I have a 5950X workstation, but would a 14900K one be an upgrade or
downgrade for computer programming/science code building/compiling/making
speed? Won't use newer AMD until there's a full or larger ATX system-/
logic-/mother-/main-board with plain PCI (currently only on micro-ATX).

On 1/5/2024 2:30 AM, David Chmelik wrote:
> I have a 5950X workstation, but would a 14900K one be an upgrade or
> downgrade for computer programming/science code building/compiling/making
> speed? Won't use newer AMD until there's a full or larger ATX system-/
> logic-/mother-/main-board with plain PCI (currently only on micro-ATX).
>

Requires a miracle to get a PCI slot now.
Even though it is a one-chip solution to have a PCI slot (PCIe to PCI bridge chip).

*******

First we check Ark on ark.intel.com .

https://ark.intel.com/content/www/us/en/ark/products/236773/intel-core-i9-processor-14900k-36m-cache-up-to-6-00-ghz.html

P&E cores, Turbo, 253 watts # Hefty load

Many programs do not benefit from multiple cores. The
basis of snappy desktop performance, is the behavior of
a single thread task.

https://www.cpubenchmark.net/singleThread.html

14900K 4,798
5950X 3,466

Multithreaded.

https://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i9-14900K&id=5717

14900K 61379

https://www.cpubenchmark.net/cpu_lookup.php?cpu=AMD+Ryzen+9+5950X&id=3862

5950X 45780

The Intel is asymmetric because of P&E. 16+16+36MB=68MB cache
L3 one blob, multiple ports ?

Cache L1: 80 KB (per core)
Cache L2: 2 MB (per core) P-Cores: 8 (x2 core count for Hyperthreaded)
Cache L3: 36 MB (shared)

E-Core L1: 96 KB (per core) E-Cores: 16 (four cores per module)
E-Core L2: 4 MB (per module)

https://www.techpowerup.com/review/intel-core-i9-14900k/2.html

Die shot

https://www.techpowerup.com/review/intel-core-i9-14900k/images/arch1.jpg

Die shot labels

https://www.techpowerup.com/review/intel-core-i9-14900k/images/arch2.jpg

"Ring bus" is not the best. On my 4930K, you get five cores performance
from a six core processor, because of "Ring bus". Unfortunately, I have
been unable to find a scaling test that increases core count one at a time.

Cinebench result here??? Can't find a methodical scaling test web page.

# Ring bus apparently has options to downclock. Tiny effect on program launch ?

https://forums.guru3d.com/threads/intel-ring-bus-oc-improvement-to-pc-responsiveness.447118/

Intel had done "mesh" as a technical improvement over it's previous
series-connected ring buses. No mention of mesh bus on this one.
The ring structure is unstated (some large 28 core CPUs used to have three rings
joined at edges).

Now the L3 is labeled as "shared", but is it "segmented" ?

On the 5950X, there are two silicon dies, and each one has L3, and
to maintain their state, in cross-town traffic, would require snoop
traffic. This means on the 5950X, it may have more L3, but the L3 is
"segmented".

On the 5950X, you can take it out of dual channel mode, and
run the two channels separately. That's shown in your
"AMD Ryzen Master" application output. I've not investigated that,
like whether I can even find the setting in the BIOS screen.

[Picture]

https://i.postimg.cc/prKSTdqr/AMD-Ryzen-Master.gif

DDR5 memory, a DIMM has two channels on it. A dual channel Intel board
with DDR5 then, counts as four channels. But whether it can be
uncoupled, I don't know if that is possible. Historically, the
industry has been "a bit bus-dishonest", in that a quad channel
board in the past, really only fetched on two channels at
a time, and behaved like a dual channel board. Although there
have been diagrams showing mesh bus plus two reservation stations
for a quad channel board, it's not clear that the design chooses to
use two reservation stations (dual channel per each) at exactly the
same time.

