Please stop giving Tony Yu airtime/coverage. He is obviously misusing the delidding device. After being corrected the first time he should know better, so I must assume that he's now intentionally breaking chips for the publicity. Stop giving him that, he's being reckless and people need to start ignoring him.
yea what killed the socket is how thick the heat spreader is and the compatibility with am4 coolers... the socket of thread reapers and epyc was just better.
That Tony Yu seems like an idiot trying to trash your delid tool reputation? He obviously does it wrong on purpose, one run then hacks at it with a scraper?
He's not stupid (breaking it because he doesn't know what he's doing) Or malicious (trying to ruin reputation) He's just annoying and being a tiktokker out for attention.. and everyone took the bait.
Gamer Nexus said AMD told him to use the X3D special chipset driver regarding Ryzen 9 dual CCD CPUs. Just 2 days before his review went live, and he had to throw out a lot of results. Maybe this can have something to say for your results
Same happened to almost everyone, apparently Jay2cents did a step by step install guide (I'm not watching his channel for that) but GN and HB Unboxed did a review using the installation procedure of the 3D Vcache parts and the performance was "normal" Wendell at L1Tech noticed performance issues under Windows and faster performance under Linux 😂
yeah, was curious if der8auer got the same recommendations related to chipset driver and core parking etc (similar to 7950x3D and 7900x3D requirements) as I don't think he mentioned this specifically in the review, but did say he reinstalled windows fresh. But he did analyze that games were using the proper cores with higher freqs.
After watching the Gamers Nexus review, I had the same thought, as it wasn't mentioned in this review. However, I believe Roman and Steve are somewhat close, so I assumed that when they said they 'reached out to other reviewers,' they would contact each other. I guess we'll find out soon.
@@MrMoerli he might have reached out before Steve said anything and Steve simply didnt have the time to just communicate around what he was told (assuming he was permitted to share it begin with) in time of this video to go out.
18:13 I don't know about new cpus but I remember using liquid metal to clean the solder from the die (9900k), I put a little bit on the die, moved around with a q-tip every 3 minutes, and then polished the die with a tiny amount of polish. Perfect mirror finish in the end. Zero scraping.
So with Tony, maybe the first time was user error, but now it seems likely he's deliberately doing it for the publicity. Grounds for defamation, not likely, but scummy none the less. I definitely would not give him more attention at this point in content at least. Maybe a strongly worded letter from an attorney though...
@ivangerginov5648 Trade libel is a thing, but difficult. Once is incompetent, twice seems like it's for the views. Tony is Asus' China manager, so there's also that issue.
Not exactly, when a corporate representative purchases a product to use it in a defamatory way, for indirect promotion of their brand, trade libel can come into play.
All reviewers need too do this going forward.Huge avg FPS that in reality is flip-flopping all over the place is no longer fooling gamers like it did back in the day.We want consistency.
Must add, as I have seen it many times, the matte or dull look on the dies after delid is remaining indium solder. If you don't polish it to remove absolutely all indium it will still alloy with LM, leading to a little 'dry out'. Unless you want to apply a really liberal amount of LM I recommend polishing the dies with cape cod or similar to get a perfect mirror on all the chiplets before putting on LM.
Zen5 is having issues. Wendell found that Windows is doing something seemingly wrong. Ancient gameplays found that core parking mostly hurt performance in his stuff
Nice work as always Roman, looking forward to seeing some proper direct die Results. As for Tony Yu, that clown is simply looking for his 5 minutes of fame and as others have already stated, needs to RTFM! 😒
Romain at 21:43 - for fix very long boot time - try 2 parameters in memory timings settings : 1st - Memory Contex Restore = Enabled, 2nd - Power Down Mode = Enabled
@@lolschrauberinstability only happens when memory context restore is enabled but power down is disabled. As long as you enable power down, there won’t be an issue with stability.
@@JoeMama-yl1owprobably about the time amd almost doubles the tdp of their chips after the media lauded them for their power efficiency, despite single digit performance gains. How is AMD blowing this so badly? This is an extremely disingenuous launch, when all they had to do was be honest.
Thanks to Intel forcing me to learn what exactly a load line is, and how Intel's use of Vdroop is different to Nvidia GPU's and AMDs everything.. I do find myself wondering if on the occasions performance seemed lower for no obvious reason, was there any difference between effective and reported clocks on the "more efficient" part? (If that's even a thing on AMD)
This isn't even a problem with not competing with X3D as some have stated. The 7950X competed with the 5950X, the fact that you don't even see a 5% increase is alarming and points to even X3D being trash no upgrade as well. This isn't just a bad launch it is a failure. All they had to do was deliver similar to what they did in the past and they couldn't even deliver 15%. The stagnation is real, they said Intel is dead what are you gonna buy desides us.
tbh these cpus would have been in development since before the intel saga. These would have been started to get developed a couple years ago when they where doing really well and probably let their quality slip.either way its probably close to an industry first when generational improvements where negligible. we don't even get any new technologies or features to make up for it.
Unfortunately this may be pointing to the hard truth that we are reaching silicon's limit of performance. Dropping in size isn't giving us beter performance anymore and may even be causing problems that can't be solved unless we start making cpus out of graphene.
@@darkkingastos4369not the silicon limit of performance, but the limits of size. You can't make things any smaller. We are already at traces and transistors being 3 atoms wide. You can't go any smaller. Technically you can, but you need to drop frequencies and use a lot more volume of silicon, so it's not that great.
I think it's partly because AMD skipped a learning generation , because Zen 3 was supposed to get a Zen 3+ desktop variant. AMD's memory controller isn't particularly strong to utilise DDR5 having 4 dimms populated, it runs at 3600mhz, which my B550 board can do. I like the thermal behaviour of Zen 5 over Zen 4, but that's only because AMD needlessly changed things from Zen 3. It's a failure of a product Zen 5 built on an expensive poorly selling AM5 platform, I agree.
@@darkkingastos4369 That isn't what happened here, at least not provably. Just shrinking the size alone can give performance but we don't know what changes they made under the hood. This looks like stagnation to me. It is worse than when Intel only targeted 5% IPC gains yearly. This barely hits 3% increase over the 7950X. That is what you call an abject failure. It would have been a fine launch if they had not promised 15-40% increases which were not delivered or priced everything so high, so old parts destroy the new ones. Again a marketing fail still. If the X3D variant cannot even beat the 7800X3d by 5% we will know Amd is cooked. Not because they failed to deliver on performance, but because they priced it so high when it fails to deliver. They need to fire everyone in marketing/pricing and start from scratch.
