First factor with it being faster with fewer threads is that the cores can boost their clocks higher for longer while remaining in the thermal envelope. Compiling huge projects can heat soak a system quickly, and clocks still to play a large part of performance.

@mytech6779

6 ай бұрын

Yeah ...but, there is usually more instructions per joule when running slightly slower at lower voltage, which favors more cores. Assuming the process can be divided into enough threads to keep the cores busy. This could be a case where just a few of the threads set the minimum total time making the whole process frequency bound, even though the remaining work can be divided into many shorter running threads. Reducing the number of cores will eventually find an inflection point where the longest running thread takes less time than the remaining bulk work can be processed. Another variable is the scheduler behavior, if those long running threads are being alternated with shorter threads rather than given a dedicated [virtual]core. This shouldn't be a major factor if thread count is manually limited to something less than availible virtual cores, but still a consideration. Especially if threads are bounced from core to core (or worse chiplet to chiplet) and suffer from their cache getting flushed every time. There may be thermal benefit to bouncing threads from core to core, but the net gain is probably very machine and task specific.

@mickgibson370

6 ай бұрын

I developed a processor had no clocks in 1979. The processor had 20 instructions and could hit speeds about 45 megacycles per output depending on what task it was doing. No one what it!

@reinekewf7987

6 ай бұрын

yea that is true but also consider memory controller. huge numbers of parallel calls to the memory controller can also limit your workload. i dont know how many is to much but for the controller but it depends on many things. like the internal herachy of the core bundles also speed of the interconnects, the distance to the modules, type of the DRAM chip, channels, modules per channel, ranks, the ECC factor, FSB, memory controller temperature and its power target specific not the whole cpu and many other things. cpu clock speed alone is not that important also the numbers of cores are irrelevant at one point. a cpu has its limits more then one ways. the mainbord can be a limiting factor but you can max out a cpu with only 10% load if you stress the right part of it like in this chase the memory controller. i had a similar problem with a phenom x3. the reason was the phenom x3 or x6 uses one level two cach for all 3 cores and this is bad for things hat dont need the same dataset but for gaming on the other hand it was great because L2 cach is faster the the L3 or RAM but mostly the L2 is too small and so the RAM was needed also it is a rare use case for benefit from this configuration so that is why modern cpus have only 2 cores per L2 also one core have 2 sets or ALU and som other parts for parallel work but one core can still only do one task at the time but it dont need to wait for. every command like add a or mov a , d need some cycles some of then only 1, some of them 4 or more. also be aware that a cpu have multiple command sets but a core dont understand a set but only the main feature set. for example if you give the cpu command thats convert FF to a decimel number with one command so its trigger a series of commands that arranged in the order of cpu calls thats needed for the result. so you expecting FF to dec. but in order to do this it need 100 cycles or more becuse the cpu featureset breaks the command down the the cpu native feature set so it can do the task and give you the result. modern cpusa are that complicated you are not able to programm them in assambler. but older cpus like the 8080 or the mos 6510 are simple. this means not there are not capable for doing modern stuff but you had to do more in order to do that task. the clock speed we ignore for now because it does not matter for explaining how a cpu works. becuse of the fact cpus are getting faster and programs can be even more complex because you do more cpu cycles in the same time. so feature sets are created to do tasks. and this means you dont need to write the whole code for the same task because you have a sort of shortcut for ist but sins cpus are that fast and some shortcuts that cycle consuming. the cpu manufactures thinks about this and though why i need to wait for the result and do nothing in the meantime so i can slap a second core on it and do other stuff in the meantime while i wait for the result. so a special feature set was created in order to select a core for your task and with more and more cores this can be really complicated to track down wich core is done and wich core is free to do stuff. if you try to programm a cpu on that low level with modern cpu you going crazy because you can do the same simple task with different feature sets and still cant geht the full potential of the cpu. but if you use a language like c# or python the compiler or interpreter do this for you and decides wich feature set is the best for and you can use the full potencil of the cpu minus that what the os needs. it is not that simple in the end but you get the idea

@lolilollolilol7773

4 ай бұрын

So this means he'd need a larger heatsink, or use watercooling, if that is true.