Here, someone has cranked the living shit out of the memory. Notice
how the L3 is only a little faster than main memory.

https://rog-forum.asus.com/t5/image/serverpage/image-id/85105iFE78CC481DEF6913/image-size/large/is-moderation-mode/true?v=v2&px=999

5764 MHz 14900K
L1 80K 645.5 GB/sec
L2 2M 137.7 <=== 2M per E-core
L3 36M 55.1
DDR5-8200 34.6

A 7950X, same thing. L3 not a big win necessarily. But good speed, whatever the hell they've done to it.

https://www.techpowerup.com/forums/attachments/img_2615-jpg.288099/

4500 MHz
L1 64K 302.1 GB/sec
L2 1M 127.4
L3 64M 72.5
Not sure 55.4

My lowly 5950x. Stock clock. Memtest does not seem to turbo?

[Picture]

https://i.postimg.cc/MZ7TgJSK/memtest-5950x.gif

3400 MHz
L1 32K 292 GB/sec
L2 512K 139
L3 64M 48.3
DDR4-3200 15.8

Since memtest runs on multiple cores now, there are occasional
memory bandwidth reporting errors (factor of 2). The AMD 5600G
can report "40GB/sec" for memory, if you run it XMP profile 2 (=CR1)
and the memory test process uses multiple cores. Which suggests
the default 5600G configuration might be uncoupled or something.

Conclusion: Yes, 14800K is faster. But are you going to spend a lot of
money for that much improvement ? You have to be careful
about getting "trapped in incrementalism".

A lot of lazy reviewers does not make this easy to evaluate.

I have a scaling test for the 5950x, but the OS sucks at it.
The study mostly shows strange scheduling choices. And Jim Keller
already complained that another benchmark did not show off the
best performance of his arch.

That leaves us with Passmark 61379 versus 45780.
A single point comparison, for your purchasing choice.

Paul

On Fri, 5 Jan 2024 10:54:29 -0500, Paul wrote:
> On 1/5/2024 2:30 AM, David Chmelik wrote:
>> I have a 5950X workstation, but would a 14900K one be an upgrade or
>> downgrade for computer programming/science code building/compiling/
>> making speed? Won't use newer AMD until there's a full or larger ATX
>> system-/ logic-/mother-/main-board with plain PCI (currently only on
>> micro-ATX).
>>
> Requires a miracle to get a PCI slot now.
> Even though it is a one-chip solution to have a PCI slot (PCIe to PCI
> bridge chip). [...]
>
> https://www.cpubenchmark.net/singleThread.html
>
> 14900K 4,798
> 5950X 3,466
>
> Multithreaded.
>
> https://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i9-14900K&id=5717
>
> 14900K 61379
>
> https://www.cpubenchmark.net/cpu_lookup.php?
cpu=AMD+Ryzen+9+5950X&id=3862
>
> 5950X 45780
>
> [...]
>
> Conclusion: Yes, 14800K is faster. But are you going to spend a lot of
> money for that much improvement ? You have to be careful
> about getting "trapped in incrementalism".

I was unaware of so many tests; thanks for details! Guess you meant
14900K (typographical error).

> A lot of lazy reviewers does not make this easy to evaluate.
>
> I have a scaling test for the 5950x, but the OS sucks at it.
> The study mostly shows strange scheduling choices. And Jim
> Keller already complained that another benchmark did not
> show off the best performance of his arch.
>
> That leaves us with Passmark 61379 versus 45780.
> A single point comparison, for your purchasing choice. [...]