Says a lot they didnt have the competence to figure out there's an issue, unlike L1T and KitGuru among others. As much as 20% improvements over 7k on linux or with scheduler tweaks in doze, for the same power
@@madb132 Windows is certainly an issue BUT its on AMD to sell a good product. Intel put the effort into ThreadDirector to fix the scheduling situation, while AMD hasnt and it shows.
Because it's too much work to expect a normal user to do it, a normal user will try and click on PBO but anything more that that only a very small amount of people will even attempt to try.
I see you also have picked up the planted terrarium hobby! I feel it goes hand in hand with PC building. I also picked up aquascaping of smaller nano tanks for Betta fish and Guppies.
Thank you Roman! I am thinking of getting a 9950X and delid it like my 7950X is. Now I think I will hold off for a bit to see if this gets resolved. I have 3 different versions of your delidding tools so I know they work perfectly.
PBO is very weird. On ryzen 5000 at least ive found best performance leaving EDC near stock but unlimiting PPT and raising TDC. Theres also numerous bugs with it across different bios versions
Zen 5 has severe Memory Bandwidth/access limits The bottleneck is Infinity Fabric in I/O die which is the same as for Zen 4 I/O die was even bottlenecking Zen 4, that's why X3D cache is so fast Steps to do: 1. Re-Architect I/O die Infinity Fabric 2, Switch from obsolete DDR architecture to HBM2
Regarding your second point: Do you like paying 1000$ for 32GB of HBM2? That memory is in high demand in the server market and will be super expensive for the PC DIY buyers. not a good solution.
I'm not trying to buy my motherboard and ram as a package because the two usually need to be soldered together. And I don't want to pay $2000 for a motherboard and ram. No thanks, I'm fine with DDR5.
Everyone thought we were likely to see the same improvement from 9000 as we did from 7000 over 5000. It seems the big bump from 7000 was the entire AM5 change and now that’s baked in perhaps 9000 hasn’t got that much to give.
The guy that delidded and broke his own CPU had a grace of a drunk truck driver after 18 beers That's what happens when you didn't pay for your own CPU, but got it for free I am also trying to understand that leak benchmark stating 9950X can do 48000+ score in Cinebench R23
I can explain the the benchmark result in CS2. Source 2 engine utilize as many physical cores as possible, as a result you are suffering from cross ccd latency. To avoid it you should add -threads 8 or 9 in your launch option, this way it works normally. Same for 14900k and all other intel CPUs that have E cores, you should use -threads X, X represents the amount of P cores you have in your CPU. X+1 also works similarly.
According to hardware leaker Chi11eddog on X, an upcoming AGESA 1.2.0.1a Patch "A" update, will reportedly increase the TDP of the two Zen 5 chips from 65W to 105W. EDC on the Ryzen 7 9700X and Ryzen 5 9600X
@@TheBackyardChemist Fair enough, and apologies, for my apparent misunderstanding. The patch is making a major difference to power, huge! Enabling EXPO and specifying PBO +200Mhz, helps a great deal. (Precision Boost Overdrive)
also there are handfull boards have their CPPC remove during SOC incident year ago...its missing on 50% of amd motherboard vendors ...something AMD failed to mention to their partners that its a must to have for windows 11 core parking software driver support for gamebar mode
9:35 SMT disabling on Cyberpunk and falling BELOW 9700x suggests that cyberpunk has around 16 active threads, the scheduler is spreading them evenly across all 16 cores. And the switching fabric is BROKEN (very very SLOW) like on Tyzen AI HX 370!
The thing about all the hoops other reviewers had to go through to do core parking in games just shows how bad the core management in windows is. In linux you don't have to do any of that crap and there the 9950x really kicks butt in some scenarios, especially compiling. AMD should have had a chat with Microsoft and let them do all the required updates before shipping these CPUs. Apart from a BIOS update people shouldn't have to do anything for a new CPU.
Since you mentioned slow boot time I can’t help myself but to ask here, I have had a problem with the boot time on my 7900X since I purchased it new… I have an MSI MEG ACE X670E and my memory is 6400 CL 32 TridentZ . My problem is there regardless of with or without XMP enabled. So when it’s on 4800MHrz it would take around 30 seconds to post and around 41 seconds each time until Windows boots. And when it’s on 6400 XMP it takes around 40 seconds to post and 52 second until windows startup . Now I’ve just played with the timings and put it on 6000 CL 30. I’ve enabled “memory context restore” which followed that I had to enable “power down enable” as well. Now the mobo posts around 25 seconds and OS starts at 35. At this stage I cannot put the pc on sleep mode cause if I do the mobo won’t post. If I hadn’t enabled “memory context restore”, any change from 4800Mhz to either XMP or 6000Mhrz would also render restarting the PC impossible cause again it wouldn’t post. I changed the motherboard to an Asus X670E-E, boot time still didn’t change, changed the memory kit to one exactly like it, didn’t change, changed the M.2 drive from which the OS loads, didn’t change (obviously cause the problem is the time for it to post, not the OS). The only thing I haven’t changed is the CPU so that’s what I’m left with. I really can’t understand it but guess the problem is with the CPU
Having to communicate across CCD, or migrate threads from one CCD to another, to compute a game frame could be the same old reason why dual-CCD parts lose in some games vs single CCD parts. The latency with any given app running across multiple CCDs is usually worse than the same app running strictly within one CCD, so for extremely low-turnaround-time ("low-latency") tasks like generating a game frame in just one 700th of a second (~1.42ms), +/- other random sources of system noise, every bit of latency matters there. Maybe a UEFI/firmware/OS update or two can help with the scheduling, I'm not sure.
Yeah zen 5 scales well in cinebench but for gaming it doesn't seem to scale with clock speed which is really strange. It feels to me like something else is hampering zen 5 gaming performance. Other reviewers seem to have made up their mind about it and it's super divisive, but I'm disappointed not many people are digging into it further to why this is the case.
AMD announced a new AGESA update for 9700x, also linux performance seems a lot better compared to windows (which indicates some sort of a potential scheduler inefficiency /bug ) last but not least these CPUs have been tested on what is technically last gen mobos (designed with zen4 in mind) maybe that's the reason for the strange results
Even if this is the case, it's all on AMD if their software/hardware wasn't ready for the CPU launch. All the poor release day results are accurate results and valid for review.