@reinekewf7987

4 ай бұрын

@@lolilollolilol7773 this is not that simple. he could increase the clock speed by lower the voltage and a bigger heat dissipation system and no power limits, but the performance boost is negligible and wont change anything if the limit lays on a other position internal of the cpu. a cpu has more then one clock speed. changing only the core speed does nothing except for a tiny bit more single core performance and a big power draw increase. boosting the memory controller is also a bad thing. the max memory speed depends on the way how the traces are laid on the motherbord and how long there are and how many banks/ranks you are using and what traces are nearby causing may trouble. this is always not that easy building a high performance pc. that is the reason why motherbords are that expensive. the main reasons are functionality and functional stability. the last one is the expensive one.

You old guys need to stick around doing stuff like this & sharing knowledge as LONG as you can. The IT space has spiked in complexity over the last ~25 years, and there's a huge appetite in the younger generations to learn from the OGs. Great stuff as usual, thanks Dave 🙏

The performance of Hyper-V when compared to bare metal is indeed impressive!

5 seconds is an eternity when you are trying to decide whether to cut the red wire or the white wire.

Dave, we'd like to see more C / C++ stuff on your channel 👍

@djarcadian

6 ай бұрын

I want to see more skydiving content first.

@dwight4k

6 ай бұрын

This is not an attack. But what do you make of the security issues?

@theintjengineer

6 ай бұрын

@@dwight4k, aheem... you don't have them in the first place? You study the language behaviours, follow best practices, test your code at each moment, use analysers and sanitisers, etc.? I mean, we haven't ended up with any of these security issues all these folks talk about in production. Is it difficult to get there? Yes. But at some point you set up your tooling, "models", etc., and get used to them, and so on. Can you guarantee 100% that there's no issue, whatsoever, with your application, and that it is perfect? Well, good luck with that.

@andrewsmilie140

6 ай бұрын

I would like to see you study threading performance in detail and when/where it makes since to use them in a project. I know it is good for prime number generation, but what about strings? File I/O (kind of silly, but could be interesting). Custom structures? Structures with Bit-Fields? Compiler alignment (byte/quad) etc.

The ratio of L1/L2 cache to RAM is lower when you use more RAM. The fastest parts of the RAM are those that have already been cached (they are "hot-spots"). Unless the kernel deliberately recycles those hot-spots to apps, it will cause more cache invalidations and cache loads when the backed RAM is larger. The effect would be similar to disk caching. A 2 GB disk backed by a 1 GB RAM cache would be faster to read repeatedly than a larger disk backed by the same amount of RAM (ie, the throughput would be higher when the ratio of cache/backed storage is higher). EDIT: also, denser ram requires more clock cycles to switch ranks, because there are more ranks that can be switched. Even though the transaction speed spec is the same, the extra clock cycles spent on rank switching means the total throughput is less. I don't think on its own this effect accounts for a 5% throughput difference, but it could be 1-2%. Gaming benchmark reviewers like Hardware Unboxed have tested this quite extensively.

@isoEH

6 ай бұрын

I like this reply.

@LiveType

6 ай бұрын

I think the primary issue is that there are clock speed differences not being observed. Clock speed is king. It makes everything go faster. Case and point intel macbooks. With mac os, the macbook boosts higher than if you install windows on it due to tuning by apple. If you want the same performance you have to do the tuning yourself inside windows. I'm guessing the windows build had some OS tuning that made that specific configuration boost the highest.

Having watched some overclocking/tuning content, I learned that there are memory timings that affect performance of a modern system more than the 4 that are usually advertised on the spec sheet. If I remember correctly, these other latencies are often set by the motherboard (no recommended values for these timings are contained in the profiles present in the sticks themselves), so if the memory is unknown to the UEFI, which is likely with a random set, you can expect looser timings and worse performance then with a set that is on the QVL, or, in this case, came with the system. So it's not even about one set of RAM being better than the other, it's the system (motherboard/UEFI) being more familiar with the RAM supplied by the OEM and being able to get the most of it, giving it an edge over possibly as capable Micron sticks from Walmart.

