AMD 12C/24T Zen 2 / 3xxx UserBenchmark Leak

[Arkaign]

https://www.userbenchmark.com/UserRun/16910589

Notes :

1 - This is a VERY trusted source, unlike the silly Youtube man.
2 - This looks like 3.4Ghz base / 3.7Ghz Turbo. (model is SAMPLE 2D3212BGMCWH2_37/34_N)
3 - Tested on AMD Qugir MTS motherboard (no specific info)
4 - Goblins

Always remember, we don't know the full story on how this thing was sitting for testing. Interesting results though. Look at the memory type used (unavailable in the wild, meaning it is probably also in validation phase still), and benchmarks in that area.

 

[krumme]

https://www.userbenchmark.com/UserRun/16910589

Notes :

1 - This is a VERY trusted source, unlike the silly Youtube man.
2 - This looks like 3.4Ghz base / 3.7Ghz Turbo. (model is SAMPLE 2D3212BGMCWH2_37/34_N)
3 - Tested on AMD Qugir MTS motherboard (no specific info)
4 - Goblins

Always remember, we don't know the full story on how this thing was sitting for testing. Interesting results though. Look at the memory type used (unavailable in the wild, meaning it is probably also in validation phase still), and benchmarks in that area.

Ok and what do you derive from it?

 

[Arkaign]

Ok and what do you derive from it?

Hard to say in any definitive way, thought I'd run it past some of the folks here to get some feedback.

My 2 cents :

MT looks really solid for the clock speeds!
ST not ideal, but could be a setting, or something else. Seems too low to be optimized.

This is an Engy sample that has been benched twice before, but now this was tested just last night with :

New motherboard, and new memory, both not on the market. So, maybe out on loan to partner at RAM maker to run though some paces with memory? Very interesting to be sure.

 

[dacostafilipe]

From looncraz in redit:

Excellent!



This looks like a newer firmware, and also a different test bed.



The new result is a fixed 3.4GHz, which actually was read as 3.35GHz average. The other fixed frequency benchmark we had had worse results per clock, by a large margin, than the 3.6GHz sample which was posted earlier, this looks pretty similar.



Most importantly, however, is the HUGE improvement in memory latency. Just shy of 80ns and very consistent. Memory bandwidth has improved to 40.3GB/s, with 44.7GB/s read, which is closer to what you would expect from Ryzen running at DDR4-2933 to 3000.



Properly reported dual channel memory, reporting in at 1.6GHz - which is undoubtedly DDR4-3200 given the performance. I can't find, from a quick search, any DIMMs that use the reported memory, so I'm not sure what timings are in play or if the IMC is at 0.5X or anything of the sort. Given the relatively weak bandwidth numbers for the speed, I would not be surprised if this was with a 0.5X multiplier on the IMC or the changes made to the IMC are more focused on latency rather than bandwidth (which is a very smart trade).



I will update my Zen 2 Latency Calculator to reflect the assumptions as follows:

This test has DDR4-3200 CL16, CR2, 1:1 IMC, auto IF. The IMC has been tweaked to improve its latency and the IF latency cost is higher. This will take some time to do the math and update, naturally.



EDIT: Calculator updated. Increased IF latency, decreased (slightly) IMC latency.



EDIT 2:

SMT scaling has also improved by about 10%, so AMD has built upon their strengths.

 

[BigDaveX]

Keeping the same memory latency (let alone reducing it to Coffee Lake levels) while physically separating the CPU cores and IMC was always going to be a tall order, so if AMD just manage to keep it level with the original Zen in the final product, that'll still be a very respectable result. Especially considering that the last such comparable design - the first-gen Core i3 - really took a hit to memory bandwidth and latency compared to its higher-end peers.

 

[JoeRambo]

Memory looks to be running DDR4 @3200, but the latency is not good at all for 3200 speed, i guess that is why AMD is looking into much higher clocked DRAM, need 5000 speeds to feed cores with dual channel and to lower latencies.

 

[burninatortech4]

Does this new latency information shed any light on whether the I/O chiplet has any included cache or is it still too early to say?

Memory looks to be running DDR4 @3200, but the latency is not good at all for 3200 speed, i guess that is why AMD is looking into much higher clocked DRAM, need 5000 speeds to feed cores with dual channel and to lower latencies.

w/ Pinnacle Ridge - 3200mhz CL 14 should be around 68-70ns right? That's what my 2700x and B-Die Flare X 3200mhz CL14 kit get on the same test.

 

[DrMrLordX]

w/ Pinnacle Ridge - 3200mhz CL 14 should be around 68-70ns right? That's what my 2700x and B-Die Flare X 3200mhz CL14 kit get on the same test.

My Summit Ridge (1800x) gets 71 ns running DDR4-3200 14-14-14-28 CR1.

 

[thigobr]

From the latency curve it looks like it has 16MB of L3 cache (per CCX?)... 2x16MB from each chiplet?

 

[StinkyPinky]

Single core is low, but then again that's a boost to only 3.7. My 8700k is quite a bit higher so hopefully amd can get those clock speeds up.

 

[PotatoWithEarsOnSide]

It is 3200Mhz RAM at JEDEC standards, so CAS 22-24.
Implication is that 3200Mhz at CAS 16 would be on a par with Zen+ for latency.

 

[BigDaveX]

Single core is low, but then again that's a boost to only 3.7. My 8700k is quite a bit higher so hopefully amd can get those clock speeds up.

I don't think it's just an issue of clockspeed - checking around on the site, it looks like the single-thread score is in the ballpark of what you'd expect from a Core i5 2500K, so I'm guessing there's still a lot of optimisation at silicon and firmware level that still needs to be done before the final product comes out.

 

[JoeRambo]

w/ Pinnacle Ridge - 3200mhz CL 14 should be around 68-70ns right? That's what my 2700x and B-Die Flare X 3200mhz CL14 kit get on the same test.

~65-70 sound right for Zen1 @ 3200 with sensible latencies, to have almost 80ns @3200 it would require CL30 or so. Of course there is possibility of failure to sensibly configure DRAM at all, like having very relaxed secondaries and tertiaries and CL20+.

Hard to make conclusions from a single result, but bw numbers seem OK for 3200, latency not so.

 

[coercitiv]

~65-70 sound right for Zen1 @ 3200 with sensible latencies, to have almost 80ns @3200 it would require CL30 or so. Of course there is possibility of failure to sensibly configure DRAM at all, like having very relaxed secondaries and tertiaries and CL20+.

Here's some data from this particular benchmark on my own 1600X @ stock clocks, each setting was run 2 times:

2933 CL16 -> 78 and 80ns
3200 CL16 -> 74 and 78ns
3200 CL22 -> 81 and 83ns

Don't make me run CL30

 

[PotatoWithEarsOnSide]

Seems to be on a par with Zen+ for latency; they won't be validating anything with RAM operating outside of JEDEC specs, so it has to be CAS 22/24. Not sure why everyone assumes that they must be validating on non-JEDEC specs.
Anyway, matching Zen+ for latency is fine, but we also have to consider that the larger L3 will result in fewer trips to RAM in any event. Overall, that suggests improved performance.

 

[JoeRambo]

2933 CL16 -> 78 and 80ns
3200 CL16 -> 74 and 78ns
3200 CL22 -> 81 and 83ns

That is completely expected, 6CL is ~3.75ns worth of latency @3200, so your results pretty much indicate the same. 3200 CL30 would add 5ns more.

 

[Zucker2k]

For perspective: Mobile Skylake @ DDR4 2800 AUTO