dick balaska <dic### [at] buckosoftcom> wrote:

> My oldest box is running ubuntu 14 / linux 3.13 (as opposed to linux
> 4.4) and does not report the bugs.

What's the most memory you have in an individual box?
What have you topped out at, rendering a frame?

Post a reply to this message

On 08/24/2018 05:34 AM, dick balaska wrote:
> On 08/23/2018 02:33 PM, Cousin Ricky wrote:
> 
> I don't know, man.  I like my i7, but this new i5 is *fast*.
> I reran a scene and this is the number of frames that each contributed:
> 
> name    frames  % of total
>          contrib
> james	1275	32.93     # i5-8400 @ 2.8GHz (6 cores 6 threads)
> cyd	789	20.38     # i7-4770 @ 3.4Ghz (4 cores, 8 threads)
> joe	701	18.10     # i5-7400 @ 3.0GHz (4 cores) (2017 box)
> ringo	618	15.96     # i5-4460 @ 3.2GHz (4 cores) (2015 box)
> liza	489	12.63     # i5-3330 @ 3.0GHz (4 cores) (2013 box)
> 
> That scene ran for 8 days and the new i5 was only up for 6 days. I'd bet
> the next scene he gets close to 40% of the total work.
> 

Interesting - especially it not slowing over time. Supposing:

grep MHz /proc/cpuinfo

shows all six cpu values?

Mike posted a wikipedia link for the i5 generations and something which 
caught my eye there is the turbo boost is being applied to your 
processor relatively evenly across all cores. 10/10/11/11/11/12 
multipliers of 100mhz where on older processors it was more unevenly 
boosted and by less. Expect this playing a part in your good performance 
too.

I wonder if having the hyperthreading circuitry off in each core - my 
understanding is it is there, but disabled - allows each core to run 
cooler? There would be less demand for on die core specific storage too 
I guess with just one thread.

Christoph, You wondered about articles. Found at least one of the 
articles I'd read about more cores instead of threads:

https://arstechnica.com/gadgets/2018/07/leaked-benchmarks-show-intel-is-dropping-hyperthreading-from-i7-chips/

My memory not quite right on things with respect to the current 
generation of i5s, but the more cores no threads direction at similar 
price is good news for POV-Ray.

Bill P.

Post a reply to this message

Am 24.08.2018 um 15:51 schrieb William F Pokorny:

> Interesting - especially it not slowing over time. Supposing:

Why should they?

Modern CPU cooling systems dissipate heat very quickly. If POV-Ray were
to max out the cooling system's capacity, it would show within a couple
of seconds.

> Mike posted a wikipedia link for the i5 generations and something which
> caught my eye there is the turbo boost is being applied to your
> processor relatively evenly across all cores. 10/10/11/11/11/12
> multipliers of 100mhz where on older processors it was more unevenly
> boosted and by less. Expect this playing a part in your good performance
> too.
> 
> I wonder if having the hyperthreading circuitry off in each core - my
> understanding is it is there, but disabled - allows each core to run
> cooler? There would be less demand for on die core specific storage too
> I guess with just one thread.

Probably. There's more chip area than if the HT circuitry were
completely absent, so there's more area to dissipate the heat. Also, the
components that would normally be shared between HT units won't be maxed
out and thus create less heat. So given the same cooling system, with HT
off the CPU will indeed inevitably run cooler. But the CPU may be
specified at a lower TDP, in which case a weaker cooling system may be
employed and thus resulting in the same CPU temperature.

I'm not sure whether Intel does indeed produce cores with HT circuitry
existing yet disabled, but producing non-HT CPUs in this manner does
make sense: You can just produce HT CPUs, and if during die testing you
find that a CPU has defective "semi-cores" you can just blow a fuse to
disable HT and sell it as a lower-tier CPU.

Intel has employed this strategy at least as early as the 80486 days,
when they sold dies with functional FPUs as 80486-DX, and those with
faulty FPUs got a fuse blown and sold as FPU-less 80486-SX. And the
external 80487 FPUs you could buy to upgrade such systems were actually
fully-fledged 80486-DX dies that would just take over the entire system,
disabling the 80486-SX installed in the mainboard's CPU socket.


> My memory not quite right on things with respect to the current
> generation of i5s, but the more cores no threads direction at similar
> price is good news for POV-Ray.

Please don't call it "no threads". Threads are a software thing, a means
of utilizing multiple cores by a single application without too much
overhead. The hardware thing is "hyperthreading", a means of multiple
execution units sharing certain functional blocks in the CPU to utilize
them more efficiently.

Post a reply to this message

On 8/24/2018 9:51 AM, William F Pokorny wrote:



Interesting. I would have expected the bottom CPU to perform better due 
to more threads and higher clock speed.