01:12 🖥 Dave received an HP Z6 workstation for evaluation, focusing on compile times for large software projects, particularly interested in compiling the Windows NT codebase. 03:12 🖥 Upgrading to 192GB RAM due to memory constraints during builds; initially struggled with 128GB causing paging, hindering performance. 05:45 🖥 Limiting threads during builds to 128 improved performance, avoiding memory constraints; power limitations on the system prevented utilizing all cores effectively. 11:43 🖥 Benchmarking with and without hypervisor showed comparable performance, indicating no significant performance penalties when running under hypervisor for compiling. 13:21 🖥 Increasing RAM to 384GB didn't significantly improve compile times; unexpectedly, it slightly slowed down the build process compared to the original 128GB configuration. 14:31 🖥 Despite issues, Walmart's seamless return process for the RAM was appreciated, showcasing an efficient return experience.

@IvoPavlik

6 ай бұрын

Thank you, good sir!

I love your content as an engineer myself whose been in IT in one form or another since commodore days, and with a son with autism (who is also very interested in topics like what you cover)…I appreciate all that you do. I don’t know if you take requests, but I’d love to see your take on some ESP32 topics. I’m late to the party and have been using mostly ESP8266 up until recently for my projects and I’m just now getting to seeing what the shiny “new” ESP32 can do.

Love how you explain everything in great detail.

Would appreciate if as you are talking through the numbers, they were on screen so I could also read them and compare them without having to keep them all in my memory.

Sub 20 minutes is actually quite impressive. my 5900X 12c/24t with 128GB RAM takes around 1h 20m to 1h 30m for an ungoogled-chromium. But thins includes the unpacking, patching and installing time. Since I use gentoo, this is compiled every time and the built time are logged. As for the speed stagnation, you might be power or heat limited so the more cores can't keep their clock rates or you might hit a memory bandwidth limitation. Had the issue once, ran perfectly fine on test systems but on "real" systems it was an order of magnitude slower. On the real system the memory controller was already near the limits with the "base system". My Software on top just got like 10% of memory bandwidth and since the tool made stuff on RAM, was hit by that limit.

Would be wild if there was a bandwidth limit such that using more memory is slower on account of taking longer to transfer more data

"at least my workbench supports a 128GBs of Ram" 🤣 I do so enjoy your channel!

You may have maxed out the TLB or page table capacity with the extra addressable RAM. More memory isn't always a boon due to the capabilities of the memory system, as you've discovered, since you're not running directly against physical memory. Exploring these limits is probably worth a video series on its own.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you…

My guess with the slower memory is that there are optimally configured timings with the factory installed RAM and fallback timings with unknown RAM. Maybe a bit like XMS and EXPO? Also, the installed RAM might have been carefully chosen based on benchmarks to be faster, so there could be some quite obscure memory timings that are making the difference, and then factory overclocked since they are known, which isn't happening with your replacement memory.

Kep producing amazing content! I'm just north of the boarder, and I'd love to buy you a beer in person just to say I met the amazing Dave.

Buildzoid might know the answer on your RAM questions, if he is out there, hopefully he gets involved, great video once again Dave

Yep, we used to compile a Linux kernel on a 486dx2-66, with the linux source residing on an NFS share hosted by a Sun Sparc1. Oh yeah, the whole building had just 1 "multi port bridge" , meaning we easily had some 100 hosts all sharing the same ethernet segment That compile was an afternoon job.

Much love Dave, you keep us fellow computer enthusiasts occupied and entertained. You've completely got me out of my drought of writing code with watching your videos in the background.

Fascinating to see that build task against those hardware parameters. BinTime will soon have a boatload of RAM coming to a bin!

WAAAAAAAAAAAYYYYYYYYYYYY more interesting, and relevant (which isn't saying much, but wait for it...) than any racing performed at the Irwindale race track. Yeah, in the dimmest corners of my aging (yet still useful) brain, I do recall going to the drag races, long long ago.

It was cool to watch the entire setup process. Would have been cool to have a running clock on screen (or in post) to show how long each step took in real time Also, you probably did 3 separate commands just for the video, but you can just type 'sudo apt install program1 program2 program3 ...' instead of each individually

Good stuff. Thank you for this.

How much time does it take to recompile once you change one source file? Good video btw

@DavesGarage

6 ай бұрын

Linking takes about 10 seconds, I believe!

