So, theres basically 3/4 bins of Zen 2 dies, each subdivided into 6 and 8 core variants.
At the very top, you have the Epyc/TR/3950x dies. This may be actually two bins, with the absolute highest end Epycs getting the top 1% of dies.
Next you have A grade dies, that go into the 3800x and the 3600x, and the first die of the 3900x. After that, B grade dies, that go into the 3700x and 3600, and the second die of the 3900x.
The 3950x is the weird one because it doesn't use two A dies (the 3900x uses an A+B)... it uses two Epyc/TR grade dies.
Those new XTs? Seem to be miniature 3950xes: 6+6, 8, and 6 Epyc/TR grade dies. Everyone who wanted a 3950x, but didn't want/need 16 cores? Here you go.
Although, honestly, unless you need a computer today, and as much as I love Zen 2, wait for Zen 3 as long as you can. Zen 3 uses 8 core CCXes, which eliminates the semi-NUMA split L3 latency issue.
That issue is incredibly mild, it isn't the cluster fuck that it was on Zen/Zen+, but it's enough that some especially badly optimized software (older games, mostly) is about as fast on a 3300x (single CCX quad core) as it is on the 3600 (dual 3 core CCxes), purely because of the ignorance of mem/cache locality.
That said, for a dumb desktop and PS4/XBOne-era gaming performance, the 3300X on a MSI B450 Max /w 16GB DDR4-3600 CL18 has exceptional value; just toss in a 1660 Super, and its a 1080p 60-120hz gaming king.
Why would Epyc get the best binned silicon? Would the best binned Epyc run substantially cooler at the same clock than the next worsed binned Threadripper?
Just curious.
I think you answered your own question. The reason for server chips to use better silicon is to generate less heat (and consume less power) at the same frequency. Server operators (unlike desktop users) are more concerned with non-idle power consumption. Besides, server CPUs are semi-passively cooled, there's not enough room in a server chassis to accommodate elaborate cooling setups common in workstations and especially in enthusiast machines.
Have you checked the CFM of server fans? Servers have much better cooling that a desktop, that's the reason they sound like jets during normal operation.
but a 24 core threadripper has a TDP of 280w. The 24 core epycs of the same generation have 155w or 180w (depending on clocks).
The clock speeds are a bit lower on the epyc models, but but I would be willing to bet that chips that cant hit the lower wattage targets just get pushed as high as they can go, and it is then on the consumer to manage the extra heat.
> Although, honestly, unless you need a computer today, and as much as I love Zen 2, wait for Zen 3 as long as you can. Zen 3 uses 8 core CCXes, which eliminates the semi-NUMA split L3 latency issue.
Not sure about this. Today in Zen 2, in a 4 core CCX the L3 cache is split between the 4 cores and a core can only directly access its own slice of it and has to go through the Infinity Fabric to access L3 cache of other cores in the same CCX. Just because Zen 3 has 8 cores CCX doesn't necessarily mean that they all access the entire CCX L3 in uniform manner. The latter seems much more important to me than bundling 8 cores in a CCX.
Do you have any links that support that statement that cores inside a CCX will have uniform L3 access in Zen 3? I can only find rumors, no official information on this.
Zen 2 isn't a dual ported singular 32mb L3, its two distinct 16mb L3s. If you need data from one, but your core is from the other, you have to pay the cost.
Due to Zen 2 monolithic IO dies make this cost far smaller than the Zen/Zen+ design, since it only has to route inside of the IO die, and not use valuable IF<->chiplet lanes.
AMD has soft confirmed that they are rolling with 8 core CCXs, so for desktop Ryzens that are 8 cores or less, it will be a single CCD with a single CCX inside, thus a single L3 domain.
And yes, you are correct: if they turn it into a single, but dual ported, L3, two CCXes in a CCD would also be fine and optimal. However, this seems to be more complex than just making bigger CCXes. Intel has horrifically overcomplex memory controller designs because this is how they typically design things, and AMD wants to avoid repeating that mistake.
Given that Zen3 doesn't exist yet, there's no legitimate research you can do about Zen3 (unless you are an AMD insider and have their NDA'd numbers / specifications).
Generally speaking, waiting to buy a computer is always better, because next year's tech is always better than this year's tech. So you only buy a computer when you need one.
Is your current computer slow in ways that you find annoying? Slow enough that you want an upgrade? If so, buy a new computer (or if its only slow in one or two benchmarks... like I/O or Video games, buy a faster SSD or GPU instead). If not, then wait another year or two and ask yourself the question again.
Big SV companies seem to be on a 5-year upgrade cadence. If you can't make the decision yourself, then just follow what they do and upgrade your equipment every 5 years.
There’s a general shortage. I just built a workstation using a Ryzen 9 3900x and it was very hard to find a compatible, sanely priced, current design and with a good enough spec. I was doing research first and then trying to buy the motherboard but they would be out of stock. Had to change model 3 times before wising up and just going through available inventory and then research.
Then the board wouldn’t POST :/ having a concrete issue to google for then surfaced tons of threads discussing the same or similar issues that hadn’t come up when I was doing research before buying.
