Question Zen2 vs Zen3 improvements for DC?

[Assimilator1]

Yea I know that's nearly old skool now! . But I don't keep my rig bang up to date anymore (if I ever did).

After noticing that my electricity charges have slightly more than doubled in 3yrs! (from 13.44p/kwh in 6/2020 to 30.3p/kwh last month! And that's with a subsidy, otherwise it'd be 48p/kwh!! [edit] Just discovered the subsidy is being phased out from this month! I hope the prices drop soon!).
Which means with a wall draw of 112w (CPU crunching only), running 24/7 the monthly cost has gone from ~£11/month > ~£25/month > ~£39/month! (£470/yr!).

I'm finally look at upgrading my Ryzen 3600 to maybe a 5600 or 5600X, IIRC they use a little less power (as well as being a little faster).
Anyone remember the difference? Looking at AT's review of the Zen3 the 5600X drew a little over 10% less power than the 3600, and Zen 3 has ~19% faster IPC than Zen2 (plus a little clock boost).
A post by biodoc showed his 5950X was ~11% faster than his 3950X in wanless (Mersenne+2), is this typical for DC?

If so I think I'll start hunting for a Zen3! I see the 5800 [edit, seems to be OEM only, but the 5700X is the same thing?] has the same TDP as the 5600/5600X, but will it draw more power overall under full load anyway? (I'm guessing that the 5800 all core load clocks will be less than the 5600X, but that it's extra 2 cores more than make up for that for ppd?).
I need to see if my mbrd BIOS is upto date for these newer CPUs (rig spec in sig btw).

 

[StefanR5R]

For several of PrimeGrid's CPU subprojects, Zen 3's larger core complexes with 32 MB last-level cache per CCX (in desktop Zen 3 flavors at least) is a big improvement over Zen 2's 16 MB sized last-level cache segments. E.g. from the projects of the 2023 PrimeGrid challenge season, PPS-MEGA, GFN-17-MEGA, SGS-LLR, 321-LLR, GFN-18, GFN-19 are still small enough to fit one or more of such tasks into a Zen 2 CCX. But PSP-LLR, SoB-LLR, ESP-LLR, and GFN-20 are too big for a Zen 2 CCX, and the resulting cross-CCX communications starves the vector arithmetic units. Zen 3 CCXes however can still go full throttle at these bigger projects.

 

[Assimilator1]

Interesting, and I'm guessing a very large difference in performance!

Looking on ebay for 2nd hand Zen 3s it seems the plain 5600s are rare (5800s non existent), most are the 5600Gs (any disadvantage to having one with onboard grx?)
The few 5600s and 5600Xs seem to go for about £85-95 (some of the latter going for £120! wth??[edit] actually, most are £120-130!), 5800Xs for about £150+...., 5700Xs prices some where in between at £130-150.

 

[cellarnoise]

Zen 4 is even more efficient and has avx-512 . Most of the desktop zen family you can set your own power use limits.

The desktop apus have less level 3 cache per core than the non-apu models.

 

[Assimilator1]

Yea but Zen 4 needs a new mbrd and RAM right? I can't afford to spend a load more atm.

APUs? Ah I guess the 5600G is an APU, I see it's L3 cache is 16MB vs 32MB for the 5600, good to know!

 

[cellarnoise]

Zen 4 does require new MB and ddr5.

We have found the entry level ASRock works well though and has been most affordable.

5950x works well power restricted all the way down to 65 watts even, though a bit better at 100w or so. Double your output!

 

[Assimilator1]

The 5950 would be great , but what would have happen to 3-4 core clock speeds? I do occasional game.
Actually, nm, having seen the prices for 5800Xs I wouldn't want to pay for a 5950!

 

[cellarnoise]

Maybe 5900 then? Better price per-core !

Looks like 5700x non-apu used might be best price per core?

 

[StefanR5R]

The desktop apus have less level 3 cache per core than the non-apu models.

Indeed; the 5600G in particular only has got 16 MB L3$. As do the 5700G and even the 5500 which is based on the same die but has got the iGPU disabled.

Edit,

I'm guessing that the 5800 all core load clocks will be less than the 5600X, but that it's extra 2 cores more than make up for that for ppd?

