Zen 3 speculation - Page 6

VBCurtis · 2020-11-17, 00:09

Quote:

Originally Posted by M344587487

Ah this is the good stuff. If only there were a way to pipe benchmarks directly into a vein.

I love this forum. That is all.

Xyzzy · 2020-11-17, 15:02

Ouch!

CPU = i5-10600K
RAM = 2×8GB DDR4-3200
CMD = ./msieve -v -nc -t 6
LA = 6h 06m 28s

CPU = 5600X
RAM = 2×16GB DDR4-3200
CMD = ./msieve -v -nc -t 12
LA = 4h 6m 45s

Note: The 10600K didn't scale well with HT so we ran it with six threads instead of twelve. The 10600K result was also with single rank memory. Our Z490 mobo is at Asus for an RMA right now so we can't retest yet. Also, the 10600K test was with all power and current limits removed so it used a lot more (maybe 2-3× more) electricity than the 5600X which we had set as "STK" AKA "no PBO and no ECO".

Xyzzy · 2020-11-17, 15:06

PS - Later today we will run the 5600X test again with the same single rank memory we used for the 10600K. We will also use six cores to make the comparison as fair as possible. Maybe we will even turn on "PBO" as well.

Any predictions for the time?

M344587487 · 2020-11-17, 17:45

No idea how msieve scales, so 4h30 at 12 threads SR, with 6 threads SR maybe 5h?

VBCurtis · 2020-11-17, 19:32

I don't think msieve is memory-limited with just 6 cores, so I doubt the single-rank memory will be much slower than your dual-rank benchmark. Perhaps 3-5% = 10-15 min slower?

I'm prepared to be mistaken, though.

That 5600 sure is fast for matrix-solving! Glad I didn't order one of those used Phi systems, perhaps my silicon habit shall be fed this winter by Zen 3.

PhilF · 2020-11-17, 19:38

Quote:

Originally Posted by VBCurtis

perhaps my silicon habit shall be fed this winter by Zen 3.

Dang there's a lot of addicts around here. I should know. Takes one to know one.

Xyzzy · 2020-11-17, 20:44

MAX POWER - 6 THREADS - 2×8GB DDR4-3200C14 (SR)

6h 06m 28s i5-10600K
4h 34m 46s 5600X

Xyzzy · 2020-11-17, 20:51

The six core run was with "PBO" turned on and single rank memory but the twelve core run was with "STK" power and dual rank memory.

CPU = 5600X
RAM = 2×16GB DDR4-3200
CMD = ./msieve -v -nc -t 12
LA = 4h 6m 45s

CPU = 5600X
RAM = 2×8GB DDR4-3200
CMD = ./msieve -v -nc -t 6
LA = 4h 34m 46s

So tonight we will run a single rank "STK" test to compare only the SR:DR change. Or, we could run a dual rank "PBO" test. Which do you all prefer?

VBCurtis · 2020-11-17, 22:15

Single-Rank stock, please.

Xyzzy · 2020-11-18, 00:17

Since we are using the "STK" power level, we decided to try a smaller HSF combo.

https://noctua.at/en/nh-l9x65

After letting the temp stabilize (running a small FFT torture test) we ended up with 71°C for the CPU. The "CPU Package" was 82°C - whatever that is. The motherboard's rear I/O plate has a debug LED display the shows the CPU temp after the boot sequence is done. In our case that displayed temp is the same as the "CPU" reading in the attachment below. Even if 82°C is the real temp that is easily within specs, and we rarely run the CPU with twelve threads of highly optimized AVX work.

We will need to get a 25mm thick black fan for it but otherwise it is perfect!

M344587487 · 2020-11-18, 09:53

The CPU package temp might be the junction temperature, the highest value of all the temp sensors on the SoC. Probably the temp of the cache as you're hammering it.

2020-11-18, 00:17	#65
Xyzzy "Mike" Aug 2002 2³×3×331 Posts	Since we are using the "STK" power level, we decided to try a smaller HSF combo. https://noctua.at/en/nh-l9x65 After letting the temp stabilize (running a small FFT torture test) we ended up with 71°C for the CPU. The "CPU Package" was 82°C - whatever that is. The motherboard's rear I/O plate has a debug LED display the shows the CPU temp after the boot sequence is done. In our case that displayed temp is the same as the "CPU" reading in the attachment below. Even if 82°C is the real temp that is easily within specs, and we rarely run the CPU with twelve threads of highly optimized AVX work. We will need to get a 25mm thick black fan for it but otherwise it is perfect! Attached Thumbnails

2020-11-17, 15:02	#57
Xyzzy "Mike" Aug 2002 2³×3×331 Posts	Ouch! CPU = i5-10600K RAM = 2×8GB DDR4-3200 CMD = ./msieve -v -nc -t 6 LA = 6h 06m 28s CPU = 5600X RAM = 2×16GB DDR4-3200 CMD = ./msieve -v -nc -t 12 LA = 4h 6m 45s Note: The 10600K didn't scale well with HT so we ran it with six threads instead of twelve. The 10600K result was also with single rank memory. Our Z490 mobo is at Asus for an RMA right now so we can't retest yet. Also, the 10600K test was with all power and current limits removed so it used a lot more (maybe 2-3× more) electricity than the 5600X which we had set as "STK" AKA "no PBO and no ECO".

2020-11-17, 15:06	#58
Xyzzy "Mike" Aug 2002 7944₁₀ Posts	PS - Later today we will run the 5600X test again with the same single rank memory we used for the 10600K. We will also use six cores to make the comparison as fair as possible. Maybe we will even turn on "PBO" as well. Any predictions for the time?

2020-11-17, 17:45	#59
M344587487 "Composite as Heck" Oct 2017 1011011111₂ Posts	No idea how msieve scales, so 4h30 at 12 threads SR, with 6 threads SR maybe 5h?

2020-11-17, 19:32	#60
VBCurtis "Curtis" Feb 2005 Riverside, CA 3⁵·19 Posts	I don't think msieve is memory-limited with just 6 cores, so I doubt the single-rank memory will be much slower than your dual-rank benchmark. Perhaps 3-5% = 10-15 min slower? I'm prepared to be mistaken, though. That 5600 sure is fast for matrix-solving! Glad I didn't order one of those used Phi systems, perhaps my silicon habit shall be fed this winter by Zen 3.

2020-11-17, 20:44	#62
Xyzzy "Mike" Aug 2002 2³·3·331 Posts	MAX POWER - 6 THREADS - 2×8GB DDR4-3200C14 (SR) 6h 06m 28s i5-10600K 4h 34m 46s 5600X

2020-11-17, 20:51	#63
Xyzzy "Mike" Aug 2002 2³·3·331 Posts	The six core run was with "PBO" turned on and single rank memory but the twelve core run was with "STK" power and dual rank memory. CPU = 5600X RAM = 2×16GB DDR4-3200 CMD = ./msieve -v -nc -t 12 LA = 4h 6m 45s CPU = 5600X RAM = 2×8GB DDR4-3200 CMD = ./msieve -v -nc -t 6 LA = 4h 34m 46s So tonight we will run a single rank "STK" test to compare only the SR:DR change. Or, we could run a dual rank "PBO" test. Which do you all prefer?

2020-11-17, 22:15	#64
VBCurtis "Curtis" Feb 2005 Riverside, CA 3⁵×19 Posts	Single-Rank stock, please.

2020-11-18, 09:53	#66
M344587487 "Composite as Heck" Oct 2017 3·5·7² Posts	The CPU package temp might be the junction temperature, the highest value of all the temp sensors on the SoC. Probably the temp of the cache as you're hammering it.