32 core Threadripper about 100M NPS

[Dann Corbit]

See them here: http://www.ipmanchess.yolasite.com/amd- ... -bench.php

Regards,

Zenmastur

With the same core count (32) the new threadripper gen 3 is 30.8% faster than the old version (both popcount versions):
101442355/77576847=1.3076369938056389427634252781632

This is not accounted for by the IPC improvements, but also must include the saturation at high core levels where the older version dropped off in efficiency as the thread count rose.

It also appears that both threadripper systems have been overclocked (gen 2 and 3).

[Zenmastur]

See them here: http://www.ipmanchess.yolasite.com/amd- ... -bench.php

Regards,

Zenmastur

With the same core count (32) the new threadripper gen 3 is 30.8% faster than the old version (both popcount versions):
101442355/77576847=1.3076369938056389427634252781632

This is not accounted for by the IPC improvements, but also must include the saturation at high core levels where the older version dropped off in efficiency as the thread count rose.

It also appears that both threadripper systems have been overclocked (gen 2 and 3).

What makes you think they were overclocked?

Regards,

Zenmastur

[Dann Corbit]

Ipman chess recorded the frequency for the threadripper tests.
They are above stock frequency.

[Vinvin]

See them here: http://www.ipmanchess.yolasite.com/amd- ... -bench.php

Regards,

Zenmastur

One that could be added there :
2xEPYC 7742 : 237 Mn/s http://talkchess.com/forum3/viewtopic.php?f=2&t=71518

[Werewolf]

I find myself wondering how strong a 2019 IBM chess machine would be.

^ this

[Zenmastur]

Ipman chess recorded the frequency for the threadripper tests.
They are above stock frequency.

So what? Ryzen chips are not bound by their advertised base clock frequency. They are bound by max current draw(2 different ones), max core voltage, power targets, thermal constraints and lastly their max boost clocks. As long as none of these limits are violated they are operating within design specs and ARE NOT over clocked even if all cores are operation at max boost clock frequency. I.E. there only frequency limit is max boost clock. As long as all other parameters are with in specified limits it's not over clocked. Chess isn't particularly demanding on a CPU as far as power required. This means that less current and core voltage will be required therefore less power and less heat. So, while it could have been overclocked “IF” they forced the violation of one or more of the constrains listed above there is no way to know for sure. The “natural” behavior of the CPU is to continue boosting cores until it hits one of the above listed constraints. If no constraint is reached it will stop boosting when the max boost frequency is reached.

In almost all cases it will reach a thermal limit first and this is generally the controlling parameter on how high the boost clock will reach. If you can keep the CPU cool it will boost to higher frequencies until one of the other constraints are reached. Depending on how good the silicon is and the workload it's executing this could be one of the current limits, TDP, or VCORE voltage.

Without more information from the source there is no way to tell if it was overclocked. My guess is that it wasn't and that the reason it stopped boosting at 3.9 Ghz is that it reached a thermal limit. My guess is that if you put it on a chilled water loop it would boost to 4.2 Ghz while running SF or ASMfish ( or there about) while staying within all other specs.

Regards,

Zenmastur

[dragontamer5788]

Ipman chess recorded the frequency for the threadripper tests.
They are above stock frequency.

That's quite typical these days.

The implementations are different, but both Intel (Turbo Boost) and AMD (Turbo Core) automatically overclock your systems even at stock settings. All modern CPUs (everything after 2011 or so) has this effect.

Ultimately, if you run a modern CPU at 20C / 68F ambient, you'll get better performance than a CPU held at 30C / 86F ambient (AMD, Intel, or even ARM / Cellphones have this effect). The CPUs include temperature and power-sensors that automatically overclock themselves for the best performance given the current ambient conditions. As such, a "realistic" test should take an ~1-hour or so, which heats up a CPU to the point where you'll get more consistent data. Most tech-review sites only hold ~1 minute or 10-minute benchmarks however, so we're not really getting apples-to-apples comparisons anymore.

As annoying as it is for benchmark / reviewers to test for... this kind of behavior is ideal for the end-user. If ambient conditions are all well, CPUs should detect it and boost its own performance, and take advantage of the cold air or relatively cool heatsinks that exist in the room.

[MikeB]

https://www.phoronix.com/scan.php?page= ... 970x&num=9
Both for SF and AF.

Pretty accurate subject heading:
This is official stockfish compile:

Code: Select all

 stockfish bench 2048 64 26 >/dev/null

===========================
Total time (ms) : 109100
Nodes searched  : 10023226905
Nodes/second    : 91871923

and this is the functionally equivalent stockfish when you use the 'honey' source' and compile stockfish (code includes fast magics and huge pages for Linux and if set your Linux machine to use HUGE PAGES)

Code: Select all

Stockfish-010120 b 2048 64 26 >/dev/null 
===========================
Total time (ms) : 64542
Nodes searched  : 6994715982
Nodes/second    : 108374639

as soon as they release SF 11, executables will be made available

[corres]

...
Ultimately, if you run a modern CPU at 20C / 68F ambient, you'll get better performance than a CPU held at 30C / 86F ambient (AMD, Intel, or even ARM / Cellphones have this effect). The CPUs include temperature and power-sensors that automatically overclock themselves for the best performance given the current ambient conditions. As such, a "realistic" test should take an ~1-hour or so, which heats up a CPU to the point where you'll get more consistent data. Most tech-review sites only hold ~1 minute or 10-minute benchmarks however, so we're not really getting apples-to-apples comparisons anymore.
...

This is the cause why we need manually overclocked CPU without turbo boost and HT/SMT for testing.
For manual OC especially important the effective cooling system: good cooler + well ventilated case.
Water coolers are very effective but there is a disadvantage too: They cool only the CPU meanwhile the RAMs and
the VRC blocks also need cooling. For this complete tasks a good air cooler is more appropriate.

[corres]

...
Ultimately, if you run a modern CPU at 20C / 68F ambient, you'll get better performance than a CPU held at 30C / 86F ambient (AMD, Intel, or even ARM / Cellphones have this effect). The CPUs include temperature and power-sensors that automatically overclock themselves for the best performance given the current ambient conditions. As such, a "realistic" test should take an ~1-hour or so, which heats up a CPU to the point where you'll get more consistent data. Most tech-review sites only hold ~1 minute or 10-minute benchmarks however, so we're not really getting apples-to-apples comparisons anymore.
...

This is the cause why we need manually overclocked CPU without turbo boost and HT/SMT for testing.
For manual OC especially important the effective cooling system: good cooler + well ventilated case.
Water coolers are very effective but there is a disadvantage too: They cool only the CPU meanwhile the
RAMs and the VRC blocks also need cooling. For this complete tasks a good air cooler is more appropriate.