Mr. Fox

One thing that I find difficult with overclocking Ryzen versus Intel is how fussy and sensitive they are in comparison. There are lots of settings that do not play nice together and if one thing is just slightly off it is like a house of cards. Performance and stability can go from great to garbage if all of the stars are not aligned. The CPU overclocking on Ryzen is about as fickle and finicky as RAM tuning is on Intel, and RAM tuning is a tedious process on either platform. In a way it is fun, but can be the opposite of fun (a really huge pain in the butt) when finding the right lever to pull is elusive. When you factor in the wildcards of the silicon lottery (trashy CPU and memory quality samples are common) it becomes even harder to be pleased with results. The crazy memory overclocking results posted places like overclock.net are fringe examples of the small winner count in the silicon lottery and they are not representative of the norm. Over the past week or so I have moved around CPUs and memory between the Strix and Apex. The Strix CONSISTENTLY has higher Cinebench scores compared to the Apex regardless of the CPU sample (9950X SP118, 9950X SP119 and 4585PX SP120). The memory overclock limits are exactly the same on all three CPUs no matter which motherboard they are installed in. Nothing I have tested in terms of settings allows the Apex to beat the Strix X870E-E on Cinebench. EVER. This was equally true comparing Cinebench scores again the X870E AORUS Master. The Apex loses 100% of the time in Cinebench. Why? I wish I knew. Is the the idiotic PCIe lane allocation on the Apex? The 2-DIMM memory topology ASUS used on the Apex? Is it the DIMM.2 somehow dragging down the CPU? Don't have a clue. Whatever it is, Cinebench doesn't like it. As far as memory overclocking in concerned, either the Apex is hardly any better or both of my memory samples have reached their functionality threshold. Same threshold with four different CPUs. The Apex maxes out at 8200 stable with my G.SKILL 6000 C26 1.450V 32GB kit and the Strix maxes out at 8000 stable with that kit. The Apex maxes out at 8400 with sloppy timings that gives performance the same as 8000 with tight timings using the Kingbank 8400 C40 1.450V 48GB kit (which has both XMP and EXPO 8400 profiles). With that kit the Strix maxes out a 8000 with the same tight timings that work on the Apex. The Strix cannot boot 8400 stable but the end result is the same memory performance on both systems. The latency and speeds follow the CPU and RAM kit to achieve identical results in each motherboard. This essentially makes the Strix the winner overall. The fact that it is roughly $300 cheaper, depending on where and when you buy it, makes it an absolute no brainer. The Apex is a great motherboard, but it doesn't deliver based on price or the fact that Cinebench scores are CONSISTENTLY 500-1000 points lower for any given CPU installed in the Apex. I captured this comparison of the small latency penalty of 3D V-Cache. I did not bother capturing screenshots of different motherboard, RAM and CPU combos because it always ends the same. The screenshots would not show anything different overall. The latency results follow the X3D and non-X3D CPU and memory combo regardless of what motherboard they are installed in. The G.SKILL 8000 C40 kit that I picked up with the Supercool Direct Touch RAM block is an XMP kit. It is virtually impossible to tune it stable at 8000 with all manual tuning in either of the AM5 motherboards. If I use timings that are stable with either of my other two memory kits mentioned above it refuses to boot and I have to clear CMOS. I cannot figure out why that is the case and it does not make any sense. Maybe the kit is rubbish, but I don't have an Intel motheboard to test it in to find out. However, I have experienced the exact same problem trying to use XMP memory that I had that were stable on Z790 motherboards that were also unusable on X670E and X870E motherboards, so I think it is something peculiar with the XMP memory, even though it makes no sense that manual tuning doesn't work. I have been trying to get the memory latency on the 4585PX below 60ns and it is tough to do. This CPU is nothing more than a better binned 9950X3D and I think this one might be slightly better than average. While the memory and L1, L2 and L3 read/write/copy and latency values are a direct result of clock speed, I wondered if the 3D V-Cache was making the latency higher. It seems it is, but not by enough to be statistically relevant. I asked Google Gemini about it, and this is what the AI regurgitated for me. The "chef" analogy (below) is pretty good. I thought @Raiderman and @jaybee83 might find this particularly interesting. Per Google Gemini - for whatever that is worth... Question: Do AMD X3D CPUs have higher memory and L3 cache latency than the non-X3D processors? Yes, technically they do. It is a paradox of the architecture: AMD X3D CPUs generally demonstrate higher latency for both L3 cache (in cycles) and system memory (in nanoseconds) compared to their non-X3D counterparts. However, they are faster in gaming because their massive cache capacity prevents the CPU from needing to access system memory in the first place. Here is the technical breakdown of why this happens and why it doesn’t hurt gaming performance. 