If you are playing games on your threadripper, you (should) already be aware of this and have taken steps to address it via something like lasso. But tbh you don't buy a threadripper to primarily play games, are you?
Definitely a “because we can” moment. I don’t think anyone was seriously waiting for something like this to upgrade, but it puts them that little bit further ahead at the top of the charts.
I just wish AMD had released it far, far sooner. My main use-case for X3D was to be Lightroom, for which it can help tasks more constrained by core count and memory t/put (which includes exports) immensely. In the time AMD have taken to get this part anywhere near market though, Adobe have pulled their finger out and given us GPU-accelerated exports (along with other things), which have served to make the X3D2 part a little redundant.
Maybe there's a serious point there. AMD seems a bit too content with keeping their Ryzen chips just one or two points ahead of Intel's Core chips. AMD's rival isn't Intel's hardware though, it's every vendor's software, since that's what determines what hardware gets purchased. I'm not paying $1000 for a CPU upgrade for a 10% performance uplift now that I can spend $1000 on a GPU upgrade and get a 50% uplift. AMD really needs to court those software devs with their features and performance, not piss about delaying parts because they don't feel the need to push forward when they're already ahead.
Depends on the price. My 9950X3D is blazing fucking fast, paired with a pcie5 SDD, I see zero reason to need more CPU right now.
Nor could I afford one lmao.
But we could a had the dual instead!
P.s. mine runs rather hot actually, hitting the 90's at times even.. Might have to disassemble my loop again.. and get a damn filter xD
I had to dial downy OC so mine stopped hitting 90c, and air also got mine on a water loop. They just run hot AF when you're pushing 250+ watts through them.
Madness!
I max out my fans at 40C coolant temp :P
They're set to run at 50% during heavy load and 25% when not :)
Heavy load = 40C > coolant temp > 34C
But it's sharing the loop with a 7900XTX :)
I tested it out last night and never got higher than 73 on full load.
Thought about doing something similar but the temps were too jumpy so the fans are annoying. Instead I use a rolling average of sorts... fans will respond fast to rising temps, but take 20s of the average temp before they start ramping down.
Ya, I've set it so the fans ramp up to high load once coolant hits 34C, but don't ramp down again until coolant reaches 30C - works great :)
Thing is, both my GPU and CPU blocks are clogged due to some bad anti-vortex foam from EK that disintegrated into flakes, shite is stuck in every damn nook and cranny in the entire loop now, every time I clean it, more just keeps getting stuck in the blocks :P
I'm gonna get an on-line filter once I get some extra money
These cpus also undervolt really well and even when power limited you can still get ridiculous performance. I’m hitting 5.4-5.5ghz ish in cpu heavy games easily with a 125w power limit and a heavy undervolt, zero difference in performance compared to running it without limits.
Amazing hardware.
I would be surprised if apple silicon doesnt wipe the floor with even amd zen5 x3d when it comes to lightroom. Good if pc can stay ahead with the addition of gpu acceleration. I just have to imagine that one would be hard pressed to have a better setup for that than something like an m5 max macbook
People don't understand market segmentation. I mean, we all *want* workstation features at a consumer price, but not everybody gets that that's not gonna happen because no company is going to cut into its profits just for good will, especially with so little competition in the market.
I understand market segmentation, thanks.
It’s annoying when 16 lanes are gobbled up by the GPU and then another 4 for NVME. So you’re left with barely any full speed expansion unless you jump to a higher segment which is MUCH more expensive. These halo consumer products should split the difference a little bit.
yeah I would 100% not mind to pay similar enough prices for a 9800X3D but on a more expensive platform with something like 32 lanes of PCIE
16 for the GPU, then 16 more for all the add on and multiple full BW PCIE stuff, esp if they can do the split like 3 x4 gen 5s then 4 more x4 gen 4s (IE 16 gen 5 lanes) for 16 PCIE 5 lanes
swap those out as needed for other stuff like USBC / TB / networking and I think that for an actual consumer HEDT and not just workstation would be killer.
focused on OCing, lots of nvme, on top of real good everything else, hell give us intel wifi chip for that actual premium experience
Sounds like Arrow Lake Refresh pushed them to respond, I’d heard this was cancelled because they thought they didn’t need it. Still not gonna be a good look when their new flagship is barely an upgrade over their old one and their competitor offers a comparable product in multi core for less than half the price.
The 270KP.
If you don’t really care about gaming the difference in productivity is negligible in comparison to the price chasm between them.
If you DO care about gaming but still want good productivity, the 270KP is still fantastic performance for its price. If price isn’t an issue, just get the old 9950X3D, because the 2nd CCD having vcache makes no difference for gaming.
Of course it is comparable. According to hwunboxed reviews (which includes extreme outlier cases like BG3) the difference in 1% lows between the 270k and the 9800x 3d is 130 vs 145. That's at 1080p. with a 5090. With 3 huge outlier games. That's really comparable. It's actually the same tier of gaming performance.
How is BG3 an outlier when it is getting more players than TLOU2, HZD, Spider Man 2, Mafia, Space Marine 2 which were also tested?
If anything these CPU gaming reviews should include games like Path of Exile (1 or 2), Civ 6 or 7 turn time, One of the paradox grand strategy titles, probably HOI4 given GN test Stellaris as a tic rate test. Same for factory builders and city builders and so on.
