Topic: AMD Ryzen hashrate? - page 21. (Read 85898 times)

I have a couple of suggestions for testing. Cryptonight is a good test for cache performance but
doesn't stress the CPU in any way. Ryzen (16 MB cache) is equivalent to 2x i7-6700K (8 MB cache each).
Very competitive in cache performance.

Testing raw compute performance is better done with the deep algo. It makes generous use of AES,
AVX and AVX2 while not being bogged down by memory accesses. It's a good test of Ryzen's AVX/2
and hyperthreading performance.

Lyra2z330 (zoin) is a good test for a memory hard algo. It's extreme memory requirements blow away
the cache and make it irrelevent. This tests the memory interface but is also dependetn on the memory itself.
Ryzen went with a 2 channel controller so will be at a disadvantage over Intel E series 4 channel controller.

Lyra2rev2 has a good mix of compute intensive and memory intensive functions.

It's good to monitor CPU Usage while testing. As seen with cryptonight a memory or cache bottleneck means
the CPU spends a lot of time idle. The CPU needs to be able to hit 100% to show its full compute potential.

Anyone try cpuminer-opt-3.5.3-windows with zcoin on ryzen ? This is :

i7 5820k @ 4.3 ghz :



i7 7700k @ 5ghz :



And i5 6600 @ default :



.

At 500h/s that is good enough to go with a R7 1700 and a sub $120 mobo that supports 6 gpus imo. Better hashing density, slow depreciation and easy to resell mobo, socket valid for 4 years, able to CPU mine new coins like zcash when it launched. I would underclock and undervolt for better hash/watt.

Ryzen on cynebench :



.

Did you read the link I gave?

At first he was getting a "modest" improvement ( the first result is no longer in the list since he updated it with the new info but if I remember correctly it was ~305 H/s) over the 1700's but after he played around with the thread count in Wolfs miner he saw a large improvement so he updated the list for the 1800x, he did not update the 1700's yet.

Output of lscpu on Ryzen:
"Architecture:          x86_64
CPU op-mode(s):        32-bit, 64-bit
Byte Order:            Little Endian
CPU(s):                16
On-line CPU(s) list:   0-15
Thread(s) per core:    2
Core(s) per socket:    8
Socket(s):             1
NUMA node(s):          1
Vendor ID:             AuthenticAMD
CPU family:            23
Model:                 1
Stepping:              1
CPU MHz:               3600.000
BogoMIPS:              7180.28
Virtualization:        AMD-V
L1d cache:             32K
L1i cache:             64K
L2 cache:              512K
L3 cache:              8192K
NUMA node0 CPU(s):     0-15"

He was using the output of lscpu and it was not listing L3 cache correctly, Ryzen has TWO 8MB L3 Caches, for a total of 16 mb of L3 and he was using only 4 threads instead of 8 so the performance was crap.

Cryptonite is very sensitive with the number of threads you let it use, not enough or too many the performance is worse. The algorithm uses two mb per thread so if you give it too many cores where it's swapped to ram it slows down or if you give it less than the optimum amount it is not using all of the resources in the cache optimally so it is not running as many threads it can for best performance.

From the link:

"Added AMD Ryzen 1800X Monero mining benchmarks to the first post. Modest improvement over the 1700X. "

And:

"Uh, if you were only using 4 threads that's your problem. and why it performs at half what i expected...

Ryzen has 4 MB L2 and 16MB l3. it is 8MB L3 per cluster complex and there are 2 clusters. Not 8MB total.

You should definitely retest at 8 and 10 threads. and what the hell test some others too because you never know... "

And:

"I did not get as good of results with 10 and 16 was extremely poor."






Ryzen uses two 8 mb L3 cache's not a single 16mb, so if data is transfered from 8mb L3 to 8mb L3 it hurts performance vs a single 16 mb L3 cache.

So it looks like testing and running the optimum number of threads is critical for performance for CPU's especially Ryzen, also until things are optimized for Ryzen it could be hit or miss with the performance of mining programs.

I do know there will be improvements to the microcode and motherboard bios's also software will need some optimizations, but all that means that the performance will only get better and that is usually what happens when a new cpu is released.
 