@@AceStrife I suppose there is a small possibility that these issues are due to something in a recent Windows update or something, something that AMD could not have predicted. It is a bit odd that they suddenly decided to postpone the launch, sure there were some rumors about why AMD did this, but if AMD discovered these issues last minute, and were hoping that they could come up with a quick fix, that could be a factor.
CPU performance is better in Linux on every Ryzen generation, so it's not really unexpected to see 9000 series is doing better. It probably should be expected that 9000 series picks up a bit more performance as compilers are adjusted for it, but with the recent very poor core-to-core latency results for Zen5, it's starting to look more and more that Zen5 has some significant architectural regressions and issues. Hopefully this is just another AMD RDNA moment, and Zen6 is going to take what Zen5 laid down and create another RDNA2 moment.
Looking at the SMT results (more power used when off), it really seems like this CPU is being scheduled like the Zen 1 chips. There, because of 4 cores per CCX, there was a change that made it schedule on SMT before moving off the 4C CCX. Perhaps there is some work pending to have that fixed, with possibly some architectural changes also at play. At least on Linux, the 9900X is 25%+ over the 7900X, though the 9950X is not going over 15% consistently. They really need to improve power consumption on these cores, or this is going nowhere.
I totally agree with including the "bad" results as well, since everything was set up by AMD review guide and results were different each run. For memory training taking about 3 mins, did you turn on the context restore for DDR5 memory, sure helps me after initial training with my 7800X3D? For games, I know time is a limiting factor in these reviews, but maybe few more games with some of them more modern FPS games that really leverage more cores and higher clocks. Btw, the custom heat spreader for delidded CPUs looks great. And question for end, what is your over all conclusion about Zen5 so far, especially now that 9900x/9950x have been tested?
Based on the 2 methods it looks like Tony Yu is trying to break the lid off too quick - applying too much pressure too early. Its just a simple case of DeBauer being more experienced.
Roman, are you using the Performance or Balanced power profile? Usually, AMD multi-chiplet (also Intel hybrid chips for the matter) scheduling and power optimizations only apply when the Balanced power profile is enabled. I understand that you reinstalled the OS and the AMD CPU drivers, not just reset the OS. Bear in mind that during the Windows installation, different power and scheduling optimizations are performed if you use multi-chiplet or hybrid CPUs. These optimizations are not reconfigured after the OS installation is finished, even if you reset the OS. Also, the AMD drivers install additional components when a multi-chiplet CPU is put on the system but if you first use a single CCD chip and then switch to a multi-chiplet CPU, the driver won't be reinstalled and the additional components/settings won't be installed/configured. Also, is there any reason why you are using 6000 MT/s DIMMs with the Intel CPU? I understand that if you're looking for the best performance, you can use faster memory and if you're looking to run it without voiding the warranty, Intel currently supports up to 4400 MT/s if I recall correctly -- I know AMD officially supports up to 5600 MT/s even though AMD recommends to use 6000 MT/s to have an Auto:1:1 ratio with the Infinity Fabric Clock (FCLK), unified memory controller clock (UCLK), and memory clock (MEMCLK)...
Did you run the tests with the updated bios for the 7900x3d and 7950x3d that fixed core parking issues? Apparently you have to enable this feature for the 9900 and 9950x cpus
There is some material that 9000x has very high core to core latency even on the same die, can that be the cause of worse performance, and could that be the result of their new increased instruction speed at the cost of latency. Basically anything that needs cross core talking ( most of desktop use ) will run the same or worse than 7000x series.
Fun fact: Tony is the head manager of ASUS China😂and asus is a big customer of thermal grizzly since many rog laptop models use thermal grizzly liquid metal😂
@@coffeejn for the record he never blamed the delidder, nor had the intention to purposely show it on camera, he just didn't know how to use it and the broken CPUs also attracts clicks very well, so🤷I wouldn't be racist just because of some marketing trickery
What a difference to your TG delid tool for the 13th and 14th gen Intel chips. One single move and it was done with my 14900KS. I heated to 70C in an oven to soften the glue.
I know in recent days it's been admitted that it's Windows itself that is affecting performance. Not the AMD drivers. Remember Windows 11 was built from the ground up to support Intel chips. To the extent it was released seemingly far too early to make Intel 11th Gen work.
Can you show the single CPU/s you are actually reviewing as a different colour in the graphs? Makes it so much easier to assess the data instead of constantly hunting for the relevant CPU in a list of 20 all the same colour thanks. Keep up the good work!
I'd wait for the X870e chipset and faster EXPO II memory modules ... and of course for Microsoft to fix their threading issues on Windows 11 OS ("Game mode" shouldn't exist if the OS was developed correctly, band aid).
I just rebuilt a x670e 7950x PC after my 13900k only started bluescreening so I still haven't tried them but I've seen several videos saying you need to disable options in AMD bioses for them to boot in a normal time like secure boot etc..Wendell also talked about that in one of his amd reviews when it came to 9000 weird behavior in benchmarks
Have you done the same tests in Linux? There's a couple of sites reporting that these CPUs performs better in Linux than they do in Windows, at least for now.
Almost completely unrelated, I'd like to see your opinions on Thermalwright 360 clcs. I feel like being half the price of other brand name clcs has to be worth talking about like surely there must be a catch, bit they seem just ad high quality as anything else from the reviews I've seen so far. Like the Notte, and Prism both look top quality. Have we just been getting ripped off before?
I have yet to get a AM5 Delid tool but i definitely need to the intel V1 has been absolutely amazing as it has more then likely saved my 14900K from Certain Death with it's incrwsed surface area for cooling better. Just sucks i bought into intel right before the degradation issues surfaced 🤦♂️☠️
Seems like this cpu was a nightmare to review, I was seriously thinking about getting one as an upgrade over my 7800x3d...ill be waiting for the next x3ds instead.
The die will crack when the user considers the IHS to be sufficiently sheared off and attempts to pry the IHS off when it has not been completely sheared off the die.
If I'm reading that CS2 core utilization correctly, SMT threads are being used while the 2nd CCD is mostly idle. Are you *sure* that's the behavior you want?
Was planning to build a 9800x3d system when it releases now I'm strongly considering just building a 7800x3d rig as it doesn't seem anything from the new gen is gonna be worth the silicon it's printed on...