The fact the vast majority of reviewers mainly stick to the standard suite of AAA titles is a bit of a nonsense given how broad gaming actually is. The fact HUB and some other places do include titles like BG3 or ACC or Stellaris is a really good thing to give a broader perspective of performance. I just wish there was even more variance in the types of titles that get tested to provide an even more complete view.
An outlier is a game that it's performance is too extreme compared to the average. BG3 is a game that the 9800x 3d has double the performance of the next intel chip. That is not the norm. How many players it has is completely irrelevant.
This whole broader perspective on performance is not what Hub is doing, hub is picking games that perform well on the 3d cache and uses them. There are plenty of those smaller titles were intel performs better instead but they are not used in reviews (the whole TTW series runs like shit on AMD when you enter big 1k vs 1k fights), beamng is another title that runs much better on Intel etc.
And since you mentioned player counts, TTW and beamng get more players than BG3 according to steam.
We don't have a real average because the suites we see are too narrow to give a broad look. Also given each game is its own thing an outlier is not the correct terminology since the results stand on their own. An outlier would be if 1 result is high or low compared to other results in the same test for the same game. Further other review sites that test BG3 act 3 get similar results to HUB, if HUB had vastly different results vs those other review sites then that could be classed as an outlier vs the average of BG3 act 3 results but that is not the case either.
What a benchmark suite actually does is multiple individual gaming tests, removes outliers from those tests and then provides a meta analysis over the chosen titles to create a high level average. If a person only plays a specific title though then they can completely ignore the meta analysis and focus just on that single title for their buying decision.
It is fundamentally different from running the same test multiple times and getting erroneous results which is what the term outlier implies. Your methodology would exclude a lot of popular titles because they don't fit within the expected average deviation. That is just bad practice because those results are still valid.
Popularity is also an import factor, to a degree, because it helps prune the suite to something that will be useful to as many people as possible. It is also worth factoring in type of game, game engine, API and other things to create a spread that hits as many points as possible. For a given type of game in a given game engine using a given API you probably want to select the most popular titles within each category. BG3, being the most popular CRPG and a very popular game in its own right makes it a worthy addition to most test suites, its popularity will also remain high because it has a pretty big modding community and the fact you can easily tweak the editor to create your own campaigns will keep it going for a long long time.
BeamNG and ACC / AC are different types of game. BeamNG is a soft body physics game where you can race and that will become a bigger feature as more is added to it. ACC / AC are licenced sim racing titles that specialise in GT racing and have real race tracks. Fundamentally different titles with different aims. If a bench marker wanted to add BeamNG then sure why not, it is a popular title, more popular than a lot of other titles that get tested, and it does something different and stresses the computer in a different way to a lot of other games so it is absolutely worth having in a CPU review suite but it is not a substitute for ACC / AC / iRacing as popular sim racing titles.
I presume you mean total war by TTW as a google just gets the tale of two wastelands mod for Fallout New Vegas. Sure it can be tested as well, the series as a whole is popular enough and the game tends to have pretty long legs as most of these kind of games do.
I don't see any total war game above BG3 though, infact looking at 24h peaks I don't see any game that can be abbreviated to TTW above BG3 which is at 48.5k. Also they are fundamentally different types of titles, one being a real time battle game with a strategy layer and the other being a CRPG. They are complementary.
If I was to update the HUB test suite I would take out 1 or 2 of the multiplayer titles that they test and the likes of TLOU2, HZD, Spider Man 2 since those games are not that popular anymore and add in the likes of Beam NG, the latest Total war game, Something like Satisfactory or RimWorld or Dyson Sphere Program, I would add in HOI4 and PoE2 and I would add in Civ 6. Some of those titles would also get tic rate / turn time tests as well as fps tests because something like HOI4 is fine at 60 fps given the sort of game it is but it drags if you have a weak CPU and your tic rate is low which makes it functionally unplayable in any reasonable amount of time. I would also test those games in situations where performance has a genuine impact which would be late game for a lot of them because that is where the differences in performance will be noticed and where they actually matter as well.
That would increase how broad the suite in terms of game type, game engine and rendering layer while using popular titles that are going to help more people come to a buying decision. If GN and LTT and KitGuru and TPU etc did the same thing and chose some different AAA / Multiplayer / other titles within similar categories then across multiple reviews you would cover a lot more ground and have an even broader testing base. This is why I quite like PCGH.de and ComputerBase.de because their test suites will often include a few 'different' titles like Anno.
>the difference in 1% lows between the 270k and the 9800x 3d is 130 vs 145
I’m looking at their video right now, at 11:10 mark of their video shows the 14 Game average snd the 9800X3D shows 1% low as 190 fps compared to 157 fps to the 270K+. Going through their benchmarks, I noticed that the 9800X3D shines especially in multiplayers games
Except 5800X3D (and 7800X3D to that extent) suck at games that require high ipc, high frequency, and high multicore performance. Games like Cities Skylines 2 and BeamNG.
[https://www.youtube.com/watch?v=zbmOIOOBA3U](https://www.youtube.com/watch?v=zbmOIOOBA3U)
9800X3D is still decent even without vcache because it has good ipc and clocks very high, so it still works fine in games that can't benefit from big cache. But 5800X3D is a comet lake tier CPU paired with big cache. Well worse actually, because comet lake i7 is only slightly slower than 5800X, while 5800X3D's CPU is a voltage starved version of 5800X.
