Hi Ulli,
 I am actually running the I9-10850k. However, when running all bands, all modes including 4w on my 64 rx channels, i started having thermal issues seeing core temps hitting 100c.
I think his had a lot to do with the environment that server is sitting in which, is not climate controlled: it is in my basement which is now approaching ambient temp of 27c.

I made the decision to upgrade that I9 with a 12th Gen Intel I9-12900KS alder lake.
My Home theater pc is an aging I7, so I will Use the I9-10850K to replace that.

I completed the build yesterday and I had put the new 24 Core I9-1200KS online (this is a special edition KS release). It is indeed incredibly fast and handled all channels, all bands without issue.
I also used water cooling (360mm radiator) and 8 fans and a very large 4U, 64cm deep Rack mount enclosure for lots of airflow.
Performance comparison of old I9, and the new 12th gen, vs the other processors you had in your previous comparison.  The new 12th gen CPU mark is nearly double that of the old I9 and against any others on that list.
Intel Core i3-12100 vs Intel Core i5-12600K vs Intel Xeon E5-2670 v2 @ 2.50GHz vs Intel Core i9-10850K @ 3.60GHz vs Intel Core i9-12900KS [cpubenchmark.net] by PassMark Software

While it was online, the decodes completed in about 15-18 seconds, and Now temperatures barely went above 60c.
A few notable differences on this new build, which I *hope will be future proof for at least 5 years.

I choose the Z690 chipset with DDR5 support. Albeit the CL #'s are a bit higher than the DDR4 of the older I9, the DDR5 I purchased runs at 5600MHz, vs. the DDR4 which was running at 3600MHz.
I believe this negates any of the latency differences seen in the newer DDR5 at this point in time.

I had taken the new Server offline after a few hours last night because it appeared that WD was missing decode cycles (new server is running the latest WD 3.0.3.1). Not exactly sure what the problem is but I mentioned this to Rob along with other nuances such as the tail up monitoring no longer working.

The I9-10850k is running the older WD 3.0.3 which runs ok without all of the 4w modes, so I removed those. 
This new 12900ks was running the latest WD 3.0.3.1.  which i will possibly place back online today to monitor further

A few pictures on water colling and the new build. I need to complete cable management today, so pardon the mess.


r.

Thanks Rob,  the
for the interesting real live hardware observation.

Here is a comparison from the passmark cpu benchmark website - the mentioned i9 is what Tom is using - if I remember correctly he is not using the k-version.
Note however the very good single thread performance of the new Intel 12th gen cpu. Even the very low cost i3 performance is excellent.

Cpu comparison from passmark website

Ulli, ON5KQ

On Jun 26, 2022, at 3:55 PM, Phil Karn <karn@...> wrote:



I wonder how relevant these benchmarks are to WSPR decoding. You're probably spending nearly all of your CPU time in my Fano decoder, and one of the drawbacks of Fano decoding (or any sequential decoding algorithm) are lots of data-driven branches resulting in a poor branch prediction hit rate.

Modern CPUs are deeply pipelined for performance. Each CPU core appears to the programmer to execute an instruction stream in sequence, but it's actually executing multiple instructions in parallel, often out of sequence, each at a different stage of execution. Each instruction has to be fetched, decoded, held until any previous instructions on which it depends have been retired, its memory arguments (if any) fetched (with possible cache misses and loads), queued for whatever execution units are required, the results (if any) written to memory (cache), and then retired. All this parallelism works great until you hit a conditional branch instruction. The CPU can't know which way the branch will go before it is actually  executed, so it has to make an informed guess. If it guesses right, great. But if it guesses wrong, everything grinds to a screeching halt as the pipeline flushes and the CPU re-fetches the instruction stream at the correct address. You still get the right answers, just a lot more slowly.

This is where branch prediction comes in. The CPU keeps track of which branch instructions were (or were not) taken in the past and uses that history as hints for instruction fetching. This works fine for many cases, such as at the end of a loop with many iterations where the branch is always taken except for the last iteration. Accurate branch prediction has become so important to performance that it has become quite sophisticated, but it still performs poorly on instruction streams where the branches are (or aren't) taken essentially at random -- as is in sequential decoding on an unknown, noisy input symbol stream.

This means CPUs with good benchmark ratings aren't necessarily best at sequential decoding. If you can parallelize the load (as with lots of receiver channels) it may turn out that a lot of inexpensive CPUs may outperform a few expensive CPUs. The only way to tell is to benchmark on your actual application, not the artificial benchmarks you see in many review articles.