It took me a few days to finally be able to make it boot- it was probably an acculturation of small silly things like the board being picky about which cable/slot you plug in the monitor, removing and installing again RAM even tho it was seemingly well placed before, etc.
BUT! I have now a nice beast running Pop_OS and I couldn’t be happier. I hadn’t had this feeling of noticing a performance gain in years of upgrading from MacBook to MacBook. I’m giggling of excitement just from thinking about working on that thing :)
And what about the 4?00{H,HS,U} series? The 4700U has 8 cores and no threading, the 4800U has 8 cores with threading. Is that also a result of binning?
Frankly if they can launch I think they'd be insane to delay, every year they've had the rumours first then come out with a body blow to Intel, that momentum has been building and they've finally started to get real mind share for the first time in nearly a human generation so it would be silly to lose that now.
Also if I had to guess why they put these out it's because production improvements have meant they've got more available high quality dies than they had at the start so it makes sense to get a last hurrah off the line and it preserves the price points for each back to somewhere near where they where at launch.
Agreed - they've been releasing 3000 series CPUs all year so I wouldn't read any tea leaves based on this launch. 3300X and 3100X launched just a month ago and people weren't panicking then. They have the capability to manufacture more of Zen2 so why not do it. I'd imagine a "new" product sells much better than a few years old one even if its just a re-brand with a slight overclock.
It would be smart to capitalise while Intel is struggling because Intel won't struggle forever they have too much money and too long a history.
I was around when the AMD64 was battering the P4 and then Intel came out with the C2D an architecture they are still using which gave them a decade of owning the top/mid in desktops, pretty much all the laptop market and vast majority of servers.
Better to be out ahead when Intel comes back swinging.
Actually, there is (kind of). The AMD presentation slides that have been released after March (that I could find, and do not have Zen 4 on the unmarked timeline as well) all suggest Zen 3 as an item on a timeline ending 2021 [0][1], as opposed to earlier slides [2][3] that have Zen 3 on a timeline ending in 2020. The only Zen 3 product which was announced to be shipping this year, was Milan at "late 2020" [3], and that was before the presentation of [1] was given, so that might very well have changed in the meantime.
Looking back at it, I really haven't seen an AMD presentation with the [ 2017 < zen, zen 2, zen 3 > 2020 ]-timeline after the financial analyst day 2020, but have seen new updated versions in the form of [ 2017 < zen, zen 2, zen 3 > 2021 ].
> AMD also re-confirmed that we are set to see Zen 3 processors in 2020.
So, as of last month, Zen 3 was again confirmed to be coming this year.
It's absolutely still just a rumor that Zen 3 is delayed until 2021. There has been no official confirmation of any change to the schedule, and attempting to read way too much into some slides doesn't change that.
> > AMD also re-confirmed that we are set to see Zen 3 processors in 2020.
> So, as of last month, Zen 3 was again confirmed to be coming this year.
Yes, I already expected a presentation at the end of the year, though this is not explicit about _shipping_ Zen 3 this year. /cynicism
You are correct though that it's a rumor, and I'm 80% sure that Zen 3 will launch this year. My comment was mostly to provide a counterexample to the "almost no evidence for Zen 3 in 2021"... Although, this evidence is indeed not a lot...
> "The new 3000XT family of processors focuses mostly on boosting the turbo frequency by 100-200 MHz for the same power."
Except they've only done this for the high power 95W/105W models, not the 65W 3700X I actually care about, so "for the same power" isn't very exciting.
It's exciting to see AMD challenge Intel in single core speed as well. For most folks, number of cores are far more important, but there are still plenty of applications that benefit tremendously from fast single core speed. Will these speed increases affect the server lines as well?
So now there are new slightly faster processors for three Ryzen 3000 processors, two of them were always a bad deal so far and that won't change with a few MHz more. A "faster" 3900X is nice of course, the 3900X was already nice, but without the cooler this is effectively a big price increase for virtually no gain. All while making AMDs sku lineup more confusing. What an unnecessary release.
I hope that benchmarks prove me wrong here, but I doubt it.
With higher TDP they achieve higher sustained speeds than the 3600/3700X. A further boost at the same price point seems a good plan. Presumably the non-T parts are superseded by these.
Because they are not significantly faster in practice. The higher TDP does not help as much as you would think. Gamersnexus called it "paying for the letter X"[0]. I see the same in my meta benchmark in games especially[1], but also in apps[2]. There are differences, but if you are not in a situation where every single inch of performance counts that's not worth the higher price. It's not like that's just a few cents.
And now you have a price increase of $50 (for a good air cooler) to $100 (if you follow AMDs advice of buying a 280mm AIO) on the 3800XT. Makes the deal a lot worse for most, if there are not hidden performance advantages not visible so far.
There's a fairly small performance increase, around 4% sure, but the price difference is only about 10% in the UK at present. yhese XT processors increase it a little more at the same price point.
Really not sure what there is to complain about...
0% for the Ryzen 7 in games, 2% for the Ryzen 5s. In applications 2% and 4%. Might differ a bit with a different benchmark selection, but that's the gist. 10% higher price for that is already bad.