Cause graphical settings (some of them) impact CPU performance as well. Especially for example RT / PT are very heavy on the CPU as well due to BVH running on the cpu.
The extra cache, obviously. The x3d chips get very high maximum fps when the data needed are already in their cache, but they get a huge penalty when the data isn't in the cache. That's why their performance is erratic, ie. there are huge differences between lows and average. Intel chips are slower but smoother.
> If you don’t really care about gaming the difference in productivity is negligible in comparison to the price chasm between them.
Then you wouldn't be comparing to the X3D chips.
Still a large jump in price between the 270KP and the 9950X. Here in Australia it’s a preorder price of $519 for the 270KP versus a retail price of $899 for the 9950X. Nearly half the price for basically identical performance.
Fair. Honestly this just makes me more excited for Nova Lake vs Zen 6, it’s probably going to be a slugfest between Intel and AMD in a way we haven’t seen since at least Alder Lake and Zen 3, which as ever is a good thing for the consumer.
Yeah, even if these prices seem tempting, I'd still advise people to wait if they don't really need an upgrade today. NVL should be an enormous upgrade for Intel in gaming, and both AMD and Intel should get big improvements in MT from more cores + new node. ST, well I think the question is where in the 6GHz range both fall.
Well at this point AVX10 is really just restoring AVX512, so I'd expect the two to be more or less at parity. APX is interesting, but adoption will be very slow, so unlikely to matter for the lifespan of the platform. In 10 years, maybe.
You're forgetting the 270KP is likely also on a dead end socket. The value is further reduced with the knowledge you'll likely need a motherboard in 2-4 years instead just socketing in a new CPU.
I care about gaming but my 4090 doesn't care that I care about gaming - it severely bottlenecks even my 5 year old 12900k. CPUs don't really matter in gaming.
They don’t yet. They will, late this year, with their next-gen Nova Lake. Ofc AMD is not sitting still, they’ll have Zen 6, but that is expected next year.
AMD told us it wasn't needed, but everyone demanded for it. So they released it. There are probably some non-gaming workloads which can benefit from large caches on all the cores. Microsoft uses Milan X v-cache chips for instance.
But for strictly gaming yeah 9950x is better suited. As AMD have been telling us for ages.
Cancelled? This was leaked a while ago and popped up on benchmark databases. Besides these were expected to launch in its own bubble for the few that wanted these for the second v cache CCDs for HPC and other workloads.
By cancelled I meant “it’s done, but AMD doesn’t think it’s worth it to launch”. Mind you IIRC there were plans for dual CCD X3D variants for both Zen 3 and Zen 4, but they didn’t go ahead with them. There’s definitely a market of people that benefit from the extra cache on both CCDs, but that’s really niche, the REAL reason is to maintain halo status, which cuts into the 9950X3D’s territory and thus harms *that* CPU.
My counter to that is what you see in this very post and what you said: "barely an upgrade over their old one". It wouldn't cut into the 9950X3D's territory had they release it before and what I assume would be a price increase over it (or if AMD decides to lower the cost of the 9950X3D and take it's cost place with the 3D2).
The timing of the announcement is no coincidence. 9850X3D and this are both responses to ARL-R, but also both ARL-R and this X3D refresh have been planned for months. They just took different approaches to their interim refreshes.
Intel chose to go with a good enough CPU at a *very* good price. AMD choose to raise the bar and release the best client desktop CPU on the market (again). I think under normal circumstances, AMD's approach would've been the more exciting one. But in these specific market conditions, I think ARL-R's low cost is more noteworthy.
Announcement video - https://www.youtube.com/watch?v=_ErnOjwcWK8
Q2 2026, no pricing; announcement focused on workstation use cases.
From the professional focus, I'd expect a price around 1k, and gaming to still be limited by cross-CCD latency and scheduling between CCDs, given there's no mention of it in the announcement.
I'm not so sure, maybe you can get a TR CPU near the same price but the rest of the TR platform more expensive. And if you don't need ECC or a ton of memory and/or your use case can fit into the cache it won't be DOA in comparison.
Don't get me wrong, the use case is super narrow and I see people buying it for their gaming rigs where it won't much of anything
Yeah there is huge crowd of "only the best." The amount of people willing to pay for an 5090 is astounding. Nothing wrong with it if you want to have it but it is a lot of money.
I split the difference on that by only ever considering a 1999 msrp FE model from best buy. It's running paired with my 5800x3d in a 7 liter case. It started becoming easier to snag last year when i got mine, it's depressing that the prices are still continuing to tick up. The best buy drops seem to be drying up too.
You can't come close to a 5090 for overall capability esp with ai use cases incorporated. i as a rule am only interested in "good deals". I'd be pretty pissed and probably wouldve caved and got some second hand mildy scalped FE card if best buy didnt come through for me.
Yeah, if it were available and accessible at say US$1600 in line with gaming performance that’d be one thing, but gaming is probably a small fraction of actual sales.
The 5090 is faster, but it’s also a step change in capability above the 5080 for workloads beyond gaming.