@Dave's Garage Dave nice to know that you can admit when you don't know the reason or answer to a problem. T.B.H. I look forward to a video from you in part to ASD and Microsoft and your generally likeable way of doing your videos. Keep on keeping on, you rock

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you…

Hyper-v for the win. Any way you could do the same with esx underneath? And what would this look like on a monster (bare metal) blade in azure?

Dave, are you using a CRT shader on your terminal? Or is it just the font?

The memory may be of a different RANK. There is single rank, dual rank and quad rank. There is a speed difference between them as far as a I know.

@dysphunkional6295

6 ай бұрын

Dave mentioned that they are "2 rank dimms" but didn't mention if the original sticks are single rank or "2 rank" as well. If the originals are single ranked and all the other timings are the same then the larger sticks will be slower due to the ranking.

Have you tried compiling chromium under windows? It would be interesting for comparison. While the code may differ some per platform, it would give an idea of the overhead for the operating systems.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you.

Its good to see HP didn't waste their investment in Compaq

Lots of horsepower in that garage !!!

Thank you sir You're a videos are always a pleasure at least the ones that I like are, like this one! I'll have to fix the typos later sorry

I'm old enough to remember when I "upgraded" to 128 kilobytes (yes kids, kilobytes) - on the Sinclair ZX Spectrum+2, the successor to the Spectrum+ which was in turn the successor to the Spectrum itself, which I believe was known in North America as the Timex Sinclair 1000. Of course in those days, for home users, upgrading the RAM meant upgrading the whole computer, much like on a Mac in...2023. Oh. I then upgraded to an Amiga 500 (on which you could, and I did, upgrade the RAM from 512K to 1M) and then the 1200 (yes, I was a bit spoilt as a child!), so like you have fond memories of those systems. And the rest, as they say, is history. It's VERY interesting to me that Hyper-V has such little overhead; when I run things like FreeBSD in VirtualBox on the Linux box I'm typing on now (which is by no means a speed demon but is usually fast enough for my purposes), the speed penalty is significant. Anyway, long story short, very interesting video, Dave, as always.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you..

i m pretty sure the bios trains the memory with looser timings when using mem sticks with denser capacity

I wonder how an Optane scratch disk would affect the swapping, Dave. Got any hanging around?

My knowledge of hardware is less than rudimentry basics but at a guess (and, honestly, I am literally guessing) it's to do with the power draw being limited to 350w. Also, I believe, limiting in this way is not an exact science, hence the slightly different results every time as well.

How does it perform when running without hyperthreading or just a subgroup of all hyperthreading threads? I've found if a process maxes out a core, running it on all threads of that core will actually make it slower. But my testing was done on older hardware.

@Takyodor2

6 ай бұрын

Sorry for being pedantic, but hyperthreading is Intels name for simultaneous multithreading.

Throttling? Did you compare the temperatures between the various hardware and thread count options?

You were running a 4th gen i7 up until 3 years ago??? Makes me feel not so bad just updating to an X99 chipset rig.

@mistahHeffo Exactly, heat.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you

"build-essential(s)", catches me every damn time.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you..

Clicked on this for the racing game stayed for the benchmarks :D

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you.

20 minutes on such a beast. Insane how large modern browsers are.

I am not alone! I fricken love compiling things - even when I don't need to. I'm not quite so bad that I use Gentoo, but I do love compiling stuff.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you.

13:47 Might be CAS, in went CL48, what was the timing of sticks which went out? Not sure HP actually supports 48, it's perfectly valid size, but they might have done some weird stuff in mem controller too allocate physical chip on stick to different channel (but then hit would be much bigger I'd thought) + you said it's 5200MT, but theirs spec says that for some config frq goes from 4800 down to 4400, not that correlates with slowdown observed, just a few random ideas (btw it's build, who needs ECC for build 🤪)

At a hardware layer, I would look at thermals and power which you have some control over. Memory bandwidth by a limitation. Then we can get into things like BIOS and motherboard things. As a software layer, good be a kernel issue. Would be interesting to compare against different OS's or distro's or your own kernel build.

I wonder how well optimized the compiler itself is....

