Nvidia GTX 1180 Founders Edition to arrive in July, claims report

[HEXUS]

Launch will be followed up by third party 'Turing' graphics cards in August or September.

Read more.

[lumireleon]

is it a new architecture or Pascal on steroids? (pascal++)

[3dcandy]

New architecture. GDDR6 memory as said.

[DanceswithUnix]

So, 1180 is basically the 1080ti spec but with GDDR6 ram?

Edit:

New architecture. GDDR6 memory as said.

That's just the memory controller. Well I say "just", it's a lot of work to get a good memory controller, but it could be plugged onto the existing 1080ti silicon design. But then if they were doing that, I would expect them to shrink the design to 12nm at the same time.

[3dcandy]

Lots of tweaks I guess, yes 12nm and a new memory controller. More CUDA cores expected and more TMU's and ROP's so should be a pretty big jump. Bit like Volta but I guess nvlink 2 will be removed and won't use HBM2 memory. I say new architecture because it's different enough to not be just a tweaked design.
Although I also guess lower end models with ship with GDDR5(x) and culled parts to fit budget

[DanceswithUnix]

Lots of tweaks I guess, yes 12nm and a new memory controller. More CUDA cores expected and more TMU's and ROP's so should be a pretty big jump. Bit like Volta but I guess nvlink 2 will be removed and won't use HBM2 memory. I say new architecture because it's different enough to not be just a tweaked design.
Although I also guess lower end models with ship with GDDR5(x) and culled parts to fit budget

But we just don't know the architecture. It could be a cost reduced 1080ti for all we know, or the cores could be something completely new.

I haven't looked at the DGGR6 spec, if it is close enough to GDDR5X then perhaps the existing memory controllers work just fine and this is a re-brand of the 1080ti. Not that graphics companies ever do re-brands

[Ozaron]

I mean, it's not like all NVidia needs to do is lower operating temperatures, raise efficiency, add a few more cores per tier and implement fancy new memory. Why try harder. I'd be somewhat confused if they actually redesigned architecture (or, more accurately, released and sold a new architecture) while Radeon is providing less of a fight than a wet paper bag.

[3dcandy]

I mean, it's not like all NVidia needs to do is lower operating temperatures, raise efficiency, add a few more cores per tier and implement fancy new memory. Why try harder. I'd be somewhat confused if they actually redesigned architecture (or, more accurately, released and sold a new architecture) while Radeon is providing less of a fight than a wet paper bag.

To be fair I don't think they need to. We are again at the point that games are still driven by consoles to some extent and pushing great 4k for example is not there yet. I could point the finger at Intel who have sat on 4c8t designs for far too long but I'm kinda of the thinking that we are pretty much ok as we are in many respects. AMD have other fights on their hands - the cpu side is resurging right now and I think they have been in the doldrums for so long that is taking a great amount of their energy

[Corky34]

But we just don't know the architecture. It could be a cost reduced 1080ti for all we know, or the cores could be something completely new.

Yes we don't know but we can probably make a good guess that it's Volta, TSMC have said (iirc) that they stated volume production of Volta GPUs at the end of 2017 and that ties in with the release of Titian V and Quadro GV100, I'm guessing these will be down binned versions of those.

Having said that Volta isn't exactly a gaming orientated design what with tensor "cores" (does real time raytracing make use of tensor cores?), NVLink (same as Pascal), and HBM (same as Pascal) so they've adapted previous non-gaming designs for gaming in the past.

[lumireleon]

AMD Zen @ 12nm is new architecture? what do we mean by architecture?

[DanceswithUnix]

AMD Zen @ 12nm is new architecture? what do we mean by architecture?

Architecture is what something does. Basically the instruction set it runs.

AMD64 is an architecture, Zen is an implementation of the AMD64 architecture, 12nm is a process that the implementation is made on. It's a fairly simple hierarchy which the marketing types like to try and make sound more impressive by mixing it up.

If you look at ARM chips it tends to be clearer. You want an ARMV7 architecture cpu, so you choose an A9 core, and you get it implemented on a 28nm process to keep it nice and cheap. You want performance? Then choose an ARMV8 architecture, the A72 implementation is pretty quick, and get built on 14nm.

Edit: Graphics chips are more fluid than CPUs when it comes to their instruction sets as only the drivers need to know. That tends to push a new architecture every generation of graphics chips, even if heavily derived from the last.

[Corky34]

It's more than just the instruction set it runs, it's a combination of the ISA and microarchitecture, sort of like how you can drive all cars by only knowing one set of instructions but some come with 4 or 5 gears, some have bigger engines and go faster, some can go round corners quicker, etc, etc.

[Zak33]

i think we're going to see some odd memory configurations - slightly less ram on some cards, slightly higher ram speeds but cut back memory controller speeds. We've had it before.

[CAT-THE-FIFTH]

Its most likely a further evolved Pascal but with tweaks to run ray tracing better since games like Metro:Exodus will use it.

Also since 12nm is denser,and a 256 bit GDDR6 memory controller will be smaller than a 384 bit GDDR5X one the gpu size should drop further.

[DanceswithUnix]

It's more than just the instruction set it runs, it's a combination of the ISA and microarchitecture, sort of like how you can drive all cars by only knowing one set of instructions but some come with 4 or 5 gears, some have bigger engines and go faster, some can go round corners quicker, etc, etc.

5 or 6 gears would just be an ISA extension. Bigger engine is an implementation detail. Cornering faster is a result of implementation. Architecture is the basics: wheels, 3 pedals in this order, steering wheel. People like to confuse the boundary though, in cars people often even now call the accelerator pedal a "throttle", despite no diesels and increasingly fewer petrol engines actually having a throttle valve (I gather BMW still have that valve, but generally the ECU doesn't use it and it certainly isn't connected to any pedal).

Microarchitecture is an X86 thing, because no-one else has an instruction set so messed up it needs translating into micro-ops to run on an internal RISC core. Graphics chips certainly can't afford that sort of overhead.

[kompukare]

Also since 12nm is denser,and a 256 bit GDDR6 memory controller will be smaller than a 384 bit GDDR5X one the gpu size should drop further.

Is it though? Denser that is. I did this calculation on another forum:
TSMC's 16nm to 12nm doesn't seem to offer any major density increases. A quick comparison of GV100 and GP100 shows around 3% increase in density

And since different sections would have different density, we don't even know if that 3% is because GV100 has more cache or something.
I also consider that it's unlikely that Nvidia will find lots of low-hanging-fruit like they did Kepler<>Maxwell by removing some compute features or going somewhat tile-based.
Ergo, since driving new sales requires a speed gain (min 20% I guess), Nvidia may have no option but to do what they'd rather not: increase die sizes. Since GP104 is only 312mm² and GP102 471mm², they have plenty of scope to do so.