Mike

Post a reply to this message

Le 24/08/2018 à 20:51, Mike Horvath a écrit :
> On 8/24/2018 9:51 AM, William F Pokorny wrote:
>> james    1275    32.93     # i5-8400 @ 2.8GHz (6 cores 6 threads)
>> cyd    789    20.38     # i7-4770 @ 3.4Ghz (4 cores, 8 threads)
> 
> Interesting. I would have expected the bottom CPU to perform better due
> to more threads and higher clock speed.
> 
> 
> Mike

But nobody looks at the speed of the memory (from tomhardware website):
* i5-8400 uses ddr4 2666
* i7-4770 uses ddr3 1600

DDR4 2666 has a bandwidth of 21333 MB/s
DDR3 1600 has a bandwidth of 12.8 GB/s

12.8 vs 21.3 !

That's the advantage of the new generation.

Post a reply to this message

Am 24.08.2018 um 20:51 schrieb Mike Horvath:
> On 8/24/2018 9:51 AM, William F Pokorny wrote:


> 
> Interesting. I would have expected the bottom CPU to perform better due
> to more threads and higher clock speed.

i7-4770 has a max clock of 3.9 GHz,

Despite being rated at 2.8 GHz, the i5-8400 can actually go up to 4.0
GHz, given adequate cooling. The i7-4770 can only go as high as 3.9 GHz.

The i5-8400 also has more cores, which is more important than the number
of concurrent threads. Each thread beyond the number of cores only
translates to about 20% to 25% a core's worth of added computing power,
but only about 1/5 to 1/4 a core's worth, so the i7-4770 is effectively
worth about 5 threads.

In addition, the i5-8400 is four generations newer than the i7-4770,
which presumably translates to some additional performance improvements
here and there from miscellaneous optimizations.

Post a reply to this message

Am 24.08.2018 um 21:39 schrieb Le_Forgeron:
> Le 24/08/2018 à 20:51, Mike Horvath a écrit :
>> On 8/24/2018 9:51 AM, William F Pokorny wrote:
>>> james    1275    32.93     # i5-8400 @ 2.8GHz (6 cores 6 threads)
>>> cyd    789    20.38     # i7-4770 @ 3.4Ghz (4 cores, 8 threads)
>>
>> Interesting. I would have expected the bottom CPU to perform better due
>> to more threads and higher clock speed.
>>
>>
>> Mike
> 
> But nobody looks at the speed of the memory (from tomhardware website):
> * i5-8400 uses ddr4 2666
> * i7-4770 uses ddr3 1600
> 
> DDR4 2666 has a bandwidth of 21333 MB/s
> DDR3 1600 has a bandwidth of 12.8 GB/s
> 
> 12.8 vs 21.3 !
> 
> That's the advantage of the new generation.

I'm not sure that translates to much of a performance gain in POV-Ray.

The memory bandwidth limits the peak throughput under sequential read
access.

Most of POV-Ray's memory accesses are probably random access though, and
there memory latency is the limiting factor, not bandwidth.

In terms of latency, DRAM hasn't seen any noticeable improvement for ages.

Also, RAM access doesn't seem to be a bottleneck for POV-Ray anyway:
There are few applications (even computation-heavy ones) that drive CPU
temperatures quite as high as POV-Ray, indicating that the CPU spends a
lot of time doing actual work rather than waiting for data to be fetched
from memory. Presumably POV-Ray's most heavily accessed data typically
fits well within the L3 cache.

Post a reply to this message

dick balaska <dic### [at] buckosoftcom> wrote:
> Any overclocking I do is provided by official intel.  I.e. the new i5
> overclocks itself. "turbo mode"

By "turbo mode," do you mean turbo boost?  That is not the same as overclocking.

Post a reply to this message

On 08/24/2018 01:58 PM, clipka wrote:
> Am 24.08.2018 um 15:51 schrieb William F Pokorny:
> 
>> Interesting - especially it not slowing over time. Supposing:
> 
> Why should they?
> 
> Modern CPU cooling systems dissipate heat very quickly. If POV-Ray were
> to max out the cooling system's capacity, it would show within a couple
> of seconds.
> 

I've not seen that behavior for an extended render on a machine with 
turbo boost active is the short answer.

My old i7-920 - before I over-clocked it with the turbo mode turned off 
- didn't hold the turbo boost frequency for long. I was doing long 
single image isosurface stuff at the time. Dick appears to be getting a 
3.8GHz 'core' over time and under load.

A reason I suggested the grep command is I wonder if all the cores are 
maxed out with respect to the multipliers. Are some cores in fact being 
throttled? The 3.8ghz number is a core seeing a 10x 100Mhz but you can 
get up to 12x 100MHz on a core. I don't know if thermal throttling is by 
core or by die.