Hey Dave… have you ever done a video on the video game called ELITE from the 80’s from Britain. And maybe there computer back then vs what we had in the states ? It’s kinda cool how they made the game in assembly with only 21k memory

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you

Interesting video and surprising results, thanks for sharing! Making a couple of guesses as to why more memory gave a slower result: - whilst there are 8 memory modules we don't know how many dedicated channels there are to read and write memory. Perhaps more memory is leading to more contention on the channels slowing memory accesses - you have more main memory backed by the same amount of L1, 2 and 3 cache which may be leading to a higher rate of cache eviction, forcing the CPU to wait to read from a slower cache or even main memory I suspect that the cache explanation is more likely but who knows, perhaps the answer is something else - I'm be interested in knowing if you find a way to investigate further

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you.

Did you confirm the ram was running at it's full speed? It's possible that the ram kits had different rank populations which could affect performance.

@volodumurkalunyak4651

6 ай бұрын

It really should run st full speed. It is 1DPC of a 8 channel JEDEC DDR5-5200 memory setup. No tweaking in BIOS like enabling XMP or manually tuning memory speed is required. What is required - more memory refresh for higher memory amount. More memory refresh - slightly slower memory.

@micThurrr

6 ай бұрын

@@volodumurkalunyak4651 5200 MT is the speed set by BIOS default. Dave should have checked task manager or CPU Z to confirm ram speed. 99% chance the bios reset to 5200 MT when it discovered new ram was populated.

@volodumurkalunyak4651

6 ай бұрын

@@micThurrr CPU does support up to DDR5-5200 JEDEC. Of course XMP or manual memory OC does allow for faster memory speed, but it comes stock at 5200, even with stock memory.

@micThurrr

6 ай бұрын

@@volodumurkalunyak4651 exactly I bet it was running at 5200 MT

@volodumurkalunyak4651

6 ай бұрын

@@micThurrr at 5200MT/s for both kits of memory. I made a mistake assuming it would run at 5600 with the 384Gigs.

If you have a chance before returning the RAM you could run a program like HWINFO or CPU-Z (probably have to run on bare metal) to check the actual timings and see if the new kit is running at its rated speed. This may be fixable with different BIOS settings (if it even allows it). Fully loaded slots with dual rank chips may be forcing it down to slower speeds. Would be nice to retest the same RAM with only 256GB installed as well in case it's either an electrical load limit on the bus or some kind of cache or page table limit when accessing over 256GB.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you.

I'm still using a 4790k, and while it shows its age with multithreaded tasks, its still going about 4.7 Ghz so I'm hanging in there. Can't wait to figure out my upgrade path.

@toby9999

6 ай бұрын

I upgraded to a i7 12700. Gave the old box to a family member. Speed improvement was good without breaking the bank.

@fightingfalconfan

6 ай бұрын

If you can swing it; get yourself either a 5000 series Ryzen or a 12/13 gen intel. I went from a Ryzen 5600g to a i912900k and what a difference!! I also went from DDR4 to DDR5 plus pcie gen 4/5 and man all of it combined for a wicked fast system Ryzen 7000 just like Ryzen 1000 chips have growing pains with issues such as ram compatibility. Ryzen 7000 is not backwords compatible with DDR4 memory. So if you do upgrade, you may want to go Intel until your ready to take that DDR5 plunge. Either side you choose your looking at a motherboard swap and possibly a PSU swap. Good luck and only you can decide what your upgrade path needs to be.

@thelegalsystem

6 ай бұрын

@@fightingfalconfan yeah its gonna be a whole thing, new mobo, upgrade from the DDR3, new processor. Graphics card and ssds are relatively new, or at least only a few years old. My PSU is way overspec because I used to run 2x290x, but it's still almost 10 years old so I should probably replace that too. Its all kinda on the back burner right now while I'm in between jobs, but whenever I get around to it, it will be a blast building my own PC again rather than just computers for friends and family. I guess my media server counts but that's a whole other headache.

@PineyJustice

6 ай бұрын

@@fightingfalconfan DDR5 is cheap now, about 100$ for 32gb. It's also about double the speed of ddr4 so a very worthwhile upgrade for things like compile tasks. It just doesn't make a big difference for gaming.

