AMD: Zen 4, Speculation, Rumours and Discussion

Do we know how many generations AM5 may last? I was going to go Raptor Lake, but it's EOL for the platform.
 
Do we know how many generations AM5 may last? I was going to go Raptor Lake, but it's EOL for the platform.
No, but I doubt AMD wants to ruin the good image they got with AM4 yet so I'd expect minimum of 3 gens
 
That sounds like a reasonable stretch, 3 gens. My main goal is ensuring top AI/simulation performance on poorly optimized games that rely mainly on single threads and/or fast/fat cache like Cities Skylines, Minecraft modded server hosting, Civ VI, etc. Need to sheer brute force some of these, unfortunately...
 

Yes fellas, it's ~30% 1t uplift gen over gen.
Only 2x of what they've said during their FAD.
:^)
A significant improvement for sure.
Those are some amazingly high clock bumps o_O

Can it be? :unsure:
Finally a legit successful 'speed demon' architecture :runaway:

Obviously a TDP increase there but will it actually spend most of its time up near that limit?

And managed to do it with decent IPC gain, even adding AVX512 (double pumped AVX256 per Papermaster, always seemed like the more obvious route vs Intels dedicated hardware)

7800 not at launch, presumably will be specced based on how Intel next gen works out & if as rumoured X3D


Kinda amazing times we live in with the supposed tail end of Moores' law and yet these clock/feature bumps each generation seemingly making bigger & more impressive changes each time.
Long gone the 'Mhz wars' times of a 50Mhz bump & adding/removing a few ALU stages.
 
And managed to do it with decent IPC gain, even adding AVX512 (double pumped AVX256 per Papermaster, always seemed like the more obvious route vs Intels dedicated hardware)
Considering the disjoint cluster fuck that is the feature set of AVX-512 ( F, CD, ER, PF, VL, DQ, BW, IFMA, VBMI, 4VNNIW, 4FMAPS, VPOPCNTDQ, VNNI, VBMI2, BITALG, VP2INTERSECT, GFNI, VPCLMULQDQ, VAES) in six different micro architectures (Knights Landing, Skylake X, Cannon Lake, Knights Mill, Ice Lake and Tiger Lake) I think AMD opted for the lowest feature set that is actually used and is the most feasible to implement.

The most important ones are varying precision arithmetic for AI benchmarks/image processing, and vector INT64 division for that all important 3DPM 2.1 benchmark that shows AVX-512 being 7x faster than AVX-256. I'd imagine gather/scatter to just be micro coded loads and stores.

Cheers
 
One article somewhere suggests you could use AVX512 on the new Ryzen CPUs to boost Raytracing performance in CPU limit.

Is that true? Is AVX512 a good fit for RT workloads in games? Given we have seen some titles with RT being the limiting factor in CPU limit like Spiderman, that would be a killer feature of the new Ryzen CPUs.
 
AVX-512 is a massive advantage in some emulators as well. I hope AMD's implemented version of it works for those.

Zen4 looks like a seriously good CPU. Even though I'd planned to hold on to my 3700x for a good portion of this generation, if not the entire generation, I'm now sorely tempted to upgrade to AM5 with a Zen4.

I will certainly wait to see what Raptor Lake brings though given it should launch only about a month after Zen4 releases. With "only" an 11% average improvement over Alderlake I can see Raptor lake having a chance of taking the crown back again, albeit likely at a much higher power draw (which is ironic give it's heterogenous design).

And then I'll likely want to wait for the 3D cache based 7800x too in the hopes that can leapfrog Raptor Lake in gaming performance again.
 
AVX-512 is a massive advantage in some emulators as well. I hope AMD's implemented version of it works for those.

Zen4 looks like a seriously good CPU. Even though I'd planned to hold on to my 3700x for a good portion of this generation, if not the entire generation, I'm now sorely tempted to upgrade to AM5 with a Zen4.

I will certainly wait to see what Raptor Lake brings though given it should launch only about a month after Zen4 releases. With "only" an 11% average improvement over Alderlake I can see Raptor lake having a chance of taking the crown back again, albeit likely at a much higher power draw (which is ironic give it's heterogenous design).

And then I'll likely want to wait for the 3D cache based 7800x too in the hopes that can leapfrog Raptor Lake in gaming performance again.


I think AMD showing 32% IPC improvement at 4GHz vs Ryzen 5000 in Dolphin might point in that direction ;)
 
And managed to do it with decent IPC gain, even adding AVX512
I missed doubling L2 from 512K to 1MB/core.
When they showed that 5.5Ghz preview I was assuming they'd stuck with 512K to make those kinds of clock gains, having previously dropped down from 1MB/core Bulldozer -> Zen1.
 
Considering the disjoint cluster fuck that is the feature set of AVX-512 ( F, CD, ER, PF, VL, DQ, BW, IFMA, VBMI, 4VNNIW, 4FMAPS, VPOPCNTDQ, VNNI, VBMI2, BITALG, VP2INTERSECT, GFNI, VPCLMULQDQ, VAES) in six different micro architectures (Knights Landing, Skylake X, Cannon Lake, Knights Mill, Ice Lake and Tiger Lake) I think AMD opted for the lowest feature set that is actually used and is the most feasible to implement.
Zen4 implements at least: AVX-512F, VL, BW, CD, IFMA, DQ, VPOPCNTDQ, BITALG, VNNI, VBMI, VBMI2, BF16

That is, they have everything that's supported by Cannon Lake and Cooper Lake, and some added in Ice lake, and are completely excluding the weird Phi crap. This doesn't look to me like supporting a minimal subset, but attempting to support everything that Intel does on their new cpus, and only coming up short on the very new stuff, which probably didn't make it in time.

Of course, it's possible that are the very complex stuff gets a half-assed microcode implementation.
 
Back
Top