Sucks for gamers, but that’s how markets work.
There is a market segment for example someone who needs to compile code for a living but also wants to play games, like me. Threadripper is more expensive especially with the motherboard too and has worse single core and gaming performance.
Even though it's small, I agree.
It would still be ridiculous to slap on 300 (or even 200) bucks compared to the 9950X3D with 5-10% gains in professional workloads and likely <5% gains in gaming.
Well it depends how much money that 5 to 10% in speed saves you. Compile times are a big issue for certain things and it can legitimately save people time.
For $1500 you can get a 24 core threadripper with 128mb l3 and 80 pcie lanes, or a 16 core threadripper pro with the same cache, quad channel ram and 128 lanes.
Threadripper has abysmal single core perf in comparison.... It's better for pure compute workloads for sure but if you have mixed loads it's very noticeably slower.
Example of mixed load is game dev, lots of parallel code and shader compilations, building assets, etc. But you're still hugely limited by single core perf driving all of those things, and when you aren't doing any of that and instead just iterating on a small thing that's where TR really holds you back. A 16 core ryzen chip is superior in all aspects for this type of work.
The other problem is that nowadays basically everything that is "embarrassingly parallelizable" can be offloaded to a distributed system. Unless you are a solo dev it makes a lot more sense to have a 9950x on your main computer that can keep up with scheduling distributed work and processing the results, while doing some of it locally with spare cores.
You know... yeah you're right to a degree. It does appear that the base and boost clocks on threadrippers have been catching up.
That said, you did pick the lowest core count threadripper, 9960x has just 24 cores and has a base clock of 4.2 ghz and boosts to 5.3 ghz. The 9950x and 9950x3d have 4.3 ghz base and 5.7 ghz boost.
So not quite as fast but pretty close. Things start to get more complicated when you look at higher core count threadrippers. They have lower base clocks, like the TR 9980X has just 3.2 ghz base clock, but it boosts to 5.4 ghz. This sounds good on paper, and would probably be fine fore actual pure single thread performance, but that's not what I'm talking about. I'm talking about a scenario where your overall performance is dictated by single thread performance but you also have load on most of your other cores. This is where threadripper starts falling behind because it isn't able to maintain the boost with all cores working due to thermal throttling.
Ultimately the threadripper has to be measurably better for it to be worth the hefty premium, not just "on par". There used to be a time when this was true for game dev, before most heavy parallelizable workloads could be offloaded to a distributed system. But now the only market I see for threadrippers is for workloads that can't be distributed and for work that requires a lot of PCIe lanes or more RAM than a ryzen 9 can handle.
fwiw I actually do have a couple of threadripper machines myself(5955WX and 3970X), but I use them as servers due to the number of PCIe lanes and RAM capacity.
You can't compare the cache. The important cache number is not the total, but the number per CCD. At least for most tasks.
TR with 16c and 128MB cache is 4 ccds, only 32MB per CCD. For cache hungry tasks that would be worse than the 9950x3D2.
TR does have significantly more memory bandwidth, PCIe lanes, and can have much larger total memory pool, so it has a lot of advantages, but cache isn't one of them.
The value prop of TR was already tricky before rampocalypse, because you had to get high end cpus together with filling out all the ram slots to get the actual bandwidth you would want. It was like $4k to start.
Now... it starts closer to $10k. it may as well be irrelevant for those of us evaluating performance per dollar
$1500 for a 9960X + $775 for the cheapest TRX50 motherboard = $2275
$1000 for a 9950X3D2 + $170 for the cheapest X870 motherboard = $1170
You can drive a whole market segment through that gap, they'll do just fine.
Well, I obviously have no clue over the pricing and as a more conservative estimate I'd lean lower considering the 9950x3d, but they mentioned how it's bridging the gap to HEDT a lot in the announcement video so also doing higher pricing for the professional market that has actual use cases for these feels reasonable.
Yeah, but you know there’s people out there that had the previous generation CPU and bought the 9950 X 3-D for $675 and for them that was what a 515% game for 50% more money? I mean if Andy had names of this the 10,850 X 3-D would you still be talking about it like that I don’t know.
Obviously there's a reason why AMD didn't release it (obviously very bad value). but since many people somehow still ask for it, I guess this is all for you guys.
Exactly. People were asking for this. I got burried every time I suggested it wasn't needed. Now that it's here they are still unhappy. Proof that you don't listen to redditors for shit.
reality is that the current stuff that people perceive as flawed (9950X3D) works relatively decently. People expect that "proper" (96+96 MB) version has to be so much better, while in reality the room for improvement is much smaller so when the "dream SKU" launches, it will get booed at.
I wonder if it's the same with Bartlett Lake. Has to be so much better because no hybrid and hybrid bad, right? You probably don't to meet your hero in this case, either.
They seems to hit some production bottleneck with Zen5 x3d.
Because there is no "Turin-X" yet, while previous "Epic-X" - server Zen cpu wilth x3d on **all** CCDs¹, were released year after initial release (table below from LLM).
|Gen|Codename|Arch|Release|
|:-|:-|:-|:-|
|1st|Naples|Zen|2017|
|2nd|Rome|Zen 2|2019|
|3rd|Milan|Zen 3|2021|
|→|Milan-X|Zen 3 + 3D V-Cache|2022|
|4th|Genoa|Zen 4|2022|
|→|Genoa-X|Zen 4 + 3D V-Cache|2023|
|5th|Turin|Zen 5|2024|
|6th|Venice|Zen 6|\~2026 (expected)|
¹- Yes, all. Eg. EPYC 9684X have over 1GB of L3 cache on CPU. Thats crazy!