@timothygibney159

6 ай бұрын

I got rid of mine 2 upgrades ago lol😂. A 13900k is double the speed and no micro stutter in games like cyberpunk

Hey Dave! What was Seattle like for you tech people during the grunge years? It might be a fun video to hear what was going on in the city and in MS in parallel, if the scenes interacted, what people thought of grunge being what the world thought Seattle was, perhaps oblivious to the strong tech industry!

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you

Nice some interesting stuff 🍿

The drop in performance when doubling the ram is probably related to memory bandwidth. If the memory bus is not wide enough to support the memory calls when pressure increases it will form a bottleneck. The increased pressure of using all the cores and all the ram most likely resulted in a bottleneck and some thrashing. The original ram may also have a higher spec and changing to the other sticks can change timing, latency and other performance aspects that can compound and result in worse performance.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you.

Chromium is insane. I did a clean chromium build on an i9 notebook and it took 6 hours.

Perhaps you are limited by I/O now and disk speeds. It could be that a large write cache is needed, so that inode creation and insertion is not an issue.

@timdowns9253

6 ай бұрын

Perhaps try a ram disk, and or different nvme drives for read and write ops.

great video

It might be worth setting the memory at 256 instead of 384. There may be some odd effect of running a non-miltiple power of 2. Just a guess.

Commenting purely for the engagement. Dave's content is wildly above me, but it's still interesting

Running a 7950x and thought I was doing good. This thing is crazy!

@etherboy3540

6 ай бұрын

I have an 8-core Xeon for my home workstation, it was teh hotness 5 years ago. Time flies!

@davidspagnolo4870

6 ай бұрын

I'm running a 7950x as well. I don't know what I'd do with 160 more threads.

@Takyodor2

6 ай бұрын

You're doing fine 😁

Any chance the 'bigger' setup caused thermal throttling or more cores=lower clock?

LETS GOO he did it yay

Nice video.

Did you check such things like core clocks, throttling and so on?

Very interesting🎉

one handed dude nice !

Possibly hitting a memory bandwidth ceiling? At that point extra threads is just more overhead?

@brylozketrzyn

6 ай бұрын

I don't know how the build system is designed. If it relies on storage or pipes. For many tasks it is actually the IOPS ceiling that you hit with an increased thread count.

@SianaGearz

6 ай бұрын

@@brylozketrzyn But why 128 threads, same everything, different RAM, more RAM, slower?

@brylozketrzyn

6 ай бұрын

@@SianaGearz because with more threads you have more time spent to wait for IO. I ran benchmarks on S2D recently, latency skyrockets with 96 threads doing just 3IOs simultaneously and with 31 it is measured in seconds. All flash servers, RDMA, 100Gbps direct connection. Most build servers use RAM drives or pipes to speedup build, because RAM latency is nothing even compared to multicolumn NVMe storage.

@SianaGearz

6 ай бұрын

@@brylozketrzyn Reading comprehension. Not more threads. Same threads.

@berndeckenfels

6 ай бұрын

@@brylozketrzynthe puzzling part is that even with 128 limit the old ram was faster. But I would agree could be a rank or (secondary) CL timing thing,

I do not like most of HP's stuff these days, at least on their consumer side. It is nice to see they put more thought into their workstations still.

Hey Dave, how about trying thin on Ampere's Altra and Altra Max Arm processors that HP offers in their HPR ProLiant RL300 servers?

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you.

The i7, 4770k is arguably intels best cpu, it stood the test of time better than any other cpu. We only saw real upgrades in the last few years, but that cpu would still handle 90% of use cases. What a great processor

@timothygibney159

6 ай бұрын

It's slow as hell in game. Cyberpunk 2077 is a stutter chop fest even with a top GPU and it would bog down with 30 tabs from Chrome easily

Seems a lot of muscle. Tried running 2 VMs in parallel vs one machine compiling 2 times ?

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you.

The Windows savior. Legend.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you

I think the main issue is the cross-cpu link which is the bottleneck. If the compilation shared across different nodes, and it needs to go other CPU to the other bank of memory to fetch data, that will be slow.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you..

Great now there's drool on my phone from looking at that ram.....

