Zen 2 details announced

[henryzz]

Originally Posted by mackerel

Yes, definitely. Dual channel 3000 ram isn't enough for even a fast quad core. It would take about double that to feed the 8086k to be close to unlimited.



I think I saw a commercial Intel tool that did that, but it is likely to be a case of if you have to ask, you can't afford it. Something for big businesses to pay big money for.


Edit: if anyone has specific test requests let me know and I will see what I can do. I'm hoping for an updated bios to allow the faster ram to be used and see how that helps (or not). Beyond that I need to move to more controlled tests.

There is a free version of Vtune available currently.

[Mysticial]

I'm actually curious, and this is more of a question for George.

What's the optimal FLOPS/byte-of-memory-access for prime95? For the stuff that I work on, it's about 10 FLOPs/byte with the current cache sizes.

Based on my estimates:

• Haswell-E is 10 FLOPs/byte
• Skylake X is ~20 FLOPs/byte depending on core count
• 16-core Zen 2 is 30 FLOPs/byte
• 64-core Zen 2 Epyc is 30 FLOPs/byte
• 64-core Zen 2 TR is 60 FLOPs/byte

Where does prime95 with FMA3/AVX2 sit? What about AVX512? ~3 FLOPs/byte?

[M344587487]

Originally Posted by mackerel

...
Edit: if anyone has specific test requests let me know and I will see what I can do. I'm hoping for an updated bios to allow the faster ram to be used and see how that helps (or not). Beyond that I need to move to more controlled tests.

Single channel 1 worker at 3000 DDR4 please. That'd be a rough indicator of what to expect from half of a 3900X and it'd be interesting to see how the read and/or write bandwidth comes into play.

[nomead]

Originally Posted by M344587487

Single channel 1 worker at 3000 DDR4 please. That'd be a rough indicator of what to expect from half of a 3900X and it'd be interesting to see how the read and/or write bandwidth comes into play.

That won't reflect the double amount of L3 cache, though. In this case, that may be really significant.

[M344587487]

Originally Posted by nomead

That won't reflect the double amount of L3 cache, though. In this case, that may be really significant.

Not for 1 worker over both chiplets on the 3900X, but it would simulate one worker per chiplet.

[maxzor]

Originally Posted by M344587487

Not for 1 worker over both chiplets on the 3900X, but it would simulate one worker per chiplet.

It would be cool to use one chiplet for GIMPS and the other for a daily system :) , been thinking about it for a few months

[mackerel]

Originally Posted by M344587487

Single channel 1 worker at 3000 DDR4 please. That'd be a rough indicator of what to expect from half of a 3900X and it'd be interesting to see how the read and/or write bandwidth comes into play.

Same setup, apart from pulling out one stick of ram.

4096k FFT: 197.3 (was 309.7 dual channel)
5120k FFT: 103.2 (was 193.9 dual channel)

I also ran y-cruncher 1B at 59.258s.

[M344587487]

Originally Posted by mackerel

Same setup, apart from pulling out one stick of ram.

4096k FFT: 197.3 (was 309.7 dual channel)
5120k FFT: 103.2 (was 193.9 dual channel)

I also ran y-cruncher 1B at 59.258s.

Nice, that's a good result. Assuming there are no major contention issues that sets the 3900X up to be ~394 it/s at DDR 3000 with two workers, 1 worker may be much better TBD.



The only other test I can think of is running dual channel at slower RAM speeds, say 2400 and 2666. It's not something you'd want to actually do but in the absence of being able to run at 3200 and higher it may be an indicator of how the 3600 scales with RAM. RAM scaling should be important as always but it's hard to predict whether the scaling is less or more impactful relative to non-zen2 chips. On one hand it should be less impactful due to the larger cache, on the other IF speed is tied to RAM speed. I predict scaling will be less impactful but still important.

[maxzor]

Maybe also try to overclock IF to 1900MHz and get a 1:1:1 run?

[mackerel]

To both the last responses, looking at how the new CPUs respond to ram speed is a part I want to look at. Unfortunately my current system does not seem to work beyond 3000 ram. I can go slower, but I'd want to do faster at the same time for a good overview. I'm half hoping a future bios update will improve ram, as my mobo AGESA combo version 1.0.0.1 is quite behind 1.0.0.3a if we're not on something even newer already. That was supposed to improve boot clocks in come cases.

There is also some chance my other mobo might work better with fast ram. The new CPUs weren't on the supported list at launch but was added recently. The latest bios doesn't specify what AGESA version it has, but it is still something I could try. This is a bigger job for the weekend.

My gut feeling is that faster ram would help at the largest sizes exceeding the cache. I wouldn't expect a significant difference below that.

[mackerel]

CPU availability is getting better. I just ordered a 3700X (8 core) for testing to see how those extra cores affect the balance. My gut feeling is the 6 core might remain the value sweet spot though.