The demand for x3d is a peculiar case.
[https://blog.cloudflare.com/gen13-launch/](https://blog.cloudflare.com/gen13-launch/)
Cloudfare for example moved on from Genoa-X to Turin and got their generational uplift without needing all the v-cache. That said, it wasn't like they needed it for HPC workloads but their previous stack with their old handling layer FL1 caused a lot of regressions without the v-cache and FL2 solved it.
It's more balancing supply and demand, more customers preferred Zen5/c CCDs without the need for v-cache and Genoa-X customers more or less are probably waiting for a bigger uplift to warrant the TCO and meet targets. AMD and it's customers are always communicating and is mostly the deciding factor for how they design their arch and stack before trickling it down to regular consumers. Anyways, AMD has already teased "Venice-X". That said we don't know if it's for regular Zen6 CCD with 12cores, or Zen6c with 32cores (said to also have per l3/core parity with classic cores this time around)
Okay, so it is not production issue, bud demand issue. That makes sense now, and explains why there is now this 9950X3D2, while there were no consumer double x3d before, and there are rummors about threadripper with x3d - because there is higher production capacity of CCDs with x3D. So the other way round than i though.
excuse my ignorance but what would Turin-X be exactly? Aren't there already a ton of Zen5 x3d CPUs available eg. 9800x3D?
Without wikichip I really don't know where to look up codename information :(
Turin-X would be Turin with all 12x CCDs with x3D.
There are consumer zen5 with x3d, but until now they were only single CCD x3d. For some reason they didnt made now those 8x or 12x x3D Epyc CPUs.
On wikipedia generations are[ nicely listed](https://en.wikipedia.org/wiki/Epyc#List_of_Epyc_processors).
Back when Zen 5 X3D came out, someone linked an analysis about the new V-cache. Unfortunately I can't remember where it was.
From what I recall, putting the v-cache on the bottom of the package had some disadvantage that made it less suitable for server CPUs, even though it fixed the thermal issues. Iirc it had something to do with routing the power through the v-cache, but I'm not sure.
Whatever the reason was, their take was that AMD designed Zen 5 V-cache for desktop first and had no intention of releasing Turin X, hence the configuration that is advantageous for desktop.
This is what I want to know. I dislike the core "parking" because if I'm watching a youtube video on the other window, when it changes to a new video, I get a mega in-game stutter as it's unparking for a few microseconds. I never had this problem on my 5950x system.
I’d be interested in running this CPU for gaming to have 2 VMs, each with their own ccd with 3D v-cache similar in level to a 9800X3D/9850X3D (and each with their own GPU and external display as well)…
…in addition to its workstation workload improvements.
The former IS more interesting to me than the latter because it would be a new capability not found on any other CPU really of its architecture whereas the latter is seemingly a marginal (albeit still significant) performance gain.
I get that gaming performance for this new chip won’t increase much if at all. But wouldn’t it vastly outperform the 9950X3D in my hypothetical use case of running 2 gaming-focused VMs?
I didn't expect much but this is still disappointing.
Obviously, gaming performance doesn't increase, the driver will still disable one CCD while gaming anyway.
But only \~7% in their own chosen tests with an increased TDP? quite underwhelming.
I was hoping that they'd do one X3D chiplet and one Z5C with at least 14 if not all 16 cores enabled
so you get 8+16 set up, best of both worlds, to match intel's P+E set up for the people who wants MC and ST gaming perf.
But alas, this is the well I hope someone who works with HPC or some sort of special thing will want it dealie.
One of the reasons why people aren't buying Intel CPUs is because they want predictable performance of a homogeneous architecture.
Interestingly, the industry seems to be moving back to where AMD currently is. With Intel working on a unified core again, and Qualcomm going from a firm "big + small cores" setup, to using one architecture in two roles.
Having two core types can increase MT performance while using less area than doing so with only big cores would, but introduces complexity and number of things that can go wrong, and a still pretty common scenario where something you want done fast is done at the pace of the slowest core on the package.
You gain some, you lose some, especially in the desktop space where you aren't as power or area limited.
the thing is, I'd agree with you if not for the fact that current 9950X3D already has the same issue
the X3D and normal non C cores are causing problems for gaming, if you dont lock your games to the X3D cores via process lasso you will have a degraded experience (or for some of the games that wants the extra clock speed, the non X3D cores).
so this isn't new with AMD, and AMD should still make single CCD processors for people who want that kind of performance, as well as dual NON X3D chiplets for people who just want more cores
but a Big.little arch for X3D + Z5C makes a lot of sense for people who just want both and if they are already going to deal with heterogenous compute units anyways, may as well make it good righht.