Have you checked if the original OEM is is single rank? Because AFAIK there’s DDR5 16GB single rank sticks, but no single rank 48GB sticks.

@SpaceCop

6 ай бұрын

This gets into territory that took me a while to research last time I upgraded.. Kids.. don't sleep on the single rank

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

@@SpaceCop I wanna talk to you..

I recall reading something about only being able to use dual channel speed if you only have a certain setup of memory modules on the ram. Your higher capacity ram probably isn't compatible with dial channel mode. ill check when I get home. I'm on my phone right now.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you..

Speaking of paging to disk, could you make a video on Optane?

@DavesGarage

6 ай бұрын

I've thought about it, but "you can't get that stuff no more", can you?

thats what i always tell people, modern cpus have turbo boost modes, if you use too much core/power they will clock down, depending on the application, sooner or later you will have reduced performance if you use too many cores/power

Bass pro shop was out of DDR5 ECC RAM :(

@dave what os do you run as the main OS? Win 11 pro?

My intuition was that HyperV vs baremetal would be a 5-10% speed hit. Seeing its less than 1%. I’d be interested to see how well this holds for ESXI, VMware workstation, virtualbox, and proxmox. Any chance we will see a virtualization drag race?

@timothygibney159

6 ай бұрын

The reason HyperV smokes is it's a type 1 hypervisor and guests access the hardware directly. The others are type 2 emulated in software where peripherals are emulated and hardware calls go through a layer instead. Vmware workstation will be much slower and bigger. Esx is also a type 1 where the hypervisor runs on the silicon directly

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you..

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

@@timothygibney159 I wanna talk to you..

Lots of focus on the memory but would be interesting to see the builds with fewer cores. Sometimes multi-core systems can trip over each other accessing memory and other resources such that the optimal number is lower than expected. Maybe do 64 and 96 cores and see if there is a useful comparison.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you.

Hey Dave, figured this idea is better in your hands: It would be nice if task manager in its "mini view mode" if it could lay out long ways instead of Tall So I can lay it along my deskbar or anywhere else while squeezing it as tight as possible I've always enjoyed just watching Task manager tick to try and theorize what each thing could be doing, why it was, and if it was really necessary

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you…

Compiling is surprisingly I/O heavy task, so the more RAM for the disk cache and faster NVMe disk will help a lot. But this comes after having a *fast* CPU.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you

What was the ram product you used?

“Uh oh” says the wife! Could this be the beginning of an interest in FPGA development and Hardware Description Configuration Optimization? This might be the next interesting step for a guy like you!

I read that Linux has a hard limit on how many cores can be used for scheduling.

Rather than disabling cores maybe it would be much better to disable hyperthreading?

I would run a memory focused microbenchmark first and foremost to confirm the result is not due to that build system shenanigans. I'd also check how physically the memory has been installed, and fetch all the memory related statistics and configuration, including hugepages configuration and numa configuration, check proc/meminfo and numactl --hardware as starting points.

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you..

amazing that a Brazilian made htop x3 I'm also Brazilian

Try compiling clang or gcc next

About RAM speed: Both RAM configurations used all CPU channels and came with the same timing / bandwith. My theory is, that the DRAM refresh cycle requeres more bandwith in the large configuration, therefore cutting a little bit of the bandwith compared to the smaller config. What do you think about this theory?

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you

You mention ECC memory. How would you setup a test that would detect if ECC memory actually helps with fliped bits?

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you.

Someone should tell him about Gentoo.

There's a phrase, mechanics would make the best engineers, because they've seen how parts fail.

@MarkBarrett

6 ай бұрын

Rubber is historically known to wear out. Engineers may have decided it should last "so long", but no, it wore out.

I want a 7995WX. The amount of stupid stuff I’ll be able to do, while doing other stupid stuff, and doing MORE!!!!!

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you

I find this fascinating since I am not a programmer. (my brain doesn't think that way) I had watch a Mark Russinovich about how Rust is more secure than C/C++, how does Rust compare in speed preforming the same tasks as C/C++?

@ON-TELE.GRAM-DavesGarage1

4 ай бұрын

I wanna talk to you

Go Saints ⚜;)