Windows seems to have issues with these new cpus, Sometimes running as admin improves perf, and other issues. Level1Tech's mentioned these issues in their new vid.
I think we can all appreciate your demo of the delider and feeling required to address it. I think this guy Tony Yu is just wasting your time and while I'm not one for threatening legal action it might be a necessary evil to get him to stfu and stop wasting ppl's time.
What were you expecting you already saw 9700x performance and PBO results? At least AMD could improve memory controller. Or could slightly increase the core count for CCD to 10 to provide multitask performance increase. But why would they do that ? Instead do a favor to Intel and do not increase expectation.
Now I really want to know if the greater power headroom and better binned CCD0 would be worth getting the 9950X for, similar to how the 7950x3D was able to clock significantly higher than the 7800x3D... Would you please try to also measure the 0.1% lows and re-run the gaming benchmarks on Win 10 with the following settings, to see if it results in better/expected frametime stability? - Disable the 2nd CCD, quick boot, virtualization and security settings (such as TPM and secure boot), as well as all power saving "features" (e.g. C-States, Memory and PCIe power down /sleep, etc.) through the BIOS. - In Windows, set the power profile to the highest available setting, disable all power saving settings and disable or suspend all of the background processes that you can. - Also ensure that Windows isn't allowed to automatically update its drivers, before you connect to the internet after installation. - Lastly, only install the required AMD chipset drivers and *not* any drivers/apps/services/profiles that is recommended to install on the 2CCD parts for things like thread scheduling, core parking or application/game detection.
That would not be a review, but a cherrypicking exploration. You wouldn't get any useful data to compare to other CPUs. If you're into extreme tinkering, then the video would be for you, but 99% of the people wouldn't get anything useful out.
AMD should sell an already delided version for enthusiasts. Probably for a bit higher price and better binning like Intel's KS. Some people don't like the thick IHS but don't ready to delid themselves, have risk to destroy chip and lose warranty.
Those numbers are all over the place and they don't seem to correlate with raw computing power (e.g. core count x frequency). I wonder - could this be an artifact of the architectural change to the cache access? *Or* it's a bit f*ck up in the branch prediction rework, which could explain disabling SMT having an impact sometimes as I would imagine the pipeline wouldn't stall in the same way...
So now we have twice as many chips on the shelves as before with the same performance? Demand will shrivel up and prices must come down. I'm still waiting for Arrow Lake.
in the case of Tony Yu, the problem lies between chair and delidder
between the ears of the problem between the chair and delid tool you mean 🤣
there is no problem, he did it on purpose to create hype and get a lot of views
Please stop giving Tony Yu airtime/coverage. He is obviously misusing the delidding device. After being corrected the first time he should know better, so I must assume that he's now intentionally breaking chips for the publicity. Stop giving him that, he's being reckless and people need to start ignoring him.
He's the General Manager of ASUS China. This should be covered until he's fired, at least.
@@aflury you mean ASUS who burned my 12900k z690 asus extreme pc in 2022 because they soldered components backwards ? thanks tony yu
I was your 100th like :P
AMD promised AM5 would be supported for several years, they never said performance was going to improve during that time 😂.
Intel and AMD bamboozling consumers right now
yea what killed the socket is how thick the heat spreader is and the compatibility with am4 coolers... the socket of thread reapers and epyc was just better.
@@alrizo1115 totally :C
That’s why I run Intel.
@@Typhon888 I hope not the latest 2 series of CPUs for your sake though. Both companies have their low moments.
That Tony Yu seems like an idiot trying to trash your delid tool reputation? He obviously does it wrong on purpose, one run then hacks at it with a scraper?
Just assumed he is stupid. Of course could be stupid with malice.
He's actually the head of Asus China branch😂
He's not stupid (breaking it because he doesn't know what he's doing)
Or malicious (trying to ruin reputation)
He's just annoying and being a tiktokker out for attention.. and everyone took the bait.
@@lyteness859 given how bad ASUS has been lately, that really doesn't surprise me at all.
Tony is just an amateur, with lumpy hands and strength. Doesn't matter if he works for Asus.
This industry really needs better naming conventions.
I'm old enough to remember the Intel 9700, Nvidia 9700, and the ATI 9700.
Why? It is just a name on the product stack. Only an idiot reads meaning into the name.
I still have a working ATI sapphire 9800 pro with the zalman orb cooler on it, good old days...
I remember Nvidia 9800GTS, there was a 9700?
Almost like numbers are numbers
ATI 9700 was the best of the mentioned :D
Tony Yu obviously hasn't RTFM and didn't do sufficient side to side moves to completely free the IHS from the die. Never any leverage with a spatula!
Gamer Nexus said AMD told him to use the X3D special chipset driver regarding Ryzen 9 dual CCD CPUs. Just 2 days before his review went live, and he had to throw out a lot of results. Maybe this can have something to say for your results
Same happened to almost everyone, apparently Jay2cents did a step by step install guide (I'm not watching his channel for that) but GN and HB Unboxed did a review using the installation procedure of the 3D Vcache parts and the performance was "normal" Wendell at L1Tech noticed performance issues under Windows and faster performance under Linux 😂
yeah, was curious if der8auer got the same recommendations related to chipset driver and core parking etc (similar to 7950x3D and 7900x3D requirements) as I don't think he mentioned this specifically in the review, but did say he reinstalled windows fresh. But he did analyze that games were using the proper cores with higher freqs.
Yeah the discussion at 11:10 makes me think this might be the issue
After watching the Gamers Nexus review, I had the same thought, as it wasn't mentioned in this review. However, I believe Roman and Steve are somewhat close, so I assumed that when they said they 'reached out to other reviewers,' they would contact each other. I guess we'll find out soon.
@@MrMoerli he might have reached out before Steve said anything and Steve simply didnt have the time to just communicate around what he was told (assuming he was permitted to share it begin with) in time of this video to go out.
18:13 I don't know about new cpus but I remember using liquid metal to clean the solder from the die (9900k), I put a little bit on the die, moved around with a q-tip every 3 minutes, and then polished the die with a tiny amount of polish. Perfect mirror finish in the end. Zero scraping.
You learned that from Roman.
@@thor.halsli Probably, I don't remember exactly.
Scraping is quicker but more dangerous. Just use liquid metal and be patient. 2 or 3 applications, then some Flitz polish to really clean the die 🙂
@@ColinDyckesmore like 10 applications for 13900ks
@@Typhon888 I literally did 3 for about 15 minutes each, but applied quite a lot of LM and agitated it frequently. Same result.