 

In review i read , ryzen 1700 clock at 4ghz is the same identical result with 1700x 4ghz and 1800x 4ghz , why in the table:

1x AMD Ryzen 7 1800x = 537H / s (nota avere su immagine Monero Ryzen)

1x AMD Ryzen 7 1700X = 273H / s
1x AMD Ryzen 7 1700 = 250h / s (209 utilizzando nproc)

Its not possible .

there are issues with (multiple) high clocked ram sticks installed, a friend of mine will try to get 3200mhz working, might have a great impact on cryptonight performance, sadly still no mobo arrived

LOL, well neither AMD or Intel have a good track record of sticking to the TDP figures.  For example look at the AMD TDP figures for the RX480 and RX470 GPU's.  Those figures are wishful thinking with the stock ROM's.  So I would venture to say the Ryzen CPU tdp figures are largely marketing at this point.  So far all the reviews I have read have shown much higher consumption rates than the published TDP.

You are right about the mobo's they are pretty iffy right now.  I heard this morning that the latest microcode was only finalized a few weeks ago so this is going to be a "work in progress" for a while.

133 watt with nominal tdp 95 watt , i read ryzen have a bug in all firmware bios motherboard . Anyone know this issue ? . Ryzen is a GREAT cpu , but all motherboard product is not good at moment .

Now those numbers are starting to look more realistic.  Thanks for the update.  I am sure there will be quite a few lessons learned and updates since this CPU is new to the market.  I now hear rumblings coming from Intel about more cores... Gosh, so all it takes is a little competition to get some progress and better prices??? Who knew right? 

I think I will give a Ryzen CPU a shot.  It has been a long time since I owned an AMD.  I won't be buying it for mining specifically, but I will test it and see if I can multitask it when the machine isn't busy doing other things.     It will be nice to have an AMD in the house again.  Always rooting for the underdog!

Big update on Ryzen:

"Which was keeping the thread count too low. Updated the Ryzen 7 1800X numbers and now getting over 500H/s. "

 https://forums.servethehome.com/index.php?threads/monero-mining-performance.12116/

Read page 16 for info.

He has not retested 1700 or 1700x yet.

"CPU Results (wolf's) Using (MB L3 cache/ 2) for threads
4x Intel Xeon E7-8870 V3 = 2600H/s (NB drops to 2000 over time)
2x Intel Xeon E5-2699 V4 = 1723H/s
1x Intel Xeon Phi 7210 = 602H/s (case to use nproc-1)
2x Intel Xeon E5-2698 V4 = 1572H/s
2x Intel Xeon E5-2698 V3 = 1463H/s
2x Intel Xeon E5-2690 V3 = 1100H/s
2x Intel Xeon E5-2667 V3 = 1090H/s
2x Intel Xeon E5-2658 V3 = 1050H/s
2x Intel Xeon E5-2650 V4 = 1047H/s
2x Intel Xeon E5-2650L V4 = 1031H/s
2x Intel Xeon E5-2683 V3 = 1014H/s
2x Intel Xeon E5-2680 V3 = 969H/s
2x Intel Xeon E5-2670 V3 = 969H/s
2x Intel Xeon E5-2628L V4 = 897H/s
2x Intel Xeon E5-2650L V3 = 874H/s
2x Intel Xeon E5-2630 V4 = 866H/s
2x Intel Xeon E5-2670 V1 = 829H/s
2x Intel Xeon E5-2650 V1 = 750H/s
1x AMD Ryzen 7 1800X = 537H/s (note have custom Monero Ryzen image)
2x Intel Xeon E5-2620 V1 = 480H/s
1x Intel Xeon D-1587 = 388H/s
1x Intel Xeon E5-2630L V3 = 334H/s
1x AMD Ryzen 7 1700X = 273H/s
1x AMD Ryzen 7 1700 = 250H/s (209 using nproc)
1x Intel Xeon D-1541 = 242H/s
1x Intel Xeon E3-1515M V5 = 240H/s
1x Intel Xeon D-1540 = 218H/s
1x Intel Xeon E3-1220 V3 = 186H/s (case to use either nproc/2 or nproc-1)
1x Intel Xeon E3-1245 V5 = 161H/s
1x Special CPU = 153H/s (case to use nproc)
1x Intel Atom C3338 = 77H/s (case to use nproc)
1x Intel Pentium D1508 = 47H/s (case to use nproc-1)
1x Intel Atom C2358 = 18H/s (use nproc)"

Nice increase with testing to find the sweet spot with the number of threads.

EDIT: Also to add the cpu was using 133 watts, not bad at all.  

But still behind Xeon ES's as for price/performance.

I can't wait until Naples is released.

I would also expect that the numbers will get a lot better if there is optimization for this new CPU.  Just like when new coin or new GPU is launched the first miners generally run pretty poorly.

Only ? i aspected more from ryzen , i think is not optimized...

