New Intel 12900k, anyone?

Have anyone bought new Intel 12900k proc? Interesting to compare to AMD 9550x. I got last one finally stable 6 months after trying everything. Looks like it was one of its bad initial samples, and many users complained. It was crashes with no tasks running and i finally got why: there exist defect there or in the motherboard in the mechanism of getting clock and consumption down for power efficiency. Seems it is unstable in undervoltage conditions. When i put on constant 4GHz speed by overvoltage (essentially make it always overclocked) it became stable with all other settings whatever you want. So AMD does not allow you to overclock more than it is sold, it is already on last drop of its power from start.

And i feel the company knows there exist the problem but keeps mum. Imagine, i posted about the problem on their own AMD website and no one from the company responded on any of our numerous similar claims. Fishy, never seen such odd behavior or manufacturer. So i will gladly move back to Intel if it beats AMD

Dan,

From the title of your post, I thought that you were offering a freebie ...

I would have declined the offer, as I'm happy with AMD.

When you build a computer yourself, you have to contend with all the compatibility issues between the CPU, the board and the RAM, and even the BIOS in the board. If you want extreme performance from the system, you are relying on the whole lot working together. You can't say which component is causing the problem until you have performed extensive tests, say putting the CPU in a range of boards and using different RAM.

If you just step back a couple of notches from the top of the range with every component, you get most of the performance and save a lot of money. You also get fewer compatibility issues.

The University I worked at always bought Intel in commercially-assembled machines from well-known manufacturers, and I was always appalled at how pedestrian they seemed in comparison to what I had assembled at home.

Eddie

Quoted from LitusSaxonicum The University I worked at always bought Intel in commercially-assembled machines from well-known manufacturers, and I was always appalled at how pedestrian they seemed in comparison to what I had assembled at home.Eddie

Eddie,

I have a very similar experence with my previous employer. They insisted on bulk buying lower spec Xeon 'Workstations' which had extremely poor computation performance. Now I am self employed, I have found much better options.

Dan,

I am not sure what 'AMD 9550x' you are refering to ? I bought a 5900X 12 months ago and had a lot of annoying problems, but not the one you have mentioned. After a few bios upgrades and not being so agressive on memory speed, it is now very stable. My limited experience was there really appeared to be a problem with silicon quality.

On problem I am finding with 5900X (and my earlier 8700K and 4790K !!) is the dual channel memory is not good enough for the thread count advertised. Using thread counts without hyper-threading appears to provide performance close to max thread count, although it does depend on the type of calculation you are doing. There are just too many advertised threads for dual channel memory to support.

Until I can understand how to reliably associate threads with cores, I will not be trying the 12900K. This looks like an added complexity that those more experienced than me (only) might solve. Look at Microsoft's first attempt with Windows 11. Just another uncontrollable issue !! Cache management is bad enough.

I am looking forward to DDR5 memory ... when it stabalises and when AMD provide processor support.

My processor was AMD 5950x not 9550x. I was assembling my own PCs since IBM PC XT, which i soldered from zero on the knees using IBM PC schematics and Intel processor manuals explaining all related diagrams for the low level components. Motherboard we made by ourselves too with friends which took us a year. If you all remember the original PC was running Intel 8088/8086 processor at 4.77 MHz. Such happened i did not have the appropriate crystal for the clock generator and i made it out of LC circuit and due to that was able to change the frequency to whatever value needed. So my PC was eventually working at around 9.4 MHz, twice as fast. The motherboard, computer box had very original design by the way. Funny, that the power supply was used from the color TV by little bit modifying it and monitor was usual small color TV with 12' screen. Given the whole our research institution of 1000-2000 people worked on two VAX11 computers in very busy timeshare, i alone probably had 1/3 of this power.

Same was with the floppy disks, the harddrives were still a novelty and opulence. Overclocked floppies were running probably 3 times faster and when they were reading/writing all people around were starting laughing hysterically so noisy and quick they worked making no mistakes by the way at least for reading.

Then i added floating point processor to it. This was a total supernovelty at that time, it was probably the first computer in the country with FP unit. At one point with two 8087 co-processors it was possible to buy small house or two room flat in the capital. To accommodate FP it was necessary to rebuild the motherboard which was done with the small board on top of mainboard which handled both processor and coprocessor. Then i tried to overclock FP unit too but that was only possible with next PC on Intel 80286 which allowed separate clock generator from major one. When all was almost done the 80486 came around 1990 which had integrated FP coprocessor which shocked all with its unseen before FP performance so from this point i only assembled all my PCs using pre-manufactured motherboards buying them on exhibitions directly from the Taiwanese vendors, they offered big discounts.

Before we used Digital Equipment Corporation Fortran on PDP and VAX computers which was very nice, then Microsoft Fortran which not bad too. Around 1990 we found an atomic bomb among compilers - the Salford Fortran FTN77 which was pirated copy of course like absolutely every software at that time. The compiler cost was exorbitant around $1500, as much as good entire PC, or 2-3 times of the cost of any other compilers. It was definitely worth it but this price tag made a bad joke to the company: no one knew even about the existence of Fortran compiler from Salford Software for at least a decade.

I had only pleasant experience with Intel processors. Twice i tried AMD and in both cases just got nothing but PITA

Dan,

I remember when 1 MFLOP/sec was a target performance. Now I am getting over 100 GFLOP/sec for real*8 matrix multiplication on 5900X using OpenMP, which is a staggering improvement (although I understand MKL gets over 300). I wonder what the next 10 years will bring ? I expect your 5950X would have worse problems with memory bandwidth ?

John,

The great advantage of lower specs all round to the IT support department is that they just work, and don't need 'tuning'. It was also commonplace for secretaries to be given higher spec machines with larger screens, because after all, they needed the power - not like engineers and scientists who apparently didn't!

Eddie

John, I was not even able to check how processor scales with number of cores and if memory bandwidth matters because that can be seen probably on faster Intel MKL which did not work on my Ryzen. Have you succeeded installing it MKL ver 19?

The Laipe matrix algebra scaled OK showing almost 4x performance on 16-core 4GHz 5950x vs overlocked to 4.4GHz 4-core 4770k. But Laipe probably was not faster than MKL so the memory bottleneck might not be limiting factor. What is good about Ryzen is its large cache level 3 so this also probably makes memory bandwidth not a limiting factor. Next Ryzens will even triple it

The MKL has support of AVX512 and new Intel 12900k also supports this AVX, it would be interesting what tests will show there but we do not know if MKL ver 19 has any improvement with parallel algebra.

Dan, you are probably reporting on your experiences with an older version of MKL when you mention '19' as its version.

More recent versions of MKL simply report a four-digit year as the first part of the version number. You can find out the version number of your MKL installation by running the following program (you can build using Intel Fortran, or with FTN95).

program mklver
character(200) buf
call mkl_get_version_string(buf)
write(*,'(A)')trim(buf)
end

The output that I see on the computer that I am currently using, which does not have the latest OneAPI version installed but an older one, is:

Intel(R) Math Kernel Library Version 2020.0.2 Product Build 20200624 for Intel(R) 64 architecture applications

I no longer have a working PC with an AMD processor, so I cannot run tests to diagnose problems specific to AMD platforms, but I expect that one can get MKL to work quite well on AMD.

Mecej4, Installation was in May, version is 2021.2.0 I spent hell amount of time and failed to install and then make it work. One click job took me days