So with Tony, maybe the first time was user error, but now it seems likely he's deliberately doing it for the publicity. Grounds for defamation, not likely, but scummy none the less. I definitely would not give him more attention at this point in content at least. Maybe a strongly worded letter from an attorney though...
0 grounds for defamation. The guy just seems incompetent.
@ivangerginov5648 Trade libel is a thing, but difficult. Once is incompetent, twice seems like it's for the views. Tony is Asus' China manager, so there's also that issue.
@@drewnewby It's his product though. He bought it, he can use it however way he wants.
Not exactly, when a corporate representative purchases a product to use it in a defamatory way, for indirect promotion of their brand, trade libel can come into play.
Techgate when?
Results sorted by 1% lows, amazing!
Yes as it should be. I’ll never understand why in a benchmark that has FPS avg of hundreds.. reviewers don’t typically rank by 1% lows…
@@B1u35ky Me neither, but a lot of outlets don't seem to get it.
Makes the most sense of what will affect your gaming.
All reviewers need too do this going forward.Huge avg FPS that in reality is flip-flopping all over the place is no longer fooling gamers like it did back in the day.We want consistency.
I salute you for seeing something was wrong and refusing to draw conclusions based on obviously iffy data for a deadline.
Must add, as I have seen it many times, the matte or dull look on the dies after delid is remaining indium solder.
If you don't polish it to remove absolutely all indium it will still alloy with LM, leading to a little 'dry out'. Unless you want to apply a really liberal amount of LM I recommend polishing the dies with cape cod or similar to get a perfect mirror on all the chiplets before putting on LM.
It's obvious why his delidding failed, he applied forces perpendicular to the silicon.
ideally you ONLY apply forces perpendicular to the silicon during de-lidding.
Zen5 is having issues. Wendell found that Windows is doing something seemingly wrong. Ancient gameplays found that core parking mostly hurt performance in his stuff
Nice work as always Roman, looking forward to seeing some proper direct die Results.
As for Tony Yu, that clown is simply looking for his 5 minutes of fame and as others have already stated, needs to RTFM! 😒
Romain at 21:43 - for fix very long boot time - try 2 parameters in memory timings settings : 1st - Memory Contex Restore = Enabled, 2nd - Power Down Mode = Enabled
Pretty sure most people are aware of this, especially someone with this background. It can lead to instability issues though
@@lolschrauberinstability only happens when memory context restore is enabled but power down is disabled. As long as you enable power down, there won’t be an issue with stability.
@@PowellCat745Roman would never run with Power Down enabled. Lol.
@@nexxusty power down does affect performance a bit, so I can understand people not wanting to enable it for a fast boot time.
we all know about this man, if the fix was this simple we wouldnt complain
Hope you guys have Arrow Lake delid and direct die kits ready soon. There's going to be tons of people moving into that.
nobody will be doing that as they will be failing and cooking themselves internally.
Nobody is moving to Intel 😂
@@JoeMama-yl1owprobably about the time amd almost doubles the tdp of their chips after the media lauded them for their power efficiency, despite single digit performance gains. How is AMD blowing this so badly? This is an extremely disingenuous launch, when all they had to do was be honest.
@@JoeMama-yl1ow hahaha wrong. cope harder fanboy
Nobody is buying Intel are you kidding me LOL
Thanks to Intel forcing me to learn what exactly a load line is, and how Intel's use of Vdroop is different to Nvidia GPU's and AMDs everything.. I do find myself wondering if on the occasions performance seemed lower for no obvious reason, was there any difference between effective and reported clocks on the "more efficient" part? (If that's even a thing on AMD)
This isn't even a problem with not competing with X3D as some have stated. The 7950X competed with the 5950X, the fact that you don't even see a 5% increase is alarming and points to even X3D being trash no upgrade as well. This isn't just a bad launch it is a failure. All they had to do was deliver similar to what they did in the past and they couldn't even deliver 15%. The stagnation is real, they said Intel is dead what are you gonna buy desides us.
tbh these cpus would have been in development since before the intel saga. These would have been started to get developed a couple years ago when they where doing really well and probably let their quality slip.either way its probably close to an industry first when generational improvements where negligible. we don't even get any new technologies or features to make up for it.
Unfortunately this may be pointing to the hard truth that we are reaching silicon's limit of performance. Dropping in size isn't giving us beter performance anymore and may even be causing problems that can't be solved unless we start making cpus out of graphene.
@@darkkingastos4369not the silicon limit of performance, but the limits of size. You can't make things any smaller. We are already at traces and transistors being 3 atoms wide. You can't go any smaller. Technically you can, but you need to drop frequencies and use a lot more volume of silicon, so it's not that great.
I think it's partly because AMD skipped a learning generation , because Zen 3 was supposed to get a Zen 3+ desktop variant. AMD's memory controller isn't particularly strong to utilise DDR5 having 4 dimms populated, it runs at 3600mhz, which my B550 board can do. I like the thermal behaviour of Zen 5 over Zen 4, but that's only because AMD needlessly changed things from Zen 3. It's a failure of a product Zen 5 built on an expensive poorly selling AM5 platform, I agree.
@@darkkingastos4369 That isn't what happened here, at least not provably. Just shrinking the size alone can give performance but we don't know what changes they made under the hood. This looks like stagnation to me. It is worse than when Intel only targeted 5% IPC gains yearly. This barely hits 3% increase over the 7950X. That is what you call an abject failure. It would have been a fine launch if they had not promised 15-40% increases which were not delivered or priced everything so high, so old parts destroy the new ones. Again a marketing fail still.
If the X3D variant cannot even beat the 7800X3d by 5% we will know Amd is cooked. Not because they failed to deliver on performance, but because they priced it so high when it fails to deliver. They need to fire everyone in marketing/pricing and start from scratch.
Considering gamers Nexus and Jay's two cents both had similar results, I think your results are sound.
This one kind of flopped out of the box
Windows scheduler is the problem. See Wendels review.
there's still issue to be ironed out
@@madb132 And Phoronix
Says a lot they didnt have the competence to figure out there's an issue, unlike L1T and KitGuru among others.