What about lbench (multithreaded nbench replacement) on UNIX/GNU/Linux (I
also saw other tests available like Phoronix)? Those OS will be used.
So, random access memory (RAM) may affect test speeds. Maybe
relevant: AMD Ryzen 9 5950X central processing unit (CPU) is already on
this BioStar X470GTA system-board with RAM overclocked '3200' (PC3200?)
and--if possible--I'd use Intel Core i9 14900K CPU on SuperMicro X13SAE-F
system-board, which SuperMicro technical support hasn't been clear if
14900K (14th generation but X13SAE-F webpage says up to 13th generation)
or faster DDR5SDRAM even works (with BIOS update unless too many Watts) or
I should wait for possible new version/similar system-board by them, so
maybe no way to know maximum DDR5 speed yet...
My family has older 2016 personal computer (PC) local PC shop and I
built... randomly crashes/halts/reboots maybe once/week, so wondering if I
should replace BioStar with SuperMicro or if 14900K will be slower for my
usage, in contrast to 5950X, so should be my family's new one instead
(otherwise they get BioStar... so far very rarely crashing/halting). They
don't computer program nor are power users, so don't want fastest/brand-
new, but we use cores they don't, on average, and at night, number-
crunching ( http://boinc.berkeley.edu , cryptocurrency)... we both use
plain PCI slots and sometimes (number-crunching) extra display/video/
graphics cards so these are reasons for X13SAE-F or full-/extended-ATX,
even for them.

On 1/6/2024 1:57 AM, David Chmelik wrote:
> On Fri, 5 Jan 2024 10:54:29 -0500, Paul wrote:
>> On 1/5/2024 2:30 AM, David Chmelik wrote:
>>> I have a 5950X workstation, but would a 14900K one be an upgrade or
>>> downgrade for computer programming/science code building/compiling/
>>> making speed? Won't use newer AMD until there's a full or larger ATX
>>> system-/ logic-/mother-/main-board with plain PCI (currently only on
>>> micro-ATX).
>>>
>> Requires a miracle to get a PCI slot now.
>> Even though it is a one-chip solution to have a PCI slot (PCIe to PCI
>> bridge chip). [...]
>>
>> https://www.cpubenchmark.net/singleThread.html
>>
>> 14900K 4,798
>> 5950X 3,466
>>
>> Multithreaded.
>>
>> https://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i9-14900K&id=5717
>>
>> 14900K 61379
>>
>> https://www.cpubenchmark.net/cpu_lookup.php?
> cpu=AMD+Ryzen+9+5950X&id=3862
>>
>> 5950X 45780
>>
>> [...]
>>
>> Conclusion: Yes, 14800K is faster. But are you going to spend a lot of
>> money for that much improvement ? You have to be careful
>> about getting "trapped in incrementalism".
>
> I was unaware of so many tests; thanks for details! Guess you meant
> 14900K (typographical error).

I usually re-read and fix the typing mistakes. Sorry about that.

>
>
>> A lot of lazy reviewers does not make this easy to evaluate.
>>
>> I have a scaling test for the 5950x, but the OS sucks at it.
>> The study mostly shows strange scheduling choices. And Jim
>> Keller already complained that another benchmark did not
>> show off the best performance of his arch.
>>
>> That leaves us with Passmark 61379 versus 45780.
>> A single point comparison, for your purchasing choice. [...]
>
> What about lbench (multithreaded nbench replacement) on UNIX/GNU/Linux (I
> also saw other tests available like Phoronix)? Those OS will be used.
> So, random access memory (RAM) may affect test speeds. Maybe
> relevant: AMD Ryzen 9 5950X central processing unit (CPU) is already on
> this BioStar X470GTA system-board with RAM overclocked '3200' (PC3200?)
> and--if possible--I'd use Intel Core i9 14900K CPU on SuperMicro X13SAE-F
> system-board, which SuperMicro technical support hasn't been clear if
> 14900K (14th generation but X13SAE-F webpage says up to 13th generation)
> or faster DDR5SDRAM even works (with BIOS update unless too many Watts) or
> I should wait for possible new version/similar system-board by them, so
> maybe no way to know maximum DDR5 speed yet...
> My family has older 2016 personal computer (PC) local PC shop and I
> built... randomly crashes/halts/reboots maybe once/week, so wondering if I
> should replace BioStar with SuperMicro or if 14900K will be slower for my
> usage, in contrast to 5950X, so should be my family's new one instead
> (otherwise they get BioStar... so far very rarely crashing/halting). They
> don't computer program nor are power users, so don't want fastest/brand-
> new, but we use cores they don't, on average, and at night, number-
> crunching ( http://boinc.berkeley.edu , cryptocurrency)... we both use
> plain PCI slots and sometimes (number-crunching) extra display/video/
> graphics cards so these are reasons for X13SAE-F or full-/extended-ATX,
> even for them.

