Rumor:PS5 & Anaconda Scarlet GPU on par with RTX 2080, Xbox exclusives focus on Cross gen, Developer complain about Lockhart.UPDATE: Windows Central said Xbox Anaconda target 12 teraflop

[Trumpstyle]

Bofferbrauer2 said:

Trumpstyle said:

Yep probably even higher, I always assumed Sony/Microsoft will use 7nm EUV because of my estimated die sizes, I ran the numbers for 64CU's:

251mm2 (Navi10) + 54.36mm2 (24CU) + 20mm2 (ray-tracing) + 40mm2 (cpu) + 32.64mm2 (2x 64-bit controllers) + 50mm2 (wasted die area) = 448mm2 on tsmc 7nm.

EUV reduces die area by 20% but I don't think it will have perfect scaling so I used 15% instead.

448mm2 * 0.85 = 380.8mm2 on tsmc 7nm EUV.

Ofc it could be smaller depending on how much wasted die area Microsoft can reduce, Xbox one X had about 50mm2 wasted die area and PS4 pro 70mm2 wasted die area, going from 16nm to 7nm might reduce wasted die area, I don't know.

What's interesting is that his TF number for Xbox series X is almost certainly a Devkit number and not retail version, the verified insider says PS5 has higher TF than Xbox series X but the retail version might be lower, all from 10.1-11.2TF as the memory bandwidth for PS5 just isn't there for 12TF.

I don't think they will make one monolithic chip with CPU and GPU in one. More likely they will put CPU and GPU under one hood (heatspreader) but keep them separated chips to save on costs.

Also, how did you get to 40mm² for the CPU? A Zen 2 chiplet, which is pretty much what would get used in the consoles and the part used in an APU, is 80mm². 40mm² would only result into 4c/8t unless you'd remove most of the L2+L3 caches - but that would drastically slow down the chip. You might have confused the size of a chiplet with the one of a CCX, of which there are 2 in a chiplet.

A similar point could be made for the I/O, which is a bit more than just the 64-bit controllers. For reference, the I/O chip in Ryzen, while still produced in a 12nm process, is still 125mm² and has over two thirds of the amount of transistors of a Zen 2 chiplet. Of course much of that is unnecessary on a console, but not all. Expect more like 40-50mm2 for the entire I/O than just 32mm².

The wasted area is mostly over-provisioning. In this case for instance, instead of having all 64CU working, they could only have 56-60 of them running (at a higher clock speed to compensate) and have the rest as reserve to allow even not fully functioning chips to be used for the consoles, drastically increasing yield. With 64 working CU that would certainly mean 72 CU in production and a chip die larger than the one of the Vega VII, which was already 336mm²... in 7nm I might add.

As you can see, we would be pushing close to 500mm² by now, way to large for a single chip in a console. Putting CPU and IO into a different chip would make it much cheaper in production and keep the sizes below 400mm².

The GPU (Navi10) already has a I/O die, there's no need for another one. As for the CPU, 2x zen2 CCX are 70mm2, you cut off 24mb of the GameCache it should land around 40mm2, and I think the devkit has all 64CU's activated so Microsoft will simply disable 4/8 CU's and up the clock-speed to hit 12TF. There's no 72CU's.

Wasted die area, what I mean by that is you can't pack everything into a perfect Square/rectangle, here's a die shot of Xbox one X:

Even though we can't see the I/O stuff, you can see the CU's, the 32-bit controllers and the 2 CPU clusters aren't packed into a perfect Square/rectangle that's what I mean by wasted die area. You will always have wasted die area.

Edit: Forgot about double chip, Microsoft has already shown a monolith die in there E3 video, it will be a APU/SOC.