I don't disagree that such a product could exist and benefit a subset of users **in addition** to the single CCD lineup. The 9950X is a product exactly for the people who need MT performance even if it comes at the expense of cross-CCD latency and added complexity and occasionally erratic gaming performance. It's also a more niche product that beyond cost, people specifically choose not to buy as they see value in the simplicity and reliable performance of a single CCD chip with the fastest cores around (7800X3D, 9800X3D). These are the bread and butter consumer desktop chips. What I'm saying is that transitioning this type of product by slapping an extra small CCD of ZxC cores would be a mistake.
I'd like to say you are worrying too much, but I am not part of AMD and I can't tell the future.
But FWIW, the chiplet system was always designed in a way to allow for single or multiple chiplet styles, while intel went full in with an integrated design where they have individual design components that is then glued together into one soc.
if AMD stay the course, for desktop, it would be normal for single CCD with no mixed core as an option.
But if they want to have laptop specific or combined laptop also designs then yeah, it could happen.
AMD approach is derived more from servers, where well less space and power constraints, while intel has more dedicated lines and decided that their desktop chips are going to follow more laptop design.
Which apple is also doing with their M chips.
So who knows if AMD will eventually switch to that kind of design, or at least two lines.
AMD C-cores are homogenous. They have less cache and lower clocks, but are the same architecture.
That's the big difference with Intel's approach, which even forgoes AVX-512 as a result — something Zen C cores handle just fine.
They are the same architecture, with the same front end, same ALUs, SMT. They are a homogenous architecture. This is in contrast to Intel's e-cores which are a different architectures.
The 9950X's CCDs are binned differently and have different "performance characteristics". CCDs themselves have different top-performing cores. Using your logic the entirety of Zen is heterogenous.
Kernel schedulers can handle these just fine, and having more nT performance on AMD to compete with Intel would be good for consumers. If you really hate having c-cores I'm sure there would be an option to gimp the CPU in the bios.
There is a difference between a single digit difference in performance, and different if you're getting a double digit drop from significantly lower clocks or smaller cache.
You can see that a good chunk of the market rejects even the idea of heterogeneous cores split by a latency bridge, going for single CCD solutions instead of taking minimal complexity bump of having a two CCD CPU, even though it means double the MT performance. The 9900X often sells for about as much as the 9800X3D, showing that the desktop market rejects 50% extra MT performance if it's at the expense of small decrease in ST/gaming performance and just the added complexity of cross-CCD communication, not even heterogeneous cores.
Likewise, people aren't buying Intel chips which now have vastly higher MT performance than comparable AMD chips, and aren't far behind in ST/gaming.
The path for AMD to increase MT performance is to increase the number of unified cores per CCD, which they are already planning on. They know that a good chunk of the desktop market will still buy whichever CPU gives them the best reliably performant big-core-only CPU, even if it's "just" 16/24 cores, and not 32 or 48 that they may not even have a need for all at the same time.
Several points to address here and don't have the time to address them all in detail, but I'll do my best.
Lower clocks and less cache still doesn't make it a heterogenous architecture. Kernel schedulers have gotten sufficient at handling non-uniform resources.
Market specifics here are difficult to pull a conclusion from. Remember that Intel still ships more CPUs period than AMD. So really we're talking about the DIY market. It's not that surprising that the better gaming chip is in higher demand, consumers who want more nT performance would just get the full fat 16-core dual CCD option instead anyway.
I don't disagree that AMD is likely to put more cores on their Zen 6 CCD. However they are ignoring huge performance benefits if they also don't include a C-core die on many of those SKUs; they'll likely have a bunch of defective c-cores dies anyway they can fuse off to recoup costs by adding them to the lower consumer SKUs.
There just aren't reasons not to do it, for the 6/8 (or whatever the equivalent is for Zen 6) count CPUs they already have pads for the second CCD, tossing a 10/12 c-core CCD greatly improves nT potential, heads off Intel at the pass, and lets them salvage defective c-core CCDs.
If enthusiasts really are worried about 1T performance they can disable the second CCD, or just set process affinity.
Parking is not disabling. The 3D V-Cache Driver and Windows Scheduler send 3D-cache-benefited threads (like games) to the V-Cache CCD, while others are prioritized to the non-V-Cache CCD (like Discord, Chrome, etc.)
It's still very much in use and enabled.
I would imagine highly multi-threaded uses of this for certain games (like simulators) might see a performance boost here since there are more, potentially faster cores that can also use a similar amount of v-cache.
> Effectively is.
Effectively, it's not.
> And no, it's not while gaming.
Yes while gaming.
This is my core usage [right now](https://i.imgur.com/c0sEZqf.png) while gaming.
Do you?
Core parking is *supposed* to be marking cores for non-use to reduce power consumption and heat production.
What AMD and MS have bastardized it to mean is marking cores for non-use by certain classification of apps through the scheduler and 3D V-Cache driver. That is not what Core Parking is supposed to do.
This would be more apt to call it Core Prioritization.
???
You're the one who told me I'm wrong.
I gave you information. This is incredibly evident of how the CPU usage looks during normal usage.
If we want to go there, *you* are the one who both incorrectly said "the driver disables the CCD" **and** it "effectively disables it" so I'm going to need *you* to provide your source on that claim.
The 9850X3D is already just a top-binned CCD, which would typically go on a 9950X as CCD0, but with added cache. This is probably two of those, but for gaming you have to disable one CCD or you will run into intra-CCD latency, which is the same as going out to RAM (actually worse if you have tuned memory timings). That means that the main demographic that wants this CPU is literally just throwing money away if they buy it.
