Ryzen 7 5800x3d but w/GDDR6 system memory?

[magellan]

AMD already manufactures 8 core consoles that use GDDR6 memory, but only 16 GiB. Would it be possible for AMD to manufacture a motherboard/CPU combo that has soldered GDDR6 memory (maybe with differing amounts of GDDR6 memory) but socketed CPUs? DDR5 still doesn't approach the bandwidth of GDDR6 right?

 

[mackerel]

I think the difficulty in wiring means if it happens at all, it'll be soldered CPU + soldered ram. Even if the ram is soldered, having the CPU on a socket would be horrible to manage, but not impossible.

Keep in mind the bandwidths are primarily for the benefit of the integrated GPU, not the CPU. Beyond a point it is effectively unlimited for the CPU cores.

I'm not too familiar with GDDR myself, but the PS5 implementation for example is supposed to be on 256-bit bus. This compares with a typical dual channel consumer mobo reaching 128-bit. So bus width doesn't by itself explain all the difference. Looking it up, even though it is called GDDR, it is actually QDR. So at comparable clocks it should be roughly 4x DDR as we know it.

I have been wondering if the PC space is overdue moving to a unified memory model. The problem with it is that what we expect as upgradable now will unlikely be possible. Not a problem in laptop use cases, but perhaps a harder sell for desktops. I saw speculation that AMD may be bound not to compete with a similar offering to console chips so we might not see this from them. Intel with Arc has the option to do something like that in future.

 

[magellan]

@mackerel
With your x299 do you get even more bandwidth than the highest overclocked dual channel DDR5 rigs? Did you tune your memory for the least amount of latency vs. bandwidth (since I figure you don't need any more bandwidth)?

 

[Woomack]

Quad-channel DDR4 ends at about 110GB/s. This is only when you use good RAM and a motherboard with optimized BIOS. Typical 24/7 results are closer to 95GB/s and 55-65ns latency (looking at AIDA64 results). Actually, results in benchmarks don't change much above 3600 on X299, so even in 3D world records, RAM was at 3600 CL14.
Dual-channel DDR5-8000 at XMP settings has 120GB/s and 60ns latency. DDR5-6800 at XMP should make around 100GB/s and 60ns (the best DDR5-6400 kits too).

In most tests or games, cache speed and access time make a bigger difference than RAM bandwidth or its latency. For some generations, memory controller, cache speed, and additional factors count as much (or even more) as RAM speed.

 

[mackerel]

With your x299 do you get even more bandwidth than the highest overclocked dual channel DDR5 rigs? Did you tune your memory for the least amount of latency vs. bandwidth (since I figure you don't need any more bandwidth)?

I don't tune it for 24/7 running. Just turn on XMP and forget about it. I'm running 8x8 3000 which isn't that fast, but it is good enough. I don't even know what it benchmarks at.

For my uses, when I need bandwidth, I can't get enough bandwidth. On paper DDR5 at 6000 in a comparable configuration should match what I have now. This is in part why I haven't bothered to replace my desktop yet. Consumer CPUs of today are a bit of a sidegrade and not worth the bother.

Also I gave up on Aida bandwidth results. They don't match any real world task I run like Prime95, which strictly speaking is more a mixed read/write pattern. For something easier to run then I'd take Y-cruncher as indicative but they don't behave the same.