It looks like WSPR is going to be around a while, so it might be worthwhile for me to see if I can re-implement my Fano decoder (which I wrote in the mid 1990s!) to avoid as many data-dependent branches as possible. When pipelining was introduced to the Intel P6 processor in the mid 1990s, a bunch of conditional move instructions were introduced to avoid the need for branching on many common operations. E.g., the code

if(a > b)

    c = a;

else

    c = b;

would ordinarily involve a data-dependent branch that can be avoided with a conditional move. All modern compilers will detect common sequences like these and issue conditional moves instead. But a Fano decoder is probably too complex for the compiler to avoid a lot of data-dependent branching without help from the programmer. It's been a long time since I looked at this so I don't know how much of an improvement I can get. But I can try.

73, Phil

On 6/26/22 03:43, ON5KQ wrote:

Thanks Rob,  the
for the interesting real live hardware observation.

Here is a comparison from the passmark cpu benchmark website - the mentioned i9 is what Tom is using - if I remember correctly he is not using the k-version.
Note however the very good single thread performance of the new Intel 12th gen cpu. Even the very low cost i3 performance is excellent.

Cpu comparison from passmark website

Ulli, ON5KQ

I wonder how relevant these benchmarks are to WSPR decoding. You're probably spending nearly all of your CPU time in my Fano decoder, and one of the drawbacks of Fano decoding (or any sequential decoding algorithm) are lots of data-driven branches resulting in a poor branch prediction hit rate.

Modern CPUs are deeply pipelined for performance. Each CPU core appears to the programmer to execute an instruction stream in sequence, but it's actually executing multiple instructions in parallel, often out of sequence, each at a different stage of execution. Each instruction has to be fetched, decoded, held until any previous instructions on which it depends have been retired, its memory arguments (if any) fetched (with possible cache misses and loads), queued for whatever execution units are required, the results (if any) written to memory (cache), and then retired. All this parallelism works great until you hit a conditional branch instruction. The CPU can't know which way the branch will go before it is actually  executed, so it has to make an informed guess. If it guesses right, great. But if it guesses wrong, everything grinds to a screeching halt as the pipeline flushes and the CPU re-fetches the instruction stream at the correct address. You still get the right answers, just a lot more slowly.

This is where branch prediction comes in. The CPU keeps track of which branch instructions were (or were not) taken in the past and uses that history as hints for instruction fetching. This works fine for many cases, such as at the end of a loop with many iterations where the branch is always taken except for the last iteration. Accurate branch prediction has become so important to performance that it has become quite sophisticated, but it still performs poorly on instruction streams where the branches are (or aren't) taken essentially at random -- as is in sequential decoding on an unknown, noisy input symbol stream.

This means CPUs with good benchmark ratings aren't necessarily best at sequential decoding. If you can parallelize the load (as with lots of receiver channels) it may turn out that a lot of inexpensive CPUs may outperform a few expensive CPUs. The only way to tell is to benchmark on your actual application, not the artificial benchmarks you see in many review articles.

It looks like WSPR is going to be around a while, so it might be worthwhile for me to see if I can re-implement my Fano decoder (which I wrote in the mid 1990s!) to avoid as many data-dependent branches as possible. When pipelining was introduced to the Intel P6 processor in the mid 1990s, a bunch of conditional move instructions were introduced to avoid the need for branching on many common operations. E.g., the code

if(a > b)

    c = a;

else

    c = b;

would ordinarily involve a data-dependent branch that can be avoided with a conditional move. All modern compilers will detect common sequences like these and issue conditional moves instead. But a Fano decoder is probably too complex for the compiler to avoid a lot of data-dependent branching without help from the programmer. It's been a long time since I looked at this so I don't know how much of an improvement I can get. But I can try.

73, Phil

On Jul 3, 2022, at 16:15, ON5KQ <ON5KQ@...> wrote:

Tom,
your PC build is absolutely great !
I would  try to make it an application server, rather than using it as a workstation for just one specific application only at a time.
For example:
You can host a specific OS on the hardware, which allows to run many virtualized apps at the same time. 
For example a system as Unraid (www.unraid.net), which could serve as a mediaserver, but additionally you can run many instances of WD virtualized in parallel.
If I understand correctly from Phil comments above, it may be even more efficient in decoding, than just on single big decode job... could be interesting experiment at least. For example one kiwi per performance core per virtual machine per band and let it run all in parallel.... need to think about most efficient way to merge without dupes,,,

Homecinema streaming you could use PLEX for example...
Just idea.

regards,
Ulli, ON5KQ

On Jul 3, 2022, at 4:49 PM, KD2OM <steve@...> wrote:

 Ulli,

With these electricity problems, is Europe still pushing electric vehicles like here in the US?