Yes, within the (by default) 65 W TDP/ 88 W PPT limits, the 8 core CPUs will have higher throughput than 6 core CPUs of the same generation, but I don't know exactly how much higher.

 

[Assimilator1]

Maybe 5900 then? Better price per-core !

Looks like 5700x non-apu used might be best price per core?

Lol, I'm sure it is, but probably not enough to make it cheaper

And yea the 5700X/5800 does look like the best deal, and looking at this review of the 5800 by GN it used less power than my 3600 running their Blender benchmark!

 

[drnickriviera]

Heck i'm still wondering what improvement I would get from Zen to Zen 2 in WCG? May find another 3700 to replace my 1600

 

[StefanR5R]

6c/12t Zen+ @ 65/88 W --> 8c/16t Zen 2 @ 65/88 W should bring an improvement in many all-core workloads at an order of magnitude of +100%, shouldn't it? (More cores, much wider cores especially WRT floating point units and cache, faster clocks at same power thanks to GloFo 12nm --> TSMC 7nm.)

 

[Assimilator1]

There's a big jump in performance from Zen+ to Zen 2, I think the AnandTech article I linked in my op shows something on that.

 

[Assimilator1]

Actually, it's not as much as I thought going from Zen+ to Zen2, AMD's slide in the AT article I linked in the op shows an IPC increase of 15%, and showing about a 8% increase in max boost clocks. I looked at AT's article on the Zen 2, it sees the 3700X using 32% less power than the 2700X!

 

[StefanR5R]

For all-core workloads¹ like in distributed computing, and especially for floating-point heavy and vector arithmetic heavy workloads², the improvement from Zen+ to Zen 2 is a lot larger than what IPC increase and max boost clock increase tell us.

¹) mainly due to the GloFo 12nm --> TSMC 7nm move
²) mainly due to the ≈doubled FP bandwidth/ ≈doubled AVX bandwidth per core

I never had a Zen or Zen+ CPU myself. My own main comparison is between Broadwell-EP and Rome, and specifically the 22c/44t SKU vs. a 32c/64t SKU of these. Machine throughputs are near the order of 1:2, and power efficiency at the order of 1:2.5 in many distributed computing workloads. (Sometimes a bit more, sometimes a bit less.) — As far as I recall, Zen and Zen+ perform somewhat comparable with Broadwell(-X,-EP) when driven in similar clock ranges.

The Zen 2 to Zen 3 step is quite different: The core was made quite a bit wider, but the process node stood the same. The result was that light workloads get a pretty good uplift in Zen 3 vs. Zen 2, while power-hungry workloads, notably all-core workloads, see only mild improvements (from what I remember reading; I don't have Zen 3 CPUs myself). Special exceptions are those which profit from the different CCX organization, as mentioned in post #2.

 

[mmonnin03]

For all-core workloads¹ like in distributed computing, and especially for floating-point heavy and vector arithmetic heavy workloads², the improvement from Zen+ to Zen 2 is a lot larger than what IPC increase and max boost clock increase tell us.

¹) mainly due to the GloFo 12nm --> TSMC 7nm move
²) mainly due to the ≈doubled FP bandwidth/ ≈doubled AVX bandwidth per core

I never had a Zen or Zen+ CPU myself. My own main comparison is between Broadwell-EP and Rome, and specifically the 22c/44t SKU vs. a 32c/64t SKU of these. Machine throughputs are near the order of 1:2, and power efficiency at the order of 1:2.5 in many distributed computing workloads. (Sometimes a bit more, sometimes a bit less.) — As far as I recall, Zen and Zen+ perform somewhat comparable with Broadwell(-X,-EP) when driven in similar clock ranges.

The Zen 2 to Zen 3 step is quite different: The core was made quite a bit wider, but the process node stood the same. The result was that light workloads get a pretty good uplift in Zen 3 vs. Zen 2, while power-hungry workloads, notably all-core workloads, see only mild improvements (from what I remember reading; I don't have Zen 3 CPUs myself). Special exceptions are those which profit from the different CCX organization, as mentioned in post #2.

I've got 1950x and 2700x CPU/mobo/RAM combo's not in use if you'd want any to try.