> These XT processors increase it a little more at the same price point.
Only for the Ryzen 5s (and the 3600XT could come out a bit better looking). For the Ryzen 7s the price point is not the same, you have to factor in the cooler you have to buy now. That will make the offer even less attractive.
The CPU is less and less relevant for gaming. The reviews I've seen put it out ahead consistently but only by a 1-2fps. And as you say, a few percent in applications. If you want those couple of percent then you'll pay a 10% higher price for it. Disproportionate but to me it's still a great value processor.
And also for me the price would be the same - AMD's wraith cooler was noisy and I ditched it ASAP on both the 3700X and 3800X we have in the house here. They're in boxes in the attic so pretty much have negative value.
I wouldn't say that. We had a couple of very cpu heavy game releases recently, and with new console generations that usually makes a jump.
And you can get a bigger performance jump in games - with Intel.
> The reviews I've seen put it out ahead consistently but only by a 1-2fp
Right, rounds down to 0%.
> Disproportionate but to me it's still a great value processor.
No, by definition it's not a great value anymore if you pay that much more for that small gain, if you have equally fast alternatives after a minimal overclock.
You could have sold the Wraith cooler :) But that's a valid argument, for some the cooler inclusion or non-inclusion does not matter.
AMD won't do that change without a reason. They probably feel they are in the position for that now. They boosted the stock cooler importance for Zen 1 (that thing was great) gaining goodwill, then made it worse, now start to remove it completely. The infinite circle of corporate attention and customer focus in action.
In games the 10600K[0]. And the 10400(F) is equally fast[1]. Outside of games the 3800X wins against those two, though the 10600K could keep up with it a bit better if we would look at its overclocked results.
I was specifically talking about gaming, and mentioned in all my comments specifically what performance category I'm talking about when making performance claims.
Actually if you go back to the top of the thread there are no such claims, just that the chips are not good value.
I'd argue that going forward the intel chips are going to hold their own less well in games too, with a lower core count, especially as the core count in consoles goes up.
You do not want to ever use the stock cooler on the 3900X, it does work but at the cost of high fan RPM cool it and you'll want to stab your ears out from the fans ramping up and down cyclically every 15 seconds from the aggressive PBO bumps that translate to immediate temperature spikes and CPU fan ramp ups. You want some of the larger heatsinks available like the Dark Rock Pro which can suck up the heat without needing to spin up fans.
Yeah, while AMD's stock coolers reviewed well on other CPUs, I went a few weeks with the wraith prism on a 3900x then decided it had to go.
Though if these are the same more efficient chiplet bins as the 3950x, maybe the prism would make more sense on a 3900xt. Given they're not including it though, I guess AMD didn't think so
Same. The stock fan on the 3900x was a bit noisy - spinning up and down all the time. Replaced it with a closed loop water cooler after a couple weeks.
I bought a 3900X and have a stock cooler on it, and have a Dark Rock Pro 4 sitting next to me which I purchased because the stock cooler was way too loud. If you build a PC with a high end Ryzen in it, you should listen to this advice.
They'd already done this kind of things with the 'premium' epyc7371, a 7351 (16c 2.4GHz beating Xeon Gold 6130 for my workloads...) pushed at 3.xGHz? 13-18% perf increase on my workloads. More expensive but if you need the low-latency / high ipc (but no AVX512, well...) these AMD processors are beasts. Still love the 7551 too, for the price... And those were only Zen cores... Can't imagine Zen2/Zen3.
Yeah there's a shortage. They end the production run just before Chinese New Year and spin up a new one after. But then covid hit. And many people upgraded their gaming PCs during the lockdown.
Not sure the rules on this... but I am moving from ITX to mATX probably next week. If you're interested in miniITX b450 I don't really have a plan for mine, email is in profile.
At the very top, you have the Epyc/TR/3950x dies. This may be actually two bins, with the absolute highest end Epycs getting the top 1% of dies.
Next you have A grade dies, that go into the 3800x and the 3600x, and the first die of the 3900x. After that, B grade dies, that go into the 3700x and 3600, and the second die of the 3900x.
The 3950x is the weird one because it doesn't use two A dies (the 3900x uses an A+B)... it uses two Epyc/TR grade dies.
Those new XTs? Seem to be miniature 3950xes: 6+6, 8, and 6 Epyc/TR grade dies. Everyone who wanted a 3950x, but didn't want/need 16 cores? Here you go.
Although, honestly, unless you need a computer today, and as much as I love Zen 2, wait for Zen 3 as long as you can. Zen 3 uses 8 core CCXes, which eliminates the semi-NUMA split L3 latency issue.
That issue is incredibly mild, it isn't the cluster fuck that it was on Zen/Zen+, but it's enough that some especially badly optimized software (older games, mostly) is about as fast on a 3300x (single CCX quad core) as it is on the 3600 (dual 3 core CCxes), purely because of the ignorance of mem/cache locality.
That said, for a dumb desktop and PS4/XBOne-era gaming performance, the 3300X on a MSI B450 Max /w 16GB DDR4-3600 CL18 has exceptional value; just toss in a 1660 Super, and its a 1080p 60-120hz gaming king.