73,

Steve KD2OM

.

 

On Jul 3, 2022, at 16:15, ON5KQ <ON5KQ@...> wrote:

Tom,
your PC build is absolutely great !
I would  try to make it an application server, rather than using it as a workstation for just one specific application only at a time.
For example:
You can host a specific OS on the hardware, which allows to run many virtualized apps at the same time. 
For example a system as Unraid (www.unraid.net), which could serve as a mediaserver, but additionally you can run many instances of WD virtualized in parallel.
If I understand correctly from Phil comments above, it may be even more efficient in decoding, than just on single big decode job... could be interesting experiment at least. For example one kiwi per performance core per virtual machine per band and let it run all in parallel.... need to think about most efficient way to merge without dupes,,,

Homecinema streaming you could use PLEX for example...
Just idea.

regards,
Ulli, ON5KQ

@ tom....correct re. bitcoin market melt down...gpu based servers flooding market cheap...hence my mention and possible useage

On Sun, Jul 3, 2022 at 2:32 PM WA2TP - Tom <myis300@...> wrote:

The problem with GPU is that it had become terribly expensive, and scarce due to bitcoin.

Now that bitcoin is on the downward trend, perhaps a better time to explore.

With this new 1200ks chipset and DDR5 prices dropping, and improvements in CL,

I think it will serve well for several years.

My biggest problem has been cooling.

I have yet to find a good stable distribution that will run a water coil per on Ubuntu.

The one I tried failed to start the pump  on reboot which caused thermal protection activation from the bios (thankfully).

On Jul 3, 2022, at 4:49 PM, KD2OM <steve@...> wrote:

 Ulli,

With these electricity problems, is Europe still pushing electric vehicles like here in the US?

73,

Steve KD2OM

.

 

On Jul 3, 2022, at 16:15, ON5KQ <ON5KQ@...> wrote:

Tom,
your PC build is absolutely great !
I would  try to make it an application server, rather than using it as a workstation for just one specific application only at a time.
For example:
You can host a specific OS on the hardware, which allows to run many virtualized apps at the same time. 
For example a system as Unraid (www.unraid.net), which could serve as a mediaserver, but additionally you can run many instances of WD virtualized in parallel.
If I understand correctly from Phil comments above, it may be even more efficient in decoding, than just on single big decode job... could be interesting experiment at least. For example one kiwi per performance core per virtual machine per band and let it run all in parallel.... need to think about most efficient way to merge without dupes,,,

Homecinema streaming you could use PLEX for example...
Just idea.

regards,
Ulli, ON5KQ

@ tom....correct re. bitcoin market melt down...gpu based servers flooding market cheap...hence my mention and possible useage

On Sun, Jul 3, 2022 at 2:32 PM WA2TP - Tom <myis300@...> wrote:

The problem with GPU is that it had become terribly expensive, and scarce due to bitcoin.

Now that bitcoin is on the downward trend, perhaps a better time to explore.

With this new 1200ks chipset and DDR5 prices dropping, and improvements in CL,

I think it will serve well for several years.

My biggest problem has been cooling.

I have yet to find a good stable distribution that will run a water coil per on Ubuntu.

The one I tried failed to start the pump  on reboot which caused thermal protection activation from the bios (thankfully).

On Jul 3, 2022, at 4:49 PM, KD2OM <steve@...> wrote:

 Ulli,

With these electricity problems, is Europe still pushing electric vehicles like here in the US?

73,

Steve KD2OM

.

 

On Jul 3, 2022, at 16:15, ON5KQ <ON5KQ@...> wrote:

Tom,
your PC build is absolutely great !
I would  try to make it an application server, rather than using it as a workstation for just one specific application only at a time.
For example:
You can host a specific OS on the hardware, which allows to run many virtualized apps at the same time. 
For example a system as Unraid (www.unraid.net), which could serve as a mediaserver, but additionally you can run many instances of WD virtualized in parallel.
If I understand correctly from Phil comments above, it may be even more efficient in decoding, than just on single big decode job... could be interesting experiment at least. For example one kiwi per performance core per virtual machine per band and let it run all in parallel.... need to think about most efficient way to merge without dupes,,,

Homecinema streaming you could use PLEX for example...
Just idea.

regards,
Ulli, ON5KQ

But I still maintain that an experimentable - for the average Joe, not the enthusiast - low power FST4-in-a-box transmitter would be boon to the mode.  Simple experimentation leads to more sophisticated approaches, etc, etc... it's about onboarding people.