That's great for AMD, but sucks for everyone else because it means there is higher demand for top-binned CCDs, increasing the price for CPUs from the 9700X up, while providing absolutely zero actual consumer value for the vast majority of people buying any of those CPUs.
The only people that gain anything from this are the few that run productivity applications that leverage the extra cache without being hurt by the intra-CCD latency. As far as I know, that is a very slim selection of applications, and even there, the uplift is kinda laughable by AMD's own admission. Realistically, everyone is going to end up paying more for a very small uplift for a very small number of people, who are themselves also not getting much value from this.
It'll be great for the epyc 4005 server segment. Many are blade setups, so being able to offer 2×8 core X3D instances per blade is a boon. GSPs, some virtualized client work (e.g. cloud gaming) will really see benefits here.
It's been a long time since I've done any server work, but I have to imagine that the single NUMA node nature of the consumer CPUs presents at least a minor hurdle to that, no? At the very least, you will be reliant on your hypervisor layer to properly segregate the cores by CCD to make that viable, otherwise you will run into all the same intra-CCD latency issues. That's provided AMD hasn't thrown something more intentional into the mix specifically to prevent this as it would undercut their lower-cost-but-still-more-expensive EPYC options.
Workload dependent, of course.
Lots of game "servers" (just as games) only need a handful of cores, so if you're a GSP selling X3D game "servers" at a premium, you can now do double per rack/blade.
Similarly for client virtualization, you'd just give one VM per CCD, doubling your capacity per blade.
There's a market for low core count (4-16 core) X3D VMs/VPSes that run at higher clocks. It's one of the reasons EPYC 4005 is even a thing, you don't always need the tons of memory and pcie (and the expensive socket/board) but still want a (relative) lot of performance.
Why inst AMD placing the two ccds' closer to eachother adn then bridging the both ccds' with a big cache. This solution will still be penalised in latency sensitise applications like all the dual ccd skus before.
Well, we all know why, to do that next gen or maybe the gen after, ie milking us :D
\> placing the two ccds' closer to eachother
A vastly better interconnect (and IOD) is coming with zen 6
\> bridging the both ccds' with a big cache
Very costly, you won't see anything like this in the consumer market anytime soon. Also doesn't make much sense for most applications, and gaming can happily stay on one single ccd, so no need for that.
why would it be slower? u have a cache that allows all of the cores from both of the ccd to access the cache slices without travelling down the io die to the other ccd.
And the games workloads are only preventable when u enable game mode, ie gimping the cpu itself which many are not keen too do, and many are so against it when people with intel are doing that to run the cpu as an hexa or octa core.
The 9900X3D, 9950X3D and now this are not meant for gaming. They are productivity CPUs that can game as well as a 9800X3D, but you're stupid to buy one if you only game. It's not really "gimping"; it's getting basically two in one.
For example, I game and work on the same machine. My work demands a 9950X, so I would previously need to sacrifice some gaming performance to be able to effectively work. With a 9950X3D, I get the best of both worlds: a 9950X when I'm working and a 9800X3D when I'm gaming.
what does it matter if I am stupid for buying such a cpu or not, then no intel cpu is suited for gaming?
Nah, this is a perf cpu with a halo price, then why not having it cost a bit more if the perf is higher because of lower latency between the two ccd's.
Indeed. It is the worst time to be releasing a product like this. Intel has finally released a good CPU with respect to both performance and price, and here AMD is releasing this thing that practically serves nobody.
I actually haven’t even played a game on this pc since putting it together
Been playing my bazzite console more
But this will become a dual gaming and cpu heavy system. If I can ever pull myself away from my m5 Mac to use it
You’re not gonna like AMD’s mobile processor naming scheme. Or Intel’s. Or Qualcomm’s. Actually the only company with a sane naming scheme for CPUs now is Apple lol.