SuperMicro do not normally participate in overclocking.
This includes boosting RAM voltage. They don't normally boost RAM that way.
XMP requires boost (a boosted voltage is in the SPD table for it).

Biostar might do so.

The 5950X is at the top of the CPU support web page. But the turbo (power) value
is higher than the nominal power dissipation listed in the table here. My machine
draws 224 watts from the wall, flat out. SS-620GB Seasonic 80+ bronze. The CPU
may have a 142W limit by itself. The 105W number is an "indicator of class".

https://www.biostar.com.tw/app/en/mb/introduction.php?S_ID=955#cpusupport

The Biostar has a limited number of phases, and the VCore section got
a small heatsink. To engineer such things, you use a "power limiter", and
by limiting the power, that prevents thermal runaway on the MOSFETs.

You will likely find, that if you stick a finger on that heatsink,
while all cores are busy, it's getting hot. The temperature achieved,
the design engineer intended it.

The machine uses closed loop feedback, the way video cards have been
using closed loop feedback for a few generations. The hardware designer
can limit the power dissipated on the board, at the expense of the
performance of a high end processor. For example, the Biostar engineer could
ask the BIOS team to turn on "ECO mode" in order to keep the board cool
(no, they did not do that, that's just an example).

The 5950x is a "bucking bronco". An enthusiast board can deliver a lot
of power to it. The CPU temperature can soar, and the board could
be running with the limitation set by the CPU hitting 90C. This is why
you use Ryzen Master monitoring program, to see which operating domain
the system is in. (Linux users asked for a Linux version, but there
were no comments from AMD staff.)

You can install Windows, and use Ryzen Master for a few days, to
ascertain the control behavior of your setup.

I started with a mid-range motherboard, with no plan at the time to use
a 5950x. I started with a 5600G (6C 12T), which has tiny power consumption.
The feedback system on the motherboard is not taxed by this in the least,
and as a result, I was deceived as to the nature of the design. Fan speeds
do not modulate when the 5600G is running on the same board. You aren't even
aware there is a control system.

When I wanted to install the 5950X, the first thing was... "AMD recommends
water cooling". I kind of chuckled about this. I didn't believe them.

So I get the 5950X, fit a NH-C14S (blow-down cooler, some of the air
intended to cool the VCore heatsinks). Strictly speaking, that isn't enough
cooling (power capability not high enough). Well, I get the system running,
and I'm watching in Ryzen Master, and there are some "strange thermal transients".
The CPU was getting pretty warm, but every once in a while, it would shoot up
to 90C for a second or two, then return to the previous thermal regime. This
seemed to be some control loop artifact (the presentation seemed bang-bang
digital and not an analog ring or overshoot).

I could not figure out what the hell was going on :-) Rather than give Noctua
more money, I got a DeepCool AK620. It was $80 Canadian (perhaps a "sale" that day),
whereas a Noctua would have been a lot more.

https://www.deepcool.com/products/Cooling/cpuaircoolers/AK620-High-Performance-CPU-Cooler-1700-AM5/2021/13067.shtml

"The AK620 features a maximum heat dissipation power of 260W"

Based on volumetric limits, I was skeptical of the rating. But it does
seem to be a pretty capable cooler. While this is not "water cooling" that
AMD wanted, the thermal inertia with that thing on the system, the
control loop no longer allows "random 90C spikes". It can get hot, but
it no longer spikes.

My motherboard is likely still limiting the peak performance of the 5950X,
but I don't really want to push the thing too hard.

*******

Now, based on the "flavor" of that description, the 14900K is "more of the same".
To get the value for money, the motherboard has to feed it. Using a four phase
VCore main, for a thing like that, would be sinful. Certainly, with BIOS modification
(the so-called "ECO Mode" being an example), you can prevent these CPUs from
drawing too much power, and then an inadequate motherboard design can continue
to operate the board. But you aren't getting the value from the project, if you
buy an expensive CPU and then you dial it down to the 50% point so things
will not overheat.