I further wonder how constant the reported 3.8GHz number really is. On 
my i3 the frequency values reported in /proc/cpuinfo look to be 
relatively instantaneous. I'm not under load and those MHz values change 
dramatically every time I look. I think we'd need to monitor the 
performance over time for a full image render to be sure the 3.8GHz held.

Dick is rendering frames. Maybe there is in fact some periodicity in the 
actual performance where a frame starts with turbo boost maxed out and 
then it tails off. Maybe thermal stuff isn't an issue because he is 
rendering frames and there is a cool off period between frames.

I worked for a long time on the hardware design side of CPUs and cache 
chips and power consumption/heat generation grows exponentially with 
frequency increases. Your right, if limits kick in, it happens fast. My 
thinking is 2.8GHz is the performance Intel can guarantee under all 
loads given the manufacturing process variation and shipped cooling at 
some expected "end of life." Otherwise, they'd be claiming a better base 
performance & charging for it, or, somehow handicapping the CPUs aimed 
at lower end boxes so as not to eat into their higher end - higher 
profit - business.

Though, AMD is again a serious player, maybe they are somewhat being 
forced to eat the high end business and what we see with Dick's new i5 
is a result.

All just me pondering. I don't know anything for sure - even that. :-)

> 
> Probably. There's more chip area than if the HT circuitry were
> completely absent, so there's more area to dissipate the heat. Also, the
> components that would normally be shared between HT units won't be maxed
> out and thus create less heat. So given the same cooling system, with HT
> off the CPU will indeed inevitably run cooler. But the CPU may be
> specified at a lower TDP, in which case a weaker cooling system may be
> employed and thus resulting in the same CPU temperature.
> 
> I'm not sure whether Intel does indeed produce cores with HT circuitry
> existing yet disabled, but producing non-HT CPUs in this manner does
> make sense: You can just produce HT CPUs, and if during die testing you
> find that a CPU has defective "semi-cores" you can just blow a fuse to
> disable HT and sell it as a lower-tier CPU.
> 
> Intel has employed this strategy at least as early as the 80486 days,
> when they sold dies with functional FPUs as 80486-DX, and those with
> faulty FPUs got a fuse blown and sold as FPU-less 80486-SX. And the
> external 80487 FPUs you could buy to upgrade such systems were actually
> fully-fledged 80486-DX dies that would just take over the entire system,
> disabling the 80486-SX installed in the mainboard's CPU socket.
> 
Agree & that last bit about the 80487's being in fact 80486-DX dies I 
did not know.

> 
>> My memory not quite right on things with respect to the current
>> generation of i5s, but the more cores no threads direction at similar
>> price is good news for POV-Ray.
> 
> Please don't call it "no threads". Threads are a software thing, a means
> of utilizing multiple cores by a single application without too much
> overhead. The hardware thing is "hyperthreading", a means of multiple
> execution units sharing certain functional blocks in the CPU to utilize
> them more efficiently.
> 

Good point. I got sloppy with the wording.

Bill P.

Post a reply to this message

Am 25.08.2018 um 18:10 schrieb William F Pokorny:

> A reason I suggested the grep command is I wonder if all the cores are
> maxed out with respect to the multipliers. Are some cores in fact being
> throttled? The 3.8ghz number is a core seeing a 10x 100Mhz but you can
> get up to 12x 100MHz on a core. I don't know if thermal throttling is by
> core or by die.

Where do you get those numbers from?

From all I know (and what a quick random peek at the internet
confirmed), CPU speed is per core, not per package.


> I further wonder how constant the reported 3.8GHz number really is. On
> my i3 the frequency values reported in /proc/cpuinfo look to be
> relatively instantaneous. I'm not under load and those MHz values change
> dramatically every time I look. I think we'd need to monitor the
> performance over time for a full image render to be sure the 3.8GHz held.

What you might be seeing is the CPU not throttling due to overheating,
but due to being bored.

Also, if this is a per-CPU speed, it may be averaging over all cores,
even cores shut down entirely due to being bored.


> My thinking is 2.8GHz is the performance Intel can guarantee under all
> loads given the manufacturing process variation and shipped cooling at
> some expected "end of life." Otherwise, they'd be claiming a better base
> performance & charging for it, or, somehow handicapping the CPUs aimed
> at lower end boxes so as not to eat into their higher end - higher
> profit - business.

Certainly. At the miniscule dimensions they're currently working at, I'd
be surprised if their tests could catch all the bad apples; so it's
reasonable to assume that they're working with safety margins, and
rating the CPUs for the worst case scenario.

Post a reply to this message