1. L3 Cache Latency: The "Stacking Tax" The primary difference between an X3D chip (like the 7800X3D) and a non-X3D chip (like the 7700X) is the 3D V-Cache technology. AMD physically stacks a slice of SRAM on top of the processor die. This introduces a slight delay. The Penalty: Accessing the 3D-stacked cache typically adds a ~4-cycle penalty. The Clock Speed Factor: Latency is a product of cycles and clock speed ($Time = Cycles / Frequency$). Since X3D chips run at lower clock frequencies to manage heat, the absolute latency in nanoseconds is slightly higher than non-X3D chips that boost 400-500MHz higher. Feature Non-X3D (e.g., 7700X) X3D (e.g., 7800X3D) L3 Cache Cycles Lower (Faster) Higher (~4 cycle penalty) Core Frequency Higher (~5.4+ GHz) Lower (~5.0 GHz) Result Lowest raw L3 latency Slightly higher L3 latency 2. Memory Latency: The Benchmark Illusion If you run a synthetic benchmark like AIDA64, an X3D CPU will almost always show higher memory latency (worse performance) than a non-X3D chip. Reason 1: Clock Speed: Memory latency is sensitive to core frequency. Because X3D chips have lower peak clocks, the time it takes to request data from RAM and process the return is technically longer. Reason 2: Testing Methodology: Synthetic benchmarks deliberately force the CPU to fetch data from System RAM (DDR5) to test the worst-case scenario. They bypass the cache to measure the "wire delay" to the memory sticks. In this specific "worst-case" scenario, the non-X3D chip wins because it has faster clock speeds and no cache-stacking overhead. 3. The "Effective Latency" Paradox If the latency is higher, why are X3D chips faster for gaming? The answer lies in Hit Rate. Non-X3D: Fast access to L3, but the cache is small (32MB). The CPU frequently "misses" (can't find data) and must go to System RAM. Penalty: Trip to RAM takes ~60–70ns (eternity in CPU time). X3D: Slightly slower access to L3, but the cache is huge (96MB).7 The CPU almost never misses. Benefit: It stays within the L3 cache, which takes ~10–12ns. The Math of "Effective" Latency: Even if the X3D cache is 10% slower to access, avoiding one trip to system RAM (which is 600% slower) saves massive amounts of time. Analogy: Imagine you are a chef. Non-X3D: You have a small fridge under your counter. You can grab ingredients instantly, but you frequently have to run to the grocery store (RAM) for missing items. X3D: You have a giant walk-in pantry. It takes you 2 seconds longer to walk into the pantry than the small fridge, but you almost never have to leave the building to go to the grocery store. You cook faster overall. Summary Raw Latency: Higher on X3D. (Due to lower clocks + stacking penalty). Effective Latency: Lower on X3D. (Because you rarely hit the slow System RAM). If your workload fits entirely inside the cache (like most games), X3D wins. If your workload is random and too large for even the X3D cache (like some scientific rendering), the non-X3D chip might win due to higher clock speeds. The video below is relevant because it directly compares the older 7800X3D and 7700X, illustrating how the theoretical latency differences translate into actual gaming performance. Ryzen 7 7700X vs Ryzen 7 7800X3D - Any Difference Or Not? TL;DR - See below: Lesson #1: If you want to go with AM5 and care about Cinebench scores buy an X870E-E Strix or X870E AORUS Master and save yourself $200-$300. If you a lottery winner have an extremely rare Ryzen golden silicon CPU sample and an extraordinarily superior memory kit and only care about synthetic memory benchmarks and nothing else matters, cough up the money for an Apex then have fun watching paint dry overnight with Karhu speed tests that have no measurable bearing on normal, real-world performance results in anything other that Karhu. Lesson #2: Unless you want to waste a lot of time and drive yourself insane, don't try to get an Intel XMP memory kit to work on AM5. Unless you get lucky somehow, you are doing to hate yourself for not just buying an EXPO memory kit.

