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

Status
Not open for further replies.
Dude on this picutere there is written 2 CUs /Dual CU and these Setup can handle 256 32b FP instructions so if you have 40 of this unit you will get 10.240 instruction.
 
Because it's not really a net win until you plop 4+ stacks.
The fastest one is 2.54Gbps which isn't really *that* fast.

Samsung is already shipping 4.2gbps HBM2E, they don't care about going out of spec. The issue is more expense than anything.

Even the power efficiency is laughable between ram types unless your on like, 15 watts or less. No one cares about saving half a watt otherwise.

But you know, maybe there's 2 big Navi versions. There's 2 for this last gen, one with HBM and one without, no reason their can't be 2 versions upcoming and it would explain both commits being there.
 
XSX does the funny 7WGP per SA config that DT doesn't do and scales alright.

We have no idea how XSX scales. What we know:
  • AMD GPUs have scaled abysmally toward the high end in the past
  • We don’t know if or how AMD have fixed this for RDNA2
So basically, we do we really know about RDNA2:
  • Nothing except the DCUs are identical at a high level to RDNA1

All this hype from random people who claim to have “sources” is giving me serious flashbacks to Vega. At least this time the false expectation isn’t AMDs fault.
 
We have no idea how XSX scales
Directly with ALU count while being overweight, yeah.
You see, they're scaling out, not up here.
Issues only arose in Fiji when they've plugged 16CUs per SE and it didn't work at all.
Everything before up to and including Hawaii worked nicely.
All this hype from random people who claim to have “sources” is giving me serious flashbacks to Vega
Oh lawd, muhvega!
At least this time the false expectation isn’t AMDs fault.
Half of muhvega.
Damn, unlucky me.
(yes, you've figured the reason of them sitting silent out)
 
Last edited:
If only N10 didn't have 64 of those.

Quite literally >50% more bandwidth over N10, and that's just offchip.
They've just redesigned the RBEs for the first time in ???? years.
Definitely >8.

I have to say, I do admire your confidence in this product. Here’s an architecture that looks, by block diagrams, to be almost identical to the one that came before. And there’s the complete silence on AMDs part as well. With Zen we always knew what to expect because they told us what to expect, but here we have literally nothing. Further to that, we have AMDs pretty dicey track record with GPUs.

But here you are, bold as brass. Hopefully there are signs of life from RTG soon :D
 
I have to say, I do admire your confidence in this product
It's good shit.
The real question is not the IP or the chips actual, but prices, memory configs, dates and availability.
Here’s an architecture that looks, by block diagrams, to be almost identical to the one that came before
From the block diagram all nVidia GPUs look the same since Fermi yet here we are (nVidia even uses the same colours and all).
Yeah, high-level organization changed once a little bit in P100 but that's basically it.
And there’s the complete silence on AMDs part as well
Intentional.
Notice the lack of Milan/Vermeer leaks from big3 or Taiwan.
With Zen we always knew what to expect because they told us what to expect
Not past OG Zen, no.
Rome literally waltzed out of deep blue and made Intel comp analysis cry (and get screamed at. Loudly.).
Further to that, we have AMDs pretty dicey track record with GPUs.
They had a >20y long beyond dicey track record in mobile chips, where Intel won basically all the time yet here we are.
You gotta ditch the historical precedent and look at this in vacuum of spess.
Hopefully there are signs of life from RTG soon
Their marketing is shitlording on social media so expect a teaser + even date soon enough.
 
If only N10 didn't have 64 of those.

Quite literally >50% more bandwidth over N10, and that's just offchip.
They've just redesigned the RBEs for the first time in ???? years.
Definitely >8.

I’m not taking a shot at the design. I’m just saying you don’t add rops if you don’t have bandwidth for them. Maybe 64 rops is well suited to 750 GB/s etc.
 
Status
Not open for further replies.
Back
Top