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

Status
Not open for further replies.
It has been discussed before. Microsoft and Sony deal with the semi-custom department at AMD. Semi-custom guys get their hands on the IP blocks when they're ready to be implemented. Of course MS & Sony just like any other AMD customer can and do give input on what they would like to see and it could affect the architecture design choices, but they're not really involved in it.

Exactly, the idea that AMD 'cooked' Navi primary for PS5 (or xbox too) is hot air though.
 
Apparently AMD has decided they're called "Work group processors" instead of "double compute units", shoulda taken the latter smh.

Regardless, I'd take the CEO's statement as meaning their continued use of TSMC is uncertain. 5nm on TSMC is really odd, very very little in terms of power/performance improvement, a lot in terms of scaling improvement. But scaling improvement doesn't translate directly to cheaper per transistor, and moving to new processes is really expensive. Besides, scaling as a limit for RDNA isn't even in sight, Navi 10 can be straight double, more than in terms of performance, for die size but they'd hit power limits far quicker. AMD doesn't need scaling unless it has cost benefits.

So at the very least it's doubtful we'll see 5nm in 2021, that it's in a PR release is probably partly a tactic to put pressure on TSMC. But it probably also mean they're considering switching to Samsung. Which makes sense, Samsung's initial nanosheet presentation was really impressive, even moreso when they stated it was going to be tool compatible with their upcoming "4nm" process. Making transitioning between nodes cheaper, faster, and smoother is becoming just as important as the node's improvements itself, and so far TSMC feels like it's having trouble juggling that with "7nm+" "6nm" and "5nm" all coming up in a similar timeframe.
 
Apparently AMD has decided they're called "Work group processors" instead of "double compute units", shoulda taken the latter smh.

Regardless, I'd take the CEO's statement as meaning their continued use of TSMC is uncertain. 5nm on TSMC is really odd, very very little in terms of power/performance improvement, a lot in terms of scaling improvement. But scaling improvement doesn't translate directly to cheaper per transistor, and moving to new processes is really expensive. Besides, scaling as a limit for RDNA isn't even in sight, Navi 10 can be straight double, more than in terms of performance, for die size but they'd hit power limits far quicker. AMD doesn't need scaling unless it has cost benefits.

So at the very least it's doubtful we'll see 5nm in 2021, that it's in a PR release is probably partly a tactic to put pressure on TSMC. But it probably also mean they're considering switching to Samsung. Which makes sense, Samsung's initial nanosheet presentation was really impressive, even moreso when they stated it was going to be tool compatible with their upcoming "4nm" process. Making transitioning between nodes cheaper, faster, and smoother is becoming just as important as the node's improvements itself, and so far TSMC feels like it's having trouble juggling that with "7nm+" "6nm" and "5nm" all coming up in a similar timeframe.
I'm not sure this is the right thread for this, since she was talking about the CPU-side, not GPU.

Regardless, I don't see TSMC "having trouble"
7nm+ is already in mass production
5nm is completely new process
6nm (which comes after 5nm) is update path meant for easy 7nm (not 7nm+) shrinks with same design rules (kinda like GloFo 14 / 12 nm)
 
I'm not sure this is the right thread for this, since she was talking about the CPU-side, not GPU.

Regardless, I don't see TSMC "having trouble"
7nm+ is already in mass production
5nm is completely new process
6nm (which comes after 5nm) is update path meant for easy 7nm (not 7nm+) shrinks with same design rules (kinda like GloFo 14 / 12 nm)

Contracts pretty much have to be for both for economic reasons, at least I'd assume. Regardless, as far as I know 5nm isn't tool/rule compatible with 7nm+ isn't compatible with 7nm, which is the entire reason 6nm even exists, to make it cheaper/faster/easier for clients to migrate. In retrospect having two overlapping nodes like 6nm and 7nm+, which have fantastically close specs and processes when viewed from a distance, doesn't bespeak a foundry that's as organized as Samsung has been presenting recently.

Samsung has an incredibly detailed public roadmap with multiple chip design tools and flows being released. That's got to inspire a good amount of confidence. Though as I stated this could just as easily be public pressure on TSMC over negotiating contracts, or both.
 
Regardless, I don't see TSMC "having trouble"

Precisely, TSMC IMO rather shapes up to be the star of 2019 - 2020.
7 nms seem a good sucessful node. Capex of TSMC is as a high as ever; they are investing a lot.

Granted the sucess of 5 nm will be a very important indicator too.
 
They keep reiterating the "scalability" (since 2016, huh). So far we have seen rather conservative Navi configurations. So where to scale compared to GCN? Ultra-mobile? Ultra-HPC?
Well, Samsung is going to use RDNA and I'm pretty sure what ever they bring would be counted as 'ultra-mobile'
 
Considering that Samsung is licensing RDNA, is it possible that AMD could be setting up a potential dual source arrangement its GPU's?

While I'm at it with speculation, isn't it possible that AMD and Samsung would partner up through the Semi-Custom devision to create an improved low power RDNA? Samsung certainly wants what AMD has, but Samsung could still be holding knowledge and IP from its GPU initiative that could prove useful.
 
Considering that Samsung is licensing RDNA, is it possible that AMD could be setting up a potential dual source arrangement its GPU's?

While I'm at it with speculation, isn't it possible that AMD and Samsung would partner up through the Semi-Custom devision to create an improved low power RDNA? Samsung certainly wants what AMD has, but Samsung could still be holding knowledge and IP from its GPU initiative that could prove useful.

That Samsung is licensing RDNA is another indication AMD could just go Samsung. The foundry owner taping out relevant designs on the process obviously has the potential to lower design costs, though there's no telling if that's part of the contract.

SoC vendors will use it in 2020 so.

Too bad they're actually shit at executing on their roadmap.
How many times did 7LPE slip?
Three?

Not likely a quick release from AMD CPUs obviously, I mean that's kind of the point of this press release to begin with.

And it's not like TSMC is any better at keeping up with Moore's law, just better at not overpromising, so far.
 
The way to think about it, we want Radeon graphics to be everywhere. Everywhere means in the datacenter, in gaming, in PCs, in consoles, as well as in mobile. Samsung, as our partner in graphics — it’s both a licensing opportunity, as well as a development opportunity, where we’re jointly optimizing our Radeon graphics for lower-power applications. Again, it’s one of these areas where having the architecture span across all those product segments is a good thing for developers and a good thing for the architecture.

https://venturebeat.com/2019/11/03/...ew-2020-will-be-a-bigger-product-year-for-us/
 
Parallelism can be gained by more CU and lowering clocks to meet wattage/thermal thresholds, etc. (in leu of moving forward 7nm+/5nm, etc.) Question is, what weighted design has the highest sustained minimums..?

Power usage in rDNA is based on the shader array, not nec the amount of compute units. How far can rDNA be scaled down..? If Samsung is passing on ARM adreno, can rDNA trickle..?
 
Status
Not open for further replies.
Back
Top