Zen2's memory write gimped

8c and less got only 1 chiplet
mot than 8c are obviously 2 chiplets but don't ask me how its layed out inside the CCX's

First the value is comparatively shit.

Second when I benchmark my own usage right now I'm basically only able to push a haswell to its limits a few times a day. Really anything over a 3700x is overkill for my usage. 3950x feels like double overkill. I have friends that could abuse 3950x but I just don't use those workloads. I tend to be memory bound not CPU bound. I tend to get the most value out of old used tech with high ratios of memory to CPU performance like old 4CPU servers with 32 dimms.

Third 3950x it sort of feels like the 2nd act of Intel getting BTFO rather than the 1st act.

3800x is significantly more expensive than 3700x though, might as well be 2 chiplets? Anyone delidded it yet?

no. someone explained this on YT. this is an intentional architectural choice AMD made as it has little to no difference on real world performance. From what i understood, this was to leviate some other changes that did in fact have performance impacts

zen2 kinda works like this

3950X
(4+4) + (4+4)

3900X
(3+3) + (3+3)

3800X / 3700X
(4+4) + (0 +0)
- literally single CCD

3600X / 3600
(3+3) + (0 +0)
- literally single CCD

I'm also I/O bound quite a bit. I really just don't use CPU that much. I do big personal projects but rarely are any of the VMs or containers I spin up running very hard since they're not being used by any actual users.

If I was being practical I would buy B450, 64gb of ram, and 3700x and call it a day. Yet I do love computers and sometimes you just want to build a powerful machine for its own sake.

>32mb L3
It's single chiplet

Attached: 3800x_specs.png (1448x479, 41K)

gimped one too.

I like to know the implications of this here

>I like to know the implications of this here
Endless intel shilling over a non-issue.

More of a real world performance impact...I reckon RAM is written to fairly often and halving write speeds would add time spent buffering datas. Inside of things bottlenecked by processing speeds it would be absent, but I fear it would be perceivable in interactivity.
Obviously it's not going to be the end of the world and turn the computer into something unenjoyable...but still...I baka.

I guess that's were the "gamercache" might come in.
It's a design and went with because they say (for whatever reason). That it has barely any performance impacts on realworld applications.
Not all architectures are made alike.

what do you fucking mean. I get 15k mb/s LESS than a 2700x and you're telling me this is no performance impact for anything?!?!

This is what happens when poojeets get their hands on 7nm

ANOTHER
MASSIVE
DISAPPOINTMENT

/v/ tourists need to be gassed

>Half the nm means half the bandwidth, right Rajeesh?

1 CCD = 2 CCX
1 CCX = 4 Cores

And the CCD is one of the die packages you see when you delid a Zen2 CPU.

Calm down, take a hit off your cock bong and tell me where the bad chip touched you.

You know, I'm all rooting for AMD but they already went once with the "shit solution but doesn't matter in real life applications" route with Bulldozer's modules sharing stuff between cores. Hope this approach doesn't knock them out again.

>3900X
>(3+3) + (3+3)
Are we certain it's always like that? If it were for example sometimes that, and sometimes (4+2) + (3+3) or even (2+2)+(4+4) it would cause serious performance discrepancies between SKUs due to cache architecture. Do we know that with 100% certainty?

It doesn't. Similar to how NVME has near zero performance gains over SATA other than the fact it has more raw throughput.
Seek time and latency are all that really make a difference

>Seek time and latency are all that really make a difference
It's difficult to notice the gains when those parameters already are at levels not perceivable by human senses.

it needs to be symmetric

>needs to
is it though?

I wonder what would have happened if 3000 was a die shrink instead of the change to chiplet+die
is there some technical talk where they go though the reasons for the change to chiplet+die?

>is there some technical talk where they go though the reasons for the change to chiplet+die?
Yields

Why the fuck would they do this

Yields and I think it helps spread the heat as the heat density is pretty bad on 7nm

Because it doesn't matter, you fucking retard. Real-world performance between the 3700X and 3900X in reviews was exactly where you'd expect it to be. The 3600 also got a glowing review from Gay Man's Next Ass. Were you planning to buy one to sit there running AIDA64's memory benchmark all day every day or something?

Why even bother having support for things like pcie 4 that dont currently matter if youre going to hard gimp other areas that dont currently matter? What unforeseen problems might this bring rise to

Of course it is, dumbass. The chips literally don't work any other way. Every Ryzen chip released to date has had symmetry between CCXs and now CCDs, because it's a necessary part of the architecture's design.

There's no point in even talking to a tech-illiterate retard like you. You don't even have a basic understanding of what you're talking about. Enjoy sitting there wallowing in your fear and ignorance.

ok

Attached: thisisyou.jpg (495x638, 46K)

yes

>memory bandwidth doesn't matter
And so it begins.

not him, but there's no point in being all high and mighty. don't forget, you might be talking to a child, or worse, an indian.

how bad would content creators feel 25GB/s vs 50GB/s?
what use case does this hit the most?

cope more

What kinds of workloads would be disproportionately affected by this?

a lot. think of ramdisks, tmpfs etc etc

3800X uses a single one of the highest binned chiplets, the same with the 3950X, which uses two.
That's why 3800Xs are so scarce. And why the 3950X wont be available till September.

k? It's proven that it doesn't matter
This is like the high memory latency FUD.

anything that can use more than 25gb/s. so pretty much nothing anyone is going to ever use a 3700x for.

t. seething here but this most likely. it's weird for AMD to forcibly include a catch to lower SKUs, but if you were to be saturating this with your workloads, you'd probably be on a higher SKU already.

while it's a downgrade from 2700X, was the 2700X even reaching those 50GB numbers outside of synthetics?