As much as 20% improvements over 7k on linux or with scheduler tweaks in doze, for the same power
@@madb132 Windows is certainly an issue BUT its on AMD to sell a good product. Intel put the effort into ThreadDirector to fix the scheduling situation, while AMD hasnt and it shows.
with a small design change, you could even switch to a one screw design and still move the carriage back and forth
Why all reviewers use PBO without touching Curve Optimizer/Shaper? Makes little to no sense to me. Curve optimizer was all the rage for 7000-series.
This.
Because it's too much work to expect a normal user to do it, a normal user will try and click on PBO but anything more that that only a very small amount of people will even attempt to try.
that's like saying reviewers need to undervolt their CPU to test it properly.
@@terrylaze6247 a normal user wouldn't do anything at all in the bios.
@@MaxIronsThird define "properly". Curve Shaper is a new feature available to 9000-series Ryzens and (almost) no one "tested it properly" (yet?).
I see you also have picked up the planted terrarium hobby! I feel it goes hand in hand with PC building. I also picked up aquascaping of smaller nano tanks for Betta fish and Guppies.
Thank you Roman! I am thinking of getting a 9950X and delid it like my 7950X is. Now I think I will hold off for a bit to see if this gets resolved. I have 3 different versions of your delidding tools so I know they work perfectly.
PBO is very weird. On ryzen 5000 at least ive found best performance leaving EDC near stock but unlimiting PPT and raising TDC. Theres also numerous bugs with it across different bios versions
Tell Yu that you can't pry the ihs off. You'll definitely ruin the chip. That is something he should know. You can't blame the tool on ignorance.
The last major win 11 update caused my 3900X to lose about 5-8% performance. I wonder if that could be causing an issue with the 7950.
Zen 5 has severe Memory Bandwidth/access limits
The bottleneck is Infinity Fabric in I/O die which is the same as for Zen 4
I/O die was even bottlenecking Zen 4, that's why X3D cache is so fast
Steps to do:
1. Re-Architect I/O die Infinity Fabric
2, Switch from obsolete DDR architecture to HBM2
Regarding your second point: Do you like paying 1000$ for 32GB of HBM2? That memory is in high demand in the server market and will be super expensive for the PC DIY buyers. not a good solution.
@@SoWeird4U that was true versus DDR4 but DDR5 is almost as expensive as HBM. And yes people are willing to pay for good performance
I'm not trying to buy my motherboard and ram as a package because the two usually need to be soldered together. And I don't want to pay $2000 for a motherboard and ram. No thanks, I'm fine with DDR5.
Everyone thought we were likely to see the same improvement from 9000 as we did from 7000 over 5000. It seems the big bump from 7000 was the entire AM5 change and now that’s baked in perhaps 9000 hasn’t got that much to give.
I like how you sort by 1 % low I think all reviews should do this
The guy that delidded and broke his own CPU had a grace of a drunk truck driver after 18 beers
That's what happens when you didn't pay for your own CPU, but got it for free
I am also trying to understand that leak benchmark stating 9950X can do 48000+ score in Cinebench R23
I can explain the the benchmark result in CS2. Source 2 engine utilize as many physical cores as possible, as a result you are suffering from cross ccd latency.
To avoid it you should add -threads 8 or 9 in your launch option, this way it works normally. Same for 14900k and all other intel CPUs that have E cores, you should use -threads X, X represents the amount of P cores you have in your CPU. X+1 also works similarly.
I think the unexpected gains from PBO are likely due to the EDC increase, it allows current to spike higher
According to hardware leaker Chi11eddog on X, an upcoming AGESA 1.2.0.1a Patch "A" update, will reportedly increase the TDP of the two Zen 5 chips from 65W to 105W.
EDC on the Ryzen 7 9700X and Ryzen 5 9600X
@@mrhassell unless this was a miscommunication, I am not sure if I like this, upping the specification after the fact is risky
@@TheBackyardChemist Fair enough, and apologies, for my apparent misunderstanding. The patch is making a major difference to power, huge!
Enabling EXPO and specifying PBO +200Mhz, helps a great deal.
(Precision Boost Overdrive)
also there are handfull boards have their CPPC remove during SOC incident year ago...its missing on 50% of amd motherboard vendors ...something AMD failed to mention to their partners that its a must to have for windows 11 core parking software driver support for gamebar mode
9:35 SMT disabling on Cyberpunk and falling BELOW 9700x suggests that cyberpunk has around 16 active threads, the scheduler is spreading them evenly across all 16 cores. And the switching fabric is BROKEN (very very SLOW) like on Tyzen AI HX 370!
Pretty much looks like a scheduler issue. Try disabling a ccd and check again versus the 7700x, 8c vs 8c on single ccd
2CCD delay is somehow worse than prior zen. THere is something wrong with the zen release.
The thing about all the hoops other reviewers had to go through to do core parking in games just shows how bad the core management in windows is. In linux you don't have to do any of that crap and there the 9950x really kicks butt in some scenarios, especially compiling. AMD should have had a chat with Microsoft and let them do all the required updates before shipping these CPUs. Apart from a BIOS update people shouldn't have to do anything for a new CPU.
Love the squeaking sound during delidding! :D
PaulsHardware noted that the mobo bioses were setting the fclock to 2100 instead of 2000 and that was causing some problems for him
AM4EVER! I probably will not upgrade till AM6 comes out given how good the x3d chips are. 5700x3d for the budget win baby!
Since you mentioned slow boot time I can’t help myself but to ask here, I have had a problem with the boot time on my 7900X since I purchased it new… I have an MSI MEG ACE X670E and my memory is 6400 CL 32 TridentZ . My problem is there regardless of with or without XMP enabled. So when it’s on 4800MHrz it would take around 30 seconds to post and around 41 seconds each time until Windows boots.
And when it’s on 6400 XMP it takes around 40 seconds to post and 52 second until windows startup .
Now I’ve just played with the timings and put it on 6000 CL 30. I’ve enabled “memory context restore” which followed that I had to enable “power down enable” as well. Now the mobo posts around 25 seconds and OS starts at 35. At this stage I cannot put the pc on sleep mode cause if I do the mobo won’t post.
If I hadn’t enabled “memory context restore”, any change from 4800Mhz to either XMP or 6000Mhrz would also render restarting the PC impossible cause again it wouldn’t post.
I changed the motherboard to an Asus X670E-E, boot time still didn’t change, changed the memory kit to one exactly like it, didn’t change, changed the M.2 drive from which the OS loads, didn’t change (obviously cause the problem is the time for it to post, not the OS). The only thing I haven’t changed is the CPU so that’s what I’m left with.
