AMD Licensing RDNA Graphics IP To Samsung For Smartphones & More

[numacross]

I got that part. That's what some described as "hybrid" - GCN instruction set (ISA), but new hardware implementation of it (microarchitecture). But do they plan to change ISA with next iteration of Navi though? That wasn't clear.

In the tread I linked before bridgman didn't say, but I doubt that they will change the ISA drastically.

[shmerl]

Let's make a crude CPU analogy. It's like for amd64 ISA going from Excavator ("GCN microarchitecture") to Zen (RDNA microarchitecture). The amd64 ISA changed very little (added a few instructions) in this transition but the actual implementations in the microarchitectures are vastly different.

I got that part. That's what some described as "hybrid" - GCN instruction set (ISA), but new hardware implementation of it (microarchitecture). But do they plan to change ISA with next iteration of Navi though? That wasn't clear.

[numacross]

So GCN ISA, but new microarchitecture? I didn't really fully understand that comment. So the next generation (Navi 2?) will use new ISA and that new mircoarchitecture?

The ISA stays the same - it's still GCN. But the microarchitecture running this ISA changes.

Let's make a crude CPU analogy. It's like for amd64 ISA going from Excavator ("GCN microarchitecture") to Zen (RDNA microarchitecture). The amd64 ISA changed very little (added a few instructions) in this transition but the actual implementations in the microarchitectures are vastly different.

[lkundrak]

Note how "adreno" is an anagram of "radeon".

[AndyChow]

This sounds more like an IP block license where Samsung does nearly zero engineering just drag and drop it in they can probably configure how many CUs and if it has various sub units such as VCE etc... and such on their end but it should just be a plain AMD device sitting among the devices attached to the CPUs... typically companies doing this don't even get acess to all of the hardware design files just some netlists and enough VHDL or verilog to make it configurable.

Could easily be a pure patent deal to prevent from lawsuit. I.E. Samsung does 100% of the engineering, but uses some tech which is patented by AMD.

[shmerl]

The RDNA microarchitecture implements GCN ISA.

So GCN ISA, but new microarchitecture? I didn't really fully understand that comment. So the next generation (Navi 2?) will use new ISA and that new mircoarchitecture?

[numacross]

From what I read, RDNA is different from GCN, and first generation of Navi is a hybrid combination of both. Second generation of Navi will be just RDNA. I suppose it's good for mobile GPUs as well (and Samsung will be using just RDNA design).

The RDNA microarchitecture implements GCN ISA.

[L_A_G]

Just look at Qualcomm's recent attempts to really push ARM performance into laptops/netbooks/whatever they're called now. Maybe Samsung think that's a market they can push into with a decent (non-mobile) GPU. And AMD was talking about ARM SoCs a few years ago before reprioritizing Zen. Maybe they decided that they really don't want to get into the fast-moving mobile chip market. So AMD license their GPU IP to Samsung and let them muck about in that space.

Considering how AMD killed off their ARM efforts years ago and the one board they did end up shipping didn't even have dedicated graphics hardware I don't think that project ever produced anything worth trying to license out. Even if they did produce the IP but never put it into silicon that made it's way outside of their own labs that's probably aged to the point it's not really usable anymore.

Who knows, maybe it's another semi-custom job? Sort of like the Polaris-but-branded-as-Vega chip they made for Intel or the Polaris and Vega chips they've made for Apple over the last few years. After all, the semi-custom side of the business is something they've put a lot of effort over the last few years.

IIRC, the Freedreno driver was even able to use some parts of the Radeon driver (register definitions?) to get started. They share some low-level design elements.

I somehow doubt that old TeraScale register definitions would be usable on Freedreno when the oldest Adreno graphics that it supports are from 4 years after the split. I do remember that they were able to use some of the AMDGPU DC backend plumbing so that may be where the misunderstanding game from.

[ms178]

Just look at Qualcomm's recent attempts to really push ARM performance into laptops/netbooks/whatever they're called now. Maybe Samsung think that's a market they can push into with a decent (non-mobile) GPU. And AMD was talking about ARM SoCs a few years ago before reprioritizing Zen. Maybe they decided that they really don't want to get into the fast-moving mobile chip market. So AMD license their GPU IP to Samsung and let them muck about in that space.

Not only Qualcomm, but more so Apple. This decision from Samsung makes it clear that they want to get a piece of that pie as well. From AMD's perspective that decision makes sense for the short term but it would mean that they would compete with Samsung and their own graphics IP if they wanted to enter that market themselves at a later point thus building a higher barrier of entry for themselves. And I find this a bit sad, I saw a huge advantage for them being able to adress both the X86 and ARM market if that ISA takes off in the consumer desktop / laptop market. It seems that this decision abandons their ARM strategy alltogether.

[starshipeleven]

IIRC, the Freedreno driver was even able to use some parts of the Radeon driver (register definitions?) to get started. They share some low-level design elements.

They do because Adreno GPU is the evolution of an ATI mobile/embedded graphics design called Imageon. AMD sold off the entire embedded GPU division and the Imageon GPU to Qualcomm in 2009, which of course had to rename it.

Note how "adreno" is an anagram of "radeon".

The fact that AMD is back into the fray for embedded graphics is very interesting.