4 signals per clock instead of just 2.

Else there is a big issue with memory starvation as Vega was super memory hungry, that's another reason why it really needed HBM.

To have a huge big "game changer" card then starve its memory will be nuts...

That's a GDDR6X thing, not an nvidia thing.

And on the subject, nvidia is lying about memory bandwidth - if the signalling can't drive the voltage hard enough to go 0->3 or 3->0 in one clock then you're losing 12.5% of the entropy in possible signals and therefore 12.5% of the bandwidth. It's got the compression tech to avoid these, but compression doesn't get you more entropy

Some of the very sketchy rumour places had a block diagram that mentions HBM alongside GDDR6, which no-one really understands, along with a couple of supposed code references to HBM. Throw in a bit of fear that they're going with a 256bit bus and it becomes a bit hopeful that they have something to mitigate, leading to guesses that the HBM might be a cache.

It wouldn't be AMD's first foray into tiered RAM (or pseudo-RAM) on a GPU:
https://www.anandtech.com/show/10518...2-ssds-onboard

There was also a vega version in 2017. I wonder if there was much crossover with this and the storage access tech in the new consoles?

Have either of you got links? I've probably got better ways to spend my time but I am curious what the odds are that it's not hobbled by memory bandwidth. One thought I've had is that perhaps the firmware parameters can be interpreted another way and actually the cards are 512 Bit/384 bit. AMD have released an (underwhelming) 512 bit card in the past. IIRC it was a similar die size so enough room for the pads for memory connection.

For some reason, 28th October seems a long way away. Even though I'm not buying anything this year :-S

It could be another 290 (itty bitty chip matching a much larger nvidia die), which leads to the obvious conclusion of another 295X2 destroying every card for years after launch

[Tabbykatze]

That's a GDDR6X thing, not an nvidia thing.

GDDR6X was developed between a very close collaboration between Micron and Nvidia for Nvidia and is a non-JEDEC standard.

It's as close to an Nvidia thing as it can get without being developed solely by Nvidia.