I really can’t understand it but guess the problem is with the CPU
Having to communicate across CCD, or migrate threads from one CCD to another, to compute a game frame could be the same old reason why dual-CCD parts lose in some games vs single CCD parts.
The latency with any given app running across multiple CCDs is usually worse than the same app running strictly within one CCD, so for extremely low-turnaround-time ("low-latency") tasks like generating a game frame in just one 700th of a second (~1.42ms), +/- other random sources of system noise, every bit of latency matters there.
Maybe a UEFI/firmware/OS update or two can help with the scheduling, I'm not sure.
Yeah zen 5 scales well in cinebench but for gaming it doesn't seem to scale with clock speed which is really strange. It feels to me like something else is hampering zen 5 gaming performance. Other reviewers seem to have made up their mind about it and it's super divisive, but I'm disappointed not many people are digging into it further to why this is the case.
Really looking forward to the overclocking reviews . We will also get agesa updates .
AMD announced a new AGESA update for 9700x, also linux performance seems a lot better compared to windows (which indicates some sort of a potential scheduler inefficiency /bug ) last but not least these CPUs have been tested on what is technically last gen mobos (designed with zen4 in mind) maybe that's the reason for the strange results
Chi11eddog on X, AGESA 1.2.0.1a Patch A update - 65W to 105W. - 9600X + 9700X -
Even if this is the case, it's all on AMD if their software/hardware wasn't ready for the CPU launch.
All the poor release day results are accurate results and valid for review.
@@AceStrife well it doesnt make sense to look whos fault is what matter if they fix it or not in the near future lol
@@AceStrife I suppose there is a small possibility that these issues are due to something in a recent Windows update or something, something that AMD could not have predicted.
It is a bit odd that they suddenly decided to postpone the launch, sure there were some rumors about why AMD did this, but if AMD discovered these issues last minute, and were hoping that they could come up with a quick fix, that could be a factor.
CPU performance is better in Linux on every Ryzen generation, so it's not really unexpected to see 9000 series is doing better. It probably should be expected that 9000 series picks up a bit more performance as compilers are adjusted for it, but with the recent very poor core-to-core latency results for Zen5, it's starting to look more and more that Zen5 has some significant architectural regressions and issues.
Hopefully this is just another AMD RDNA moment, and Zen6 is going to take what Zen5 laid down and create another RDNA2 moment.
Looking at the SMT results (more power used when off), it really seems like this CPU is being scheduled like the Zen 1 chips. There, because of 4 cores per CCX, there was a change that made it schedule on SMT before moving off the 4C CCX.
Perhaps there is some work pending to have that fixed, with possibly some architectural changes also at play.
At least on Linux, the 9900X is 25%+ over the 7900X, though the 9950X is not going over 15% consistently. They really need to improve power consumption on these cores, or this is going nowhere.
That 7800x3d is the goat of them all. What a good chip.
I totally agree with including the "bad" results as well, since everything was set up by AMD review guide and results were different each run. For memory training taking about 3 mins, did you turn on the context restore for DDR5 memory, sure helps me after initial training with my 7800X3D? For games, I know time is a limiting factor in these reviews, but maybe few more games with some of them more modern FPS games that really leverage more cores and higher clocks. Btw, the custom heat spreader for delidded CPUs looks great. And question for end, what is your over all conclusion about Zen5 so far, especially now that 9900x/9950x have been tested?
Based on the 2 methods it looks like Tony Yu is trying to break the lid off too quick - applying too much pressure too early. Its just a simple case of DeBauer being more experienced.
Roman, are you using the Performance or Balanced power profile? Usually, AMD multi-chiplet (also Intel hybrid chips for the matter) scheduling and power optimizations only apply when the Balanced power profile is enabled.
I understand that you reinstalled the OS and the AMD CPU drivers, not just reset the OS. Bear in mind that during the Windows installation, different power and scheduling optimizations are performed if you use multi-chiplet or hybrid CPUs. These optimizations are not reconfigured after the OS installation is finished, even if you reset the OS. Also, the AMD drivers install additional components when a multi-chiplet CPU is put on the system but if you first use a single CCD chip and then switch to a multi-chiplet CPU, the driver won't be reinstalled and the additional components/settings won't be installed/configured.
Also, is there any reason why you are using 6000 MT/s DIMMs with the Intel CPU? I understand that if you're looking for the best performance, you can use faster memory and if you're looking to run it without voiding the warranty, Intel currently supports up to 4400 MT/s if I recall correctly -- I know AMD officially supports up to 5600 MT/s even though AMD recommends to use 6000 MT/s to have an Auto:1:1 ratio with the Infinity Fabric Clock (FCLK), unified memory controller clock (UCLK), and memory clock (MEMCLK)...
Delidding ASMR is kind of awesome, really. Thumbs up!
On benchmarks it was very important to Oc the 9000 series, 6 ghz was 3700pts compared to 3400 5ghz iirc
False voltage and power readings?
Did you run the tests with the updated bios for the 7900x3d and 7950x3d that fixed core parking issues? Apparently you have to enable this feature for the 9900 and 9950x cpus
There is some material that 9000x has very high core to core latency even on the same die, can that be the cause of worse performance, and could that be the result of their new increased instruction speed at the cost of latency. Basically anything that needs cross core talking ( most of desktop use ) will run the same or worse than 7000x series.
gotta setup core parking, Steve went through the steps to make sure it was working.
Fun fact: Tony is the head manager of ASUS China😂and asus is a big customer of thermal grizzly since many rog laptop models use thermal grizzly liquid metal😂
So he's running on ASUS settings? Maybe they should reset him to defaults.
So ASUS is trying to tarnish someone else reputation. Typical Chinese scam artists.
@@coffeejn for the record he never blamed the delidder, nor had the intention to purposely show it on camera, he just didn't know how to use it and the broken CPUs also attracts clicks very well, so🤷I wouldn't be racist just because of some marketing trickery
What a difference to your TG delid tool for the 13th and 14th gen Intel chips. One single move and it was done with my 14900KS. I heated to 70C in an oven to soften the glue.
I just did my 13900ks also. No heat one pass about 2min lol
I know in recent days it's been admitted that it's Windows itself that is affecting performance. Not the AMD drivers.
Remember Windows 11 was built from the ground up to support Intel chips. To the extent it was released seemingly far too early to make Intel 11th Gen work.