No it won't. The algo is cache bound, it has nothing to do with compute power. Once the cache overflows performance goes
down. Optimum thread count is cache size MB / 2 MB. No CPU can mine cryptonight efficiently with max threads, hyper or not.

 Better example - NVidia 750 Ti - 250 H/s or a bit more if pushed for around 50 watts actual usage - which is the same ballpark as most CPUs are in any more, is easily a match for those Ryzen results at a bit LOWER power despite being generation-older tech and quite a bit less power usage than the high-end stuff usually used for CPU mining, as well as being quite a bit less $$$ than any of the CPUs listed in the post I didn't quote that can compete with it on hashrate.

I thought about trying that with my older dual 2670 build, but that parts are back in their boxes because I wanted to put them up on ebay. 

That explains why my own CPU does not get more hashing past a certain point. I suppose the Ryzen CPU's are fairly efficient power wise for CPU mining but it would have been nice if they could simply produce double the hash rates they got right now.

I wonder now if disabling hyperthreading can be done to actually cram all its power into 8 threads. I could probably try that with my old i7

Quoted from the Serve The Home forum:

"It's definitely not a perfect formula, but the cryptonight algorith requires 2MB of memory per thread to calculate hashes. Keeping it all in L3 cache or below definitely sped it up on the platforms I've tested.

For example:

Intel Xeon E3-1240 v1 went from 129 H/s to 231 H/s (going from 7 to 4 threads, respectively). L3 cache size is 8MB."

This is the most important thing with cpu mining cryptonight, with both my E5's combined I have 70 mb of L3 so 70/2 = 35 so I can run all 32 threads and get the best performance, with my older pair of E5 2670's with a total of 40mb of L3 for both it seemed to run best with ~18 threads and any more the hash rate would drop.

Also you compare my ES 2667 V3's to the retail ones in the above list my has rate is a about 50-60 H/s faster even tho mine only boost up to 3.0 Ghz and the retail version boost up to I think 3.6 Ghz. The thing to think about is that the retail 2667's have 20mb of L3 vs 35mb of L3 with my ES cpu's.
 

Unfortunately the reviews are stating the TDP figures are pure fiction for Ryzen.  So I wouldn't use those TDP numbers to figure efficiency. 

From the awesome Servethehome forums
I'm guessing these numbers will improve as they get optimized miners from the devs.

CPU Results (wolf's) Using (MB L3 cache/ 2) for threads
4x Intel Xeon E7-8870 V3 = 2600H/s (NB drops to 2000 over time)
2x Intel Xeon E5-2699 V4 = 1723H/s
1x Intel Xeon Phi 7210 = 602H/s (case to use nproc-1)
2x Intel Xeon E5-2698 V4 = 1572H/s
2x Intel Xeon E5-2698 V3 = 1463H/s
2x Intel Xeon E5-2690 V3 = 1100H/s
2x Intel Xeon E5-2667 V3 = 1090H/s
2x Intel Xeon E5-2658 V3 = 1050H/s
2x Intel Xeon E5-2650 V4 = 1047H/s
2x Intel Xeon E5-2650L V4 = 1031H/s
2x Intel Xeon E5-2683 V3 = 1014H/s
2x Intel Xeon E5-2680 V3 = 969H/s
2x Intel Xeon E5-2670 V3 = 969H/s
2x Intel Xeon E5-2628L V4 = 897H/s
2x Intel Xeon E5-2650L V3 = 874H/s
2x Intel Xeon E5-2630 V4 = 866H/s
2x Intel Xeon E5-2670 V1 = 829H/s
2x Intel Xeon E5-2650 V1 = 750H/s
2x Intel Xeon E5-2620 V1 = 480H/s
1x Intel Xeon D-1587 = 388H/s
1x Intel Xeon E5-2630L V3 = 334H/s
1x AMD Ryzen 7 1700X = 273H/s
1x AMD Ryzen 7 1700 = 250H/s (209 using nproc)
1x Intel Xeon D-1541 = 242H/s
1x Intel Xeon E3-1515M V5 = 240H/s
1x Intel Xeon D-1540 = 218H/s
1x Intel Xeon E3-1220 V3 = 186H/s (case to use either nproc/2 or nproc-1)
1x Intel Xeon E3-1245 V5 = 161H/s
1x Special CPU = 153H/s (case to use nproc)
1x Intel Atom C3338 = 77H/s (case to use nproc)
1x Intel Pentium D1508 = 47H/s (case to use nproc-1)
1x Intel Atom C2358 = 18H/s (use nproc)