New AMD GPUs to launch July 7th.

[Pemalite]

EricHiggin said:

I really, really, like the look and styling of the card. Hopefully that blower ain't so loud with the RDNA efficiency gains.

Speaking of which, I wonder how long before the meme's show up pointing out the card 'runs so hot that the shroud partially melted on top'?

I would have rather AMD kept it square and simple. The dent on top looks bad.
Have to say... nVidia's styling is a step up over AMD's GPU's.

Looks like I am waiting for a card with an aftermarket cooler!

Ganoncrotch said:
Crazy to think that you could buy a 5700XT and the Ryzen 9 for the same price as the i9 which the Ryzen is rivaling, I mean they should have made it clearer I think that you could get pretty much the entire Ryzen PC built for less than the cost of an intel CPU, although I think if you're building a rig with that Ryzen 9 in it you're not going to be looking for 2070 levels of performance, I could be wrong.

Great showcase though, I love AMD's CEO just her cheeky grin when she reveals shit that she knows gamers are gonna love, she's very charismatic.

That was AMD's tactic during the bulldozer days actually... Where AMD marketed that you could buy an AMD CPU+GPU and get more performance than an Intel+nVidia equivalent at any given price point.

I guess they would have sussed out whether that campaign worked for them or not... And all things considered AMD is smashing it out of the park on the CPU front currently anyway.

fatslob-:O said:

RDNA has a SIMD32 and a SIMD64 mode which corresponds with their dual scalar units and is somewhat similar in concept to what Intel has been doing a long time with variable SIMD widths on their GPUs ... (Intel has SIMD32/SIMD16/SIMD8/SIMD4x2 modes)

As for primitive shaders, I'd be interested in seeing an API exposed for it to see what it's programming model is actually like as an interesting comparison to Turing's mesh shaders. "Primitive shaders" are still a massive mystery since we still don't if it's truly equivalent to Turing's mesh shaders or something completely different ... 

Really? From what I can tell AMD increased it to SIMD32 from SIMD16 on GCN and decreased wavefronts from 64 on GCN to 32.
So no longer does it take 4 cycles for a wavefront to be pushed through on GCN, it's now 1 on Navi, if they have a SIMD64 mode, that is news to me.

The Primitive Shaders are handled by the drivers compiler. I would assume developers have some aspects they can leverage from it.

Bofferbrauer2 said:

Even without any architectural improvements, we'd be almost there already from just the higher clock speed and 4 additional CU.

In fact, if you just make a simple CU*clock speed calculation, you'll get actually very close to Vega 64 (air) at boost clock, and still surpassing Vega 56 just at gaming speed. So without even any RDNA improvements, it's sitting right between Vega 56 and Vega 64 in terms of performance and far above RX 590.

So yes, I think Navi will replace both Ellesmere (RX 570-590) and Vega except VII. Baffin (RX 460/560) and Lexa (RX 540-550) are probably still getting used as low-end/entry solutions. Especially Lexa, as it just months ago finally replaced the dreaded Oland chips (and the age-old cap verde chips in mobile) in mobile and entry-level desktop OEM.

And I would be delighted if they could adapt it into a high-end Laptop GPU. It's high time AMD gets back into that market.

37.5% alone is from the clocks+CU increase. The rest of the 50% or so should come from IPC improvements.
In certain scenario's, I would still expect Vega 64 to win, in others I would expect Navi to win, some scenarios just love compute... And at the end of the day, Vega 64 is still a compute monster.

What should give Navi a small advantage is that it should sustain higher boost clocks than Vega 64 with some luck.

The big kicker that allows Navi to compete with Vega 56/64 is that it's no longer held back by a few hundred Gigaytes/per second of memory bandwidth, like Polaris. - GDDR6 has brought the mainstream forwards where they are competing with 2048-bit HBM2 implementations.