Can you show the single CPU/s you are actually reviewing as a different colour in the graphs? Makes it so much easier to assess the data instead of constantly hunting for the relevant CPU in a list of 20 all the same colour thanks. Keep up the good work!
Transistor downscaling means noise and aging at high clocks
I'd wait for the X870e chipset and faster EXPO II memory modules ... and of course for Microsoft to fix their threading issues on Windows 11 OS ("Game mode" shouldn't exist if the OS was developed correctly, band aid).
I just rebuilt a x670e 7950x PC after my 13900k only started bluescreening so I still haven't tried them but I've seen several videos saying you need to disable options in AMD bioses for them to boot in a normal time like secure boot etc..Wendell also talked about that in one of his amd reviews when it came to 9000 weird behavior in benchmarks
Have you done the same tests in Linux? There's a couple of sites reporting that these CPUs performs better in Linux than they do in Windows, at least for now.
Almost completely unrelated, I'd like to see your opinions on Thermalwright 360 clcs.
I feel like being half the price of other brand name clcs has to be worth talking about like surely there must be a catch, bit they seem just ad high quality as anything else from the reviews I've seen so far. Like the Notte, and Prism both look top quality.
Have we just been getting ripped off before?
did you make sure to set your bios to chose driver for the core parking feature?
I have yet to get a AM5 Delid tool but i definitely need to the intel V1 has been absolutely amazing as it has more then likely saved my 14900K from Certain Death with it's incrwsed surface area for cooling better. Just sucks i bought into intel right before the degradation issues surfaced 🤦♂️☠️
Seems like this cpu was a nightmare to review, I was seriously thinking about getting one as an upgrade over my 7800x3d...ill be waiting for the next x3ds instead.
The die will crack when the user considers the IHS to be sufficiently sheared off and attempts to pry the IHS off when it has not been completely sheared off the die.
CCD problems on windows and perhaps some extra windows quirks behind some of the strange results.
If I'm reading that CS2 core utilization correctly, SMT threads are being used while the 2nd CCD is mostly idle. Are you *sure* that's the behavior you want?
... only thing i'm wondering and waiting for now is a "9800X3D" ... due to how good the 7800X3D was (and still is 😅)
Is there something wrong with the beta bios? I also use the x670e hero. The latest bios is 2204/2302. The latest chipset driver is 6.07. Do they help?
12 mins in you say you're running the same BIOS as 9600x?
@der8auer Unrefined zen 5 code, but more importantly, we need deeper dive on core parking & inter-core latency. A method to disable core-parking
Was planning to build a 9800x3d system when it releases now I'm strongly considering just building a 7800x3d rig as it doesn't seem anything from the new gen is gonna be worth the silicon it's printed on...
add a little bit of lithium grease to those threads to get rid of the squeak
AMD's Rocket Lake moment.
Windows seems to have issues with these new cpus, Sometimes running as admin improves perf, and other issues. Level1Tech's mentioned these issues in their new vid.
Did you hear about this special setup you're supposed to use?
"One 7800X3D to rule them all"
I have a simple question. Why do testers seem to always use Cinebench R23 instead of the newest one (24)?
Is the thermal grizzly phasesheet ptm a rebranded honeywell ptm7950?
Really hope this PBO stuff won't cause another massive issue in the CPU market like what happened with Intel.
I wondered what anomalies will show up after the delay of launch
I think we can all appreciate your demo of the delider and feeling required to address it. I think this guy Tony Yu is just wasting your time and while I'm not one for threatening legal action it might be a necessary evil to get him to stfu and stop wasting ppl's time.
19:50 You know.. what we like to call, the right way?
Damn, that guy can't Delid anything to save his life. He for sure used a thermal paste spreader to pry it open thus causing the break.
did you run cs2 with -threads 16 on commanline? (I guess no) Default is 3 btw.
6:20 hpad? what?!? edit: ohh you mean HPET
Tony is a clown
Who's Tony? who cares?
😅
@@madb132 did you watch the video ? tony is the clown that can't delid cpu properly
@@ThaexakaMavro he's saying Tony is irrelevant...
Tony should work at a steel demolishing place. All this fine work is not for him.
@@twiggsherman3641So irrelevant that Derbauer had to talk about him
Have you tried testing it with Windows 10? I heard weird things about Windows 11 on these CPUs.
13:15 That's 1337!
7800X3D 💪
What were you expecting you already saw 9700x performance and PBO results? At least AMD could improve memory controller. Or could slightly increase the core count for CCD to 10 to provide multitask performance increase. But why would they do that ? Instead do a favor to Intel and do not increase expectation.
Now I really want to know if the greater power headroom and better binned CCD0 would be worth getting the 9950X for, similar to how the 7950x3D was able to clock significantly higher than the 7800x3D...
Would you please try to also measure the 0.1% lows and re-run the gaming benchmarks on Win 10 with the following settings, to see if it results in better/expected frametime stability?
- Disable the 2nd CCD, quick boot, virtualization and security settings (such as TPM and secure boot), as well as all power saving "features" (e.g. C-States, Memory and PCIe power down /sleep, etc.) through the BIOS.
- In Windows, set the power profile to the highest available setting, disable all power saving settings and disable or suspend all of the background processes that you can.
- Also ensure that Windows isn't allowed to automatically update its drivers, before you connect to the internet after installation.
- Lastly, only install the required AMD chipset drivers and *not* any drivers/apps/services/profiles that is recommended to install on the 2CCD parts for things like thread scheduling, core parking or application/game detection.
That would not be a review, but a cherrypicking exploration. You wouldn't get any useful data to compare to other CPUs. If you're into extreme tinkering, then the video would be for you, but 99% of the people wouldn't get anything useful out.
AMD should sell an already delided version for enthusiasts. Probably for a bit higher price and better binning like Intel's KS. Some people don't like the thick IHS but don't ready to delid themselves, have risk to destroy chip and lose warranty.
lol - I find it amusing that anyone even thinks there is any value in this. I mean... just look at those tiny little things man! haha - wow
Those numbers are all over the place and they don't seem to correlate with raw computing power (e.g. core count x frequency). I wonder - could this be an artifact of the architectural change to the cache access? *Or* it's a bit f*ck up in the branch prediction rework, which could explain disabling SMT having an impact sometimes as I would imagine the pipeline wouldn't stall in the same way...
So now we have twice as many chips on the shelves as before with the same performance? Demand will shrivel up and prices must come down. I'm still waiting for Arrow Lake.