There should be no yellow visible even when it is still in the PSU box.

Ooops, I sure did. I updated the post with this image. Pretty huge difference in high and low FPS, and those runs are 4K with no frame generation. CCD0 only (3DvCache only) - I used Process Lasso for this instead of Windoze trash

Awesome. I am looking forward to seeing how it runs, brother. I got up early to delid the CPU. It always amazes me how much better than makes things. I increased the core clock 100MHz across all 16 cores and my max temperature is still over 20°C cooler. I hate owning CPUs that are not delidded. The experience sucks with the IHS. So, I gave up on trying to use Windoze worthless trash software bloat to determine whether or not 3DvCache actually worked or not. Micro$lop butchers or botches up everything they get their hands on. I used Process Lasso to set parameters on the EXE file to force 3DvCache CCD0 use or all-core use with both CCDs. As you can see below, it DOES make a difference. All Core (no 3DvCache) CCD0 only (3DvCache only) - I used Process Lasso for this instead of Windoze trash All Core (no 3DvCache) CCD0 only (3DvCache only) - I used Process Lasso for this instead of Windoze trash

Have you identified anyone that owns a Zotac 5090 AIO version? I haven't gone looking but I don't recall seeing anyone in the forum at oc.net that has one. In general, I do not want an AIO GPU, (or CPU AIO,) but definitely think it is better than air. I do not care for the appearance of most of the AIO cooled GPUs. I think most of them are pretty garish-looking and I do not like the short and stubby look. At least the water blocked GPUs are not nearly as short and stubby as the OEM AIO models.

When you know most of the users of your product are painted into a corner and need to use your product no matter how lousy it becomes, that makes cleaning up your messes and keeping users happy a very low priority. Switching to Mac OS or Linux is not an option for 99% of Windoze users and they know it. Most either cannot make such a switch due to lack of equivalent software (which could make it virtually impossible from a business continuity perspective) or lack of technical knowledge on the part of the user base that would totally incapacitate the users (consumer and business).

The normie sheeple and ignoramus population has become so vast and prevalent the OEMs no longer give a rat's butt about enthusiast money because they don't need to. We no longer matter. There are no customers, only victims. Clueless morons pay asinine prices for things they know nothing about, and they do it with a smile. They believe all the ads they read, all the lies their "friends" on Facepoot and Reddtit tell them, and all of the disinformation they watch on the "news" so is it any wonder that clowns like A$$zeus behave as if they have a license for sodomy? We have both to blame: the stupid sheeple and the shifty thugs peddling grossly overpriced, misrepresented and overhyped crap.

Works fine on the Zotac Solid OC. So I am guessing with the shunt resistors I've got like 1600W (which I am pretty sure it cannot reach).

That Matrix GPU is only good so others can use the vBIOS. The price is extremely stupid and I think it looks stupid, too. 🤣 I have gone pee in my own backyard before. When you have a family of 7 and 3 bathrooms, there are times when you just can't hold it any longer. Now that we are down to 2 people at home (empty nesters) and 2 bathrooms I have not found myself needing to do that for a long time.

You hate cold and I hate heat. My body has adjusted to the hot weather here, but too hot is a health risk that is equally hazardous as freezing to death. Extreme heat and extreme cold are both an issue. Temperature aside, I love living in the desert because I do not like precipitation in any form. I hate rain and I hate snow. I like it to be very dry, clear skies and low/no humidity whether it is hot or cold. I find it easier to get warm if I am too cold. I can bundle up and add more blankets to the bed. If I am too hot it is harder to fix that problem. You can be wet and naked and still be too hot, LOL. Both can kill you.