217 Comments
Culbrelai@reddit
fiduce@reddit
dfv157@reddit
kikimaru024@reddit
Fr0stCy@reddit
00Koch00@reddit
kittymoo67@reddit
5u114@reddit
Reggitor360@reddit
forgottenendeavours@reddit
DeltaPeak1@reddit
Canadian_Border_Czar@reddit
DeltaPeak1@reddit
Canadian_Border_Czar@reddit
BloodyLlama@reddit
DeltaPeak1@reddit
Canadian_Border_Czar@reddit
DeltaPeak1@reddit
Aggravating_Ring_714@reddit
michaelsoft__binbows@reddit
Nuck_Chorris_Stache@reddit
Fearless-Area-532@reddit
teh_spazz@reddit
noiserr@reddit
arahman81@reddit
TheFondler@reddit
teh_spazz@reddit
Wizard8086@reddit
theholylancer@reddit
teh_spazz@reddit
__________________99@reddit
Raknaren@reddit
PlsDntPMme@reddit
milk-jug@reddit
ThiccSkipper13@reddit
Total-Guest-4141@reddit
EnglishBrekkie_1604@reddit
Slabbed1738@reddit
EnglishBrekkie_1604@reddit
trackdaybruh@reddit
Good_Season_1723@reddit
timorous1234567890@reddit
Good_Season_1723@reddit
timorous1234567890@reddit
trackdaybruh@reddit
Johnny_Oro@reddit
Good_Season_1723@reddit
Johnny_Oro@reddit
Good_Season_1723@reddit
trackdaybruh@reddit
Good_Season_1723@reddit
trackdaybruh@reddit
Good_Season_1723@reddit
trackdaybruh@reddit
Good_Season_1723@reddit
trackdaybruh@reddit
Good_Season_1723@reddit
trackdaybruh@reddit
Good_Season_1723@reddit
trackdaybruh@reddit
Exist50@reddit
EnglishBrekkie_1604@reddit
Exist50@reddit
EnglishBrekkie_1604@reddit
Exist50@reddit
EnglishBrekkie_1604@reddit
Exist50@reddit
DaMan619@reddit
AHrubik@reddit
F9-0021@reddit
Good_Season_1723@reddit
b-maacc@reddit
trackdaybruh@reddit
Good_Season_1723@reddit
trackdaybruh@reddit
Good_Season_1723@reddit
greggm2000@reddit
Jumpy-Dinner-5001@reddit
Good_Season_1723@reddit
greggm2000@reddit
kittymoo67@reddit
Jumpy-Dinner-5001@reddit
noiserr@reddit
EnglishBrekkie_1604@reddit
noiserr@reddit
Noble00_@reddit
EnglishBrekkie_1604@reddit
Noble00_@reddit
FitCress7497@reddit
soggybiscuit93@reddit
Numerlor@reddit (OP)
Admirable_Bid2917@reddit
deadbeef_enc0de@reddit
BFBooger@reddit
3G6A5W338E@reddit
deadbeef_enc0de@reddit
Sopel97@reddit
ElementII5@reddit
dgibbons0@reddit
michaelsoft__binbows@reddit
airmantharp@reddit
airmantharp@reddit
zxyzyxz@reddit
Admirable_Bid2917@reddit
zxyzyxz@reddit
froop@reddit
steik@reddit
froop@reddit
steik@reddit
froop@reddit
BFBooger@reddit
michaelsoft__binbows@reddit
Ok-Parfait-9856@reddit
Navi_Professor@reddit
zzzoom@reddit
Numerlor@reddit (OP)
capybooya@reddit
timorous1234567890@reddit
Tripleberst@reddit
Aos77s@reddit
Subtle_Tact@reddit
DaMan619@reddit
someguy50@reddit
ritz_are_the_shitz@reddit
windowpuncher@reddit
Fritzkier@reddit
noiserr@reddit
JapariParkRanger@reddit
jocnews@reddit
ElementII5@reddit
gorion@reddit
Noble00_@reddit
gorion@reddit
qscqe@reddit
gorion@reddit
jonnywoh@reddit
einmaldrin_alleshin@reddit
imaginary_num6er@reddit
Nuck_Chorris_Stache@reddit
bigMeech919@reddit
Hordsak@reddit
PrimaryRecord5@reddit
mgwair11@reddit
Jumpy-Dinner-5001@reddit
kikimaru024@reddit
TheFondler@reddit
TheDregn@reddit
airmantharp@reddit
windowpuncher@reddit
theholylancer@reddit
glitchvid@reddit
PastaPandaSimon@reddit
theholylancer@reddit
PastaPandaSimon@reddit
theholylancer@reddit
glitchvid@reddit
Jumpy-Dinner-5001@reddit
glitchvid@reddit
Jumpy-Dinner-5001@reddit
glitchvid@reddit
PastaPandaSimon@reddit
glitchvid@reddit
Berzerker7@reddit
Jumpy-Dinner-5001@reddit
Berzerker7@reddit
Jumpy-Dinner-5001@reddit
Berzerker7@reddit
Jumpy-Dinner-5001@reddit
Berzerker7@reddit
Jumpy-Dinner-5001@reddit
Berzerker7@reddit
Jumpy-Dinner-5001@reddit
Berzerker7@reddit
Jumpy-Dinner-5001@reddit
Berzerker7@reddit
Jumpy-Dinner-5001@reddit
Sworn@reddit
pythonic_dude@reddit
TheFondler@reddit
glitchvid@reddit
TheFondler@reddit
glitchvid@reddit
zakats@reddit
Paliknight@reddit
knflxOG@reddit
Pillokun@reddit
Kryohi@reddit
Pillokun@reddit
webjunk1e@reddit
Pillokun@reddit
DoktorLuciferWong@reddit
Savage4Pro@reddit
rossfororder@reddit
MOSTLYNICE@reddit
Slasher1738@reddit
Ander12391@reddit
teh_spazz@reddit
parallel_mike@reddit
Method__Man@reddit
LIGHTNINGBOLT23@reddit
SalivateTheStarfish@reddit
TheRealSeeThruHead@reddit
capybooya@reddit
TheRealSeeThruHead@reddit
webjunk1e@reddit
TheRealSeeThruHead@reddit
The-Bite_of_87@reddit
TheRealSeeThruHead@reddit
SBABakaMajorPayne@reddit
EnolaGayFallout@reddit
Vagamer01@reddit
EnglishBrekkie_1604@reddit
Cohibaluxe@reddit
Weird_Tower76@reddit
Gloriathewitch@reddit
pythonic_dude@reddit
danielkoala@reddit