MoneyWhat would be the challenges to create a next-gen machine using CPU & GPU chiplets?
Money
Really? One of the touted features of chiplets is cost savings because you can use a larger percentage of the wafer thin mint.
Yes, but building a custom chiplet designs, having wide enough busses between GPU and CPU chiplets and IO is a whole another matter than a high profit server chips whichs design can be utilized across the whole CPU-fieldReally? One of the touted features of chiplets is cost savings because you can use a larger percentage of the wafer thin mint.
All true, and it might not be any cheaper initially, but these are devices intended to sell from 6-10 years, in an era where mid-gen consoles have been established.
Just looking at the Zen 2 chiplet design, and its 16nm controller, I think a chiplet design would allow the consoles to take advantage of process and node improvements sooner and cheaper e.g. "plug in" some 5nm Zen 2 cores when they're cheap enough, and use a cheaper cooling solution.
And a PS4Pro "butterfly" design would probably be cheaper than ever to R&D.
The denser libraries did place a lower ceiling on the clocks for Excavator. Whether lower clocks or the physical customization are cause or effect isn't clear to me. The shift to bulk and the process tweaks made clock declines for the CPU likely, and once that was inevitable that could have made AMD focus the implementation more on density at the expense of clocks that weren't going to be competitive anyway.I thought the high density libraries were more a trade off of clock speed for a smaller size as they switched to a different metal stack to enable greater density.
From what I've seen DP is not important for games, or most client software in general.Is this the case for games in general?
I suppose right now it's a proven concept only for cpu levels of bandwidth, which seems to be fine with low cost organic interposers?What would be the challenges to create a next-gen machine using CPU & GPU chiplets?
I suppose right now it's a proven concept only for cpu levels of bandwidth, which seems to be fine with low cost organic interposers?
GPUs need an order of magnitude more bandwdth, maybe it would require a completely different design, instead of a central IO chip being the memory controller, each chiplet might needs to serve a fraction of the memory and the central chip would be a crossbar of sorts? Still, it multiplies the bandwidth which has to go through the interposer. Each chiplet need as much bandwith to it's section of memory as it needs to/from the central hub.
If they manage to use organic interposers in this very high bandwidth situation, maybe it solves the same issues with the long promised HBM on organic interposers.
I'm also curious if right now the center chip is 14nm because it's main purpose is to drive a lot of powerful external PHYs, making 7nm a waste of money, so it's a great cost saving to stay at 14nm.
They say it's an active interposer, so it's silicon, so maybe we should forget about it on consoles.It's more complicated than that with several GPU-chiplets, this is a diagram of how AMD thinks they could use several GPU-chiplets and a CPU-chiplet in same MCM
and the article it came from https://spectrum.ieee.org/tech-talk...iplet-revolution-with-new-chip-network-scheme
Perhaps folks are overthinking things greatly for a $399 box.
''Mark Cerny, PS4’s lead architect, told Digital Foundry last year that the realistic limits for a next-gen console GPU would most likely top out at eight teraflops''
From that link.