I delid my own with the der8auer tool. Works fantastic. You just have to spend about 30 minutes moving it back and forth with the tool until it is loose. Tedious but foolproof as long as you keep going until it falls off and don't try to lift the IHS before it is loose. Usually in the range of 500-1000 Cinebench R23 points lower with the Apex. AMD CPUs are inconsistent between Cinebench runs and can have a wide variance between runs with the same CPU. There are too many artificial algorithms in play with Ryzen. But, I can't get within 500 to 1000 points of my best scores with the Strix or AORUS Master using the Apex with the same CPU installed. It almost has to be something in the firmware. It would not seem logical for something with the motherboard hardware to affect it. Also worth mentioning that it is with the same Windows installation. To rule out an OS issue I create an image from one, restore it using Macrium Reflect on the other and reactivate. So all of the same software, tweaks, drivers and OS tuning on both machines. This makes no sense to me. There is probably something I need to do different that I don't know about. Maybe an obscure BIOS setting I need to turn on or off on the Apex. This should be an almost entirely mathematical outcome based on core count, IPC and clock speed as long as something is not interfering with the Cinebench run. This EPYC CPU seems like a really good sample and I am pretty sure I am going to keep it. It seems better than average looking at the V/F curve and voltage requirements in practice. Getting it to function like an X3D in a scenario where it should has not worked for me yet. When I try to configure it to use the 3DvCache CCD it is defaulting to the non v-cache CCD no matter what I have tried so far. Configuring it to function like an ordinary CPU it seems fine, so I suspect this is a firmware configuration problem, an OS/game/chipset driver bug issue, or even an ignorant noob issue on my part (relating to firmware configuration). I did a clean OS install with the latest chipset drivers and set everything that is supposed to be how it is done that I know of and it still parks the wrong CCD and activates the one without v-cache, LOL. I don't care about the v-cache necessarily, but if I can use it as leverage for 3DMark scores I certainly want to take advantage of it. I don't expect it to matter much at all with something like Cinebench, wPrime, Pifast, or Y-Cruncher.

One thing I find very puzzling about the Apex is that it does not matter what CPU I have installed, my Cinebench scores are always notably lower than the Strix. Every CPU that I test consistently has stronger Cinebench results when I install it in the Strix. This is my third Apex. I don't really understand why this is the case, but it is consistent. Same was true of the AORUS Master... consistently higher Cinebench scores than the Apex using any given CPU. This really bugs the crap out of me and makes me wish I had saved $300 and just purchased another Strix X870E-E or just kept the AORUS Master. I think I am spinning my wheels with the EPYC CPU. It seems like a much better sample overall, but I am not going to be able to really confirm that until it is delidded and bare die. Absolutely ludicrous thermals with the factory solder and IHS. It's unfortunate that most CPUs suck with the factory solder and IHS. It's true of both Intel and AMD, but seems much more so with AMD for some reason. They seem a lot more sensitive to high temperatures. Jury is still out whether I like 3DvCache. The 9950X3D that I RMA'd was such a pathetic silicon sample that it wasn't fair to draw any conclusions based on what a piece of garbage it was. I'm leaning toward either not liking it or viewing it as an irrelevant selling point/gimmick. (I'm a casual gamer, not a gaming enthusiast, and it seems to provide no tangible benefit for an overclocking enthusiasts.) I do see higher FPS in the few games I tested. 3DMark doesn't seem to benefit.

I still have a lot to figure out for the best settings for tuning of this new EPYC CPU but it does appear to be better than either of my good but average 9950X samples. Initial impressions are that it is a pretty decent sample. I think my memory kit could be a little better, but I was surprised it could run 6400 1:1 and FCLK 2200. I haven't spent much time with core tuning yet, just set a simple PBO, but will definitely need a good manual overclock to see any respectable Cinebench results. I will play with it more tomorrow to confirm it is worth keeping before I delid it. I always forget how crazy hot these CPUs are that have not been decapitated and run bare die. There are still many hours that need to be invested to find what works best. Requires a lot more time and effort with AMD to get things figured out.

Hopefully this was not a waste of sand. But, I will find out as soon as I button things up. If it's not better than what I already have it's going back for a refund.

I've been watching the insane increase in DDR5 prices. I don't need any, but decided to see if I could find some spares cheap to hang onto. Just scored a Trident Z5 DDR5-8000 48GB CL40 1.400V kit with Supercool direct touch cooler on eBay for $250. That same RAM kit is now about $300 sold and shipped by NewEgg. So, basically like getting the Supercool cooler kit for free, give or take a couple of bucks.