AMD Radeon 400 series is based upon 'Polaris' microarchitecture

[HEXUS]

And AMD's Raja Koduri recently Tweeted about this '2.5x brighter' North star.

Read more.

[61zone]

hopefully these will be available in Q4 2016

[CAT-THE-FIFTH]

Looking forward to it - I have been with Nvidia for nearly three years now and it would be nice to have a change.

[kalniel]

There's a star/thermal properties joke to be made here

[Corky34]

So is this just a rumored renaming of arctic island?

[lumireleon]

just give us anything faster than 950ti but costs same as it

[DanceswithUnix]

So is this just a rumored renaming of arctic island?

Sounds more like a replacement for CGN. Sort of GCNN?

[watercooled]

That's really weird, only back in the summer when I was playing KSP a lot I was saying celestial objects would be a great idea for GPU names (going on AMD's existing logic naming after Evergreen trees, groups of islands, etc), I think I may have even mentioned Polaris for the flagship!

Granted that's not exactly what's happening, it seems more like a name for the GCN replacement as you say.

[Corky34]

Sounds more like a replacement for CGN. Sort of GCNN?

It would be a major shock if it was, AMD’s Chief Tech Officer said (YouTube Link) back in may that there's been two generations of GCN with a third coming next year.

[DanceswithUnix]

It would be a major shock if it was, AMD’s Chief Tech Officer said (YouTube Link) back in may that there's been two generations of GCN with a third coming next year.

Sounds about right then.

It takes a long time to get a new architecture right. Plenty of time for a new implementation of the old GCN architecture.

[Corky34]

So it is a renaming of arctic island then?
I'm confused as previously we were lead to believe arctic island was due in 2016 with a rumored 2x increase, now this Polaris is also rumored for release in 2016 with a 2.5x increase.

[watercooled]

A ground-up replacement architecture would be unusual considering that happened with Terascale>GCN a couple of years ago, but perhaps AMD are changing how they market their architecture revisions, in which case it would be fair to call GCN 1.x > GCN 2.0/Polaris (assuming that's what's happening) a 'replacement' TBH. For comparison, Nvidia have marketed all of their architecture revisions with names rather than just numbers, but it's the same story at the end of the day - Maxwell is an iteration of Kepler, and Kepler an iteration of Tesla, the core architecture being very similar between them with changes in things like functional unit ratios, layout, cache size, etc. But it's really splitting hairs over semantics at the end of the day.

Edit: @Corky, I'm assuming, like with AMD's previous generations the group of processors is known collectively as Arctic Islands, with Polaris as the uArchitecture name. Like how HD5000's architecture was Terascale 2, but the family name was Evergreen; or HD6000 as Terascale 3 and Northern Islands, and so on.

[Corky34]

@Watercooled, That kind of makes sense but what doesn't is that previous rumors suggested Arctic Island would have three uArchitecture names, Greenland, Ellesmere and Baffin, and if IIRC Polaris is a trinary star system, coincidence?

[DanceswithUnix]

But it's really splitting hairs over semantics at the end of the day.

The proper semantics are fairly clear cut, until the marketing people start abusing the language. The definitions are in any decent processor architecture text book.

An architecture is the high level programming abstraction, the definition of how you write code to run on it. The bits of silicon that we buy are "implementation" of an architecture, possibly to a given revision of the architecture but changing architecture on every generation makes for massive software effort so you do that sparingly.

So, ARMv8 is an architecture, AMD64 is an architecture. Zen is not an architecture, it is an implementation of AMD64, it shares the same AMD64 architecture as Piledriver (same number of registers, same instruction set with same format for instruction set extensions).

If that is a techie slide, then it means a complete new shader instruction set. If it was a marketing slide, then throw it away as it doesn't mean a thing anyway.

[watercooled]

I think Greenland, Ellesmere and Baffin are the chip names, like how Pitcairn, Bonaire, Tahiti, etc were for HD7000 (Nvidia's equivalent being GM200, GM204, GM206 for Maxwell, for example). Of course I'm still making assumptions based on rumours, it just seems to line up with what they've done previously.

AMD and Nvidia seem to do things the opposite way round in that regard - Nvidia name their architectures and give their actual chips numerical names, whereas AMD have been giving their architectures numerical names and their chips names.

[watercooled]

The proper semantics are fairly clear cut, until the marketing people start abusing the language. The definitions are in any decent processor architecture text book.

An architecture is the high level programming abstraction, the definition of how you write code to run on it. The bits of silicon that we buy are "implementation" of an architecture, possibly to a given revision of the architecture but changing architecture on every generation makes for massive software effort so you do that sparingly.

So, ARMv8 is an architecture, AMD64 is an architecture. Zen is not an architecture, it is an implementation of AMD64, it shares the same AMD64 architecture as Piledriver (same number of registers, same instruction set with same format for instruction set extensions).

If that is a techie slide, then it means a complete new shader instruction set. If it was a marketing slide, then throw it away as it doesn't mean a thing anyway.

Yeah that's what I mean about the ambiguity - what one company's marketing team calls an 'architecture', 'micro-architecture' or whatever else is probably different to another company's interpretation. Zen, Haswell etc being micro-architecture where AMD64 is the instruction set architecture, the terms are used more properly on the CPU side.

But on the GPU side it's a bit more hazy because there isn't really that sort of relationship - there isn't really an 'x86' equivalent on GPUs with GPUs from each company having their own instruction sets abstracted by drivers. The closest approximation probably being that e.g. GCN is an architecture, GCN 1.0 being a microarchitecture. They're not CPUs the terminology was designed for at the end of the day.