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

Status
Not open for further replies.
The past few console SoCs GPU configs weren't so far off the respective midrange offerings.
Navi10 is supposedly midrange.
Similar configurations were just conincidences, they could have just as well be bigger or smaller. It all came down to what kind of performance console makers wanted and how big of a soc they could afford. Absolutely nothing to do with similar size in certain desktop GPUs, they're designed completely separately using same "building blocks" (as the GPU portion of the console soc).
Navi 10 might or might not be midrange, but it has absolutely nothing to do with how small/big console GPUs will be. Vega was "high-end" but in reality it's top to bottom, from tiny 3 CU GPU all the way to 64 CU Vega 10 & 20

The link in the tweet is dead. What was it about?
Here you go:

amd-navi-pcb-2-20190427.jpg

amd-navi-pcb-20190427.jpg
 
Similar configurations were just conincidences, they could have just as well be bigger or smaller. It all came down to what kind of performance console makers wanted and how big of a soc they could afford. Absolutely nothing to do with similar size in certain desktop GPUs, they're designed completely separately using same "building blocks" (as the GPU portion of the console soc).
Navi 10 might or might not be midrange, but it has absolutely nothing to do with how small/big console GPUs will be. Vega was "high-end" but in reality it's top to bottom, from tiny 3 CU GPU all the way to 64 CU Vega 10 & 20


Here you go:

View attachment 3036

View attachment 3037
The RAM ballout checks out as GDDR6. This looks legit.
 
And how much power do they have to spare.
So if Navi doesn't move perf/w needle up, then it's pretty much pointless.
Why wouldn't it? I'm pretty sure every single GCN-gen has done that (when each is run at their comfortable range, not the overtuned retail clocks/volts)
 
And how much power do they have to spare.
So if Navi doesn't move perf/w needle up, then it's pretty much pointless.

The 1X, with a single bin, offers greater than 580 performance and memory with an entire rest-of-platform attached and only pulls ~175W at the plug socket under load. Granted, that took some MS mobo level magic (4Pro is about 2/3 as fast and 155W), but clearly AMD can deliver much stronger performance per watt than they do in the PC space when it's prioritised.

However good this first Navi based chip is, I fully expect them to crank it up to 250+ Watts regardless of however fractional the absolute performance gains are.
 
The 1X, with a single bin, offers greater than 580 performance and memory with an entire rest-of-platform attached and only pulls ~175W at the plug socket under load. Granted, that took some MS mobo level magic (4Pro is about 2/3 as fast and 155W), but clearly AMD can deliver much stronger performance per watt than they do in the PC space when it's prioritised.

However good this first Navi based chip is, I fully expect them to crank it up to 250+ Watts regardless of however fractional the absolute performance gains are.
And higher binned (X dev kit) hits 6.6TF.
 
However good this first Navi based chip is, I fully expect them to crank it up to 250+ Watts regardless of however fractional the absolute performance gains are.
That's what they've been doing, yes.

Still, two 8-pin connectors doesn't bode well at all for Navi 10, unless it's going with greater than RTX 2080 performance.
 
Still, two 8-pin connectors doesn't bode well at all for Navi 10, unless it's going with greater than RTX 2080 performance.

I wouldn't read too much into that. AMD reference boards sent to AIB makers have occasionally had ridiculously overspecced power supply circuitry in the past.
 
I wouldn't read too much into that. AMD reference boards sent to AIB makers have occasionally had ridiculously overspecced power supply circuitry in the past.
TU102 bringup PCB also had like, 3*8pins and more chokes than my eyes would ever bother to properly count.
 
That's what they've been doing, yes.

Still, two 8-pin connectors doesn't bode well at all for Navi 10, unless it's going with greater than RTX 2080 performance.

Eh, 7nm from TSMC should halve the power consumption, meaning either a Navi 250 watt tdp either beats out a 2080ti, or this is just some prototype board/console board. Vega by itself was, almost, very power efficient and could be at low enough clockspeeds. I'd be very surprised if they didn't fix whatever power exponential power draw problem Vega had at higher clockspeeds to produce something far more efficient. It just seems like too obvious of a target for improvement.
 
Eh, 7nm from TSMC should halve the power consumption, meaning either a Navi 250 watt tdp either beats out a 2080ti, or this is just some prototype board/console board. Vega by itself was, almost, very power efficient and could be at low enough clockspeeds. I'd be very surprised if they didn't fix whatever power exponential power draw problem Vega had at higher clockspeeds to produce something far more efficient. It just seems like too obvious of a target for improvement.

This is an evaluation board.
 
Eh, 7nm from TSMC should halve the power consumption, meaning either a Navi 250 watt tdp either beats out a 2080ti, or this is just some prototype board/console board. Vega by itself was, almost, very power efficient and could be at low enough clockspeeds. I'd be very surprised if they didn't fix whatever power exponential power draw problem Vega had at higher clockspeeds to produce something far more efficient. It just seems like too obvious of a target for improvement.
This is an odd post from you. There is no reason 7nm ”should halve the power consumption”. Compared to what?
If nothing else Radeon VII should serve as a cautionary examply. Roughly similar complexity to Vega 64, very similar power draw and frequencies that increased by just over 10%.
Now that chip was very early out of the gate, and was aimed at providing maximum computational ability, but it still serves as an example. Moving from 14nmFF to 7nmFF is probably the largest lithographic step I’m likely to see for the remainder of my life, but it is no panacea for anything really when compared to its predecessor.

The ”exponential” frequency-vs-power wall you spoke of is not primarily a feature of the chip design, but of the lithographic process and where you choose to be on the curve. Your circuit design choices and certain lithographic adjustments will help determine just where you end up in terms of reasonable frequencies, but the overall behaviour of the frequency-vs-power curve is not something you can design away. It’s a really complex balance that determine where you end up, and ultimately the market conditions may force you to move out of the sweet spot no matter how well you planned.

All that said, if there is one thing I think can be counted on, it’s that AMD (and TSMC) has gained experience with the process, and the combination of a new design, and a firmer lithographic process footing should yield better results. Better along which axis of interest is still unknown though. Maybe they decided to focus on density to drastically increase chip yield/wafer and increase their competitiveness that way.
We just don’t know. Yet.
 
This is an odd post from you. There is no reason 7nm ”should halve the power consumption”. Compared to what?

This is an official AMD slide

InstinctMI60MI25.png
 
That's not an old slide and it's based on their current 7nm products.
Whether people interpret it correctly is another matter - general statement
Are you saying that Vega 20 clocks 30% higher than Vega 10 at ISO power for the products indicated? That’s just not the case and you know it.
That doesn’t necessarily mean all that much for Navi though, that has the benefit of over two years of development since Vega, and almost a year at the node. That should count for something.

Edit: Also, in addition to development time, Navi may benefit from not having to carry the compute baggage of Vega 20, which in and of itself should lead to higher efficiency/mm2 and /Watt.
 
Last edited:
Status
Not open for further replies.
Back
Top