Phil Spencer said:
The video that we showed is talking about Project Scarlett. That's the focus that we have, on that console and hitting that specification. That's the console that we're talking about.
I'm not sure why people have been taking this quote to mean Lockhart is dead. To me, it sounds like the exact opposite: indirect confirmation that Lockhart exists. Why say "this is the console we're talking about" unless there was another one you could be talking about?
It doesn't change my point that you do benefit from going wider and reducing clock speeds.
I agree with that point (my posted predictions rely on it!). I was only disagreeing with the specific scenario you used as an example.
Hoping for zero defects at ~250mm2 is likely to have serious yield implications.
Yep, but both the schematic block diagram from AMD and the annotated one
anexanhume shared show only 40CUs.
Here's my Anaconda prediction now:
3 shader engines, 6 workgroup processors, 1CU pair per SE disabled - 54 total CUs. 1500MHz. 10.4TF.
If you want to quit ResetEra, just do it. You don't owe anyone anything.
Microsoft has never been shy about getting people into their ecosystems with free trials. 6 months is way too aggressive. but even 1 month trial to gamepass (two weeks feels short) could be a big selling point. ...It certainly feels like they could persuade people to buy in early by alliviating day one cost (even if they have to pay it later, little by little, to keep those games).
Oh yeah, for consumers it's very attractive how Microsoft is willing to "split the bill". My point was that their long history of subsidization, across all three platforms to the tune of billions, has shaped perception to the point that cashing goodwill back out is going to be rough. I mean, you yourself think that two weeks of playing multiple AAA next-gen games, completely for free, "feels short"!
Yes, the idea of subscriptions is to draw people in by the initial offering, then keep them around long enough for ROI. And specifically for early adopters the math probably makes sense eventually. The problem I see Microsoft facing is that their services are almost constantly on sale, or in free trial periods, etc--very few people ever seem to end up paying the "regularly" price. They're too accustomed to good deals from Xbox.
And as a generation goes on, the future-purchase value of the people coming to your platform goes down. And their price sensitivity generally goes up (or else they'd have bought in earlier). It's a far from ideal situation. Even without that issue to contend with, Netflix has still encountered problems finding any profit. And Xbox doesn't have nearly as much exclusive content.
What I see recently is Microsoft committing to massive expenditures on their gaming initiatives which require massive scale to make business sense...but I don't see where that scale has materialized (yet?). You can only prime the pump so long before the thing has to run on its own.
Of course, I'm not an expert like Microsoft has working for them, and perhaps Xbox is on the verge of blowing up. Wouldn't be the first time my predictions went awry.
i have no idea what happened here overnight but when i went to sleep we were discussing 11-12 tflops, 56-60 CUs because the Scarlett APU is huge, way bigger than the base PS4, Pro and the X1X. Some estimates are 400 mm2.
That's only 10% bigger than One X. Of course the node shrink will help oodles, but there's also been an increase in the relative complexity and size of AMD's CUs.
I did calculations based on AMD's block diagram for Navi, and it seems ~450mm^2 would be the minimum size for a 64CU part. And that's without accounting for extra RT hardware. That is, a 60 active CU chip is almost certainly off the table.
Consoles with have 3 SEs and 60 CUs. 54 active like Anex said.
My size analysis--though it is if course tentative and unconfirmed--suggests that this configuration would be bigger than the Anaconda chip we've seen. A 54 active CU chip might be possible within that constraint, but it'd have to be two SEs each with more DCUs. An entire new SE requires too much support silicon.