You don't have to go crazy and buy one of the $1000 motherboards. But
on the other hand, a $90 motherboard is likely too far in the other extreme.
The $90 motherboards are a good match for the 65W CPUs, because it does
not take much to feed those. I have a 65W CPU and I've watched the
dynamics of that too.

To get the peak performance from the 14900K, you would want a decent
cooler on the thing, and a board that can provide turbo power. If SuperMicro
programs the power limitation, they'll set Turbo at 28 seconds or 56 seconds.
Some of the motherboard makers set the time to "Infinity" and they set
the power limiter to "4096 watts" (a register with 0xFFF kind of thing).
The VCore definitely does not deliver such power, but by removing a power limit
(VCore heatsink could overheat), the processor will be pushed a lot further.
Some of the enthusiasts like to do stuff like that. All I want, is for the
processor to not be "held back too much", rather than me attempting to set
a worlds record.

On Sun, 7 Jan 2024 04:33:29 -0500, Paul wrote:
> On 1/6/2024 1:57 AM, David Chmelik wrote:
>> On Fri, 5 Jan 2024 10:54:29 -0500, Paul wrote:
>>> On 1/5/2024 2:30 AM, David Chmelik wrote:
[...]
> I see what you're after -- Intel dangling their half-assed ECC again
> [...].
> While I admire your bravery going this route (to get ECC, to get one PCI
> slot), you're giving up a lot to get there. [...]

Apparently so. I just seek powerful enough full-to-extended ATX/SSI-CEB
with PCI slot. ECC is nice but not major issue (though should be). I had
ECC from 2006 to 2012 on my Tyan K8QSD Pro (SSI-MEB) workstation. Since
2012 ECC wasn't available for my system-boards (old IBM term, and Apple
used to say logic-boards). My BioStar has ECC capability but when I saw
that I'd already bought 64GB (non-ECC) RAM.
If I could find system-board (with PCI slot) for current most-powerful
Ryzen or maybe even (saving up months) AMD EPYC CPU, I'd use one, since
clearly better for building/compiling/making than Intel (lagging behind
cores/threads).
On the other hand I dislike AMD's platform security processor (PSP)
more than Intel's (can be disabled) management engine (ME) but if
building/compiling/making hours or overnight even with Ryzen 9 also using
distcc/icecc servers, then with less powerful workstations/servers, such
as capable of running LibreBoot, it'll eventually/already take maybe a
week...

> Good luck in your search for the "perfect motherboard".

Thanks! Great help reading details/tests you described. Though I'm still
learning about some of those, apparently best choice is wait. SuperMicro
system-board I mentioned clearly isn't high-Wattage enough so would only
be worth buying less-powerful CPU for, which likely is neither as powerful
as my current one nor signifcantly more powerful to justify expense.
Another computer programmer I know said DDR5 SDRAM isn't significantly
faster than DDR4 SDRAM for him to even consider upgrading, even if his PC
might be slightly old... he thought they were kind of exaggerating what
newer RAM does. I know they did various tricks to send more information
with each electrical wave, but for a long time, numbers aren't the actual
Hertz, just multiples with hardware tricks. I'd still build DDR5 SDRAM
workstation someday, but wait for system-boards that can run more powerful
CPUs full-speed.
I currently use two PCI slots: one for UMAX PowerLook 2100XL
12x18" (304x457mm) scanner (SCSI-2) which fortunately someone on Libera
Internet Relay Chat (IRC) #hardware explained can use PCI-express (PCIe)
SCSI card with SCSI-to-SCSI2 cable, and another PCI slot for my Creative
Labs SoundBlaster Audigy 2ZS sound-card with front panel with volume knobs
& audio sockets. I could use an ESI Maya44 XTE sound-card to plugin same
equipment and small volume knob boxes cables plugin each end, which is a
mess! Don't know why sound-card companies quit making front panels and
falsely call external digital-analog-converter (DAC) boxes 'sound-
cards'...