No workload that has been found so far. That means no games, no productivity software that is normally benched.

is this writing TO RAM or writing to something else from RAM (read->write)?

because if it's the former then at Ryzen 7s and down with their officially supported 128GB of RAM we're talking 5 and some change seconds to fill it up instead of half of that. And that's assuming from empty.

Yep I know, that's why (as I just said) nvme ssd vs sata ssd is hardly noticeable

What are you filling the RAM from?

you tell me, content creators are up in arms about the reduced speed

Describe to me a work that just writes to memory? Aida's copy metric is the most indicative of what real word applications actually do and read comes close second.

no they aren't

so we agree that the half speed isn't a huge issue to a majority of people

now we must decide if we are coping or seething over AMD doing this

>while it's a downgrade from 2700X, was the 2700X even reaching those 50GB numbers outside of synthetics?
Nope, that's why they cut it down in the first place, so they can make more noticable real world gains in other areas.

>implying harddrives would be faster

Attached: brainlet.jpg (326x294, 12K)

I mean, good luck getting more than 32gb of ram on a x570 board without paying out the ass for ram, four 8gb sticks only get you to 32gb. and if you're paying extra for ram and a premium mobo, then paying a little extra for the 3900x probably isn't an issue.

>performance doesn't matter

shitpost aside, come back when you can saturate it outside of synthetics.

uhm...
sweetie...
if you have half the bandwidth then all memory writes take twice as long
it's not just some bottleneck that makes things slow once you hit the maximum

yes but how much are you writing and how much can the destination hold
you seem to have trouble answering that

This is the bandwidth and not "write speed", tranny.
And it hasn't been show to affect any actual application or game so what does it matter?

that's not how memory bandwidth works, every write operation makes use of the maximum speed.

now find something that's writing that much

>This is the bandwidth and not "write speed"
>Drone thinks those are actually separate, unrelated things

It's gonna affect load times for games for sure

It seems you're lacking living brain cells to understand that EVERY SINGLE thing that does ANY WRITING AT ALL is affected

Attached: 1512731514713.png (389x432, 55K)

great, now find a group of things that are all writing that much

>OH NOES I CAN ONLY WRITE 25 GIGABYTES TO MY RAM EACH SECOND
>THAT'S A YUUUUGE PROBLEM BECAUSE MY SSD CAN READ 40 GIGABYTES FROM DISK EACH SECOND

This is how fucking stupid everyone in this thread is. Please kill yourselves immediately.

Every single application will write to memory, the ironic part is that they are shilling PCIEX4 bandwidth, which will require fast memory writes, while gimping memory to save a few cents

we'll see about that

>tfw 3750X and 3850X is a possibility

>(4+0) + (0+4) with the same amount of cache as your 3950

the absolute state of Jow Forums

The words are different for a reason. This is the maximum bandwidth. It does not mean that writing 1MB takes twice as long. That would be extremely noticeable in benchmarks...

No

Attached: chrome_2019-07-09_17-01-32[1].png (1018x626, 182K)

NOOOOOOO DELETE THIS!!!!

Attached: winning.png (351x376, 2K)

Fuck outta here. Go troll the other threads.

are you actually retarded, no one is going to get 128GB ram and use anything less than a 3900x

content creators are batshit retarded if they dont get 3900x or just wait.tm for 3950x

dmg control

I was just giving the absolute max of the Ryzen 7s 3700X/3800X and how that writing performance stacks up.

It's nothing to do about the intelligence of anyone, but maybe you just outed yourself.

Here's an idea. Maybe, just maybe, your SSD would be the greater bottleneck here. That is it would be, if decompressing the compacted data wasnt the real bottleneck. Not trying to imply that you're a literal retard, but you are.

>b-but its gimped
3700X is faster than the 2700X is literally everything. Every game, every productivity test, every professional program, even Premier.
Memory performance isn't a significant factor to performance here.

Someone explain in baby-speak for me what the fucj this means

>what this means
On paper the single die Zen2 chips have lower peak memory bandwidth.
In practice however the effect is nil. The 3700X isn't slowed down by anything, its faster than the 2700X across the board, and per clock performance is identical to the 3900X which has higher memory write speeds. So again the effect is nil.

Its just shills grasping at straws.

>Please kill yourselves immediately.
I would but I'm too stupid.

perform the same miracle you did when you survived birth

sometimes to go fast you must go slow

and at all times to be retarded you must choose intel

>asks for examples
>gets called a shill
I hate this board sometimes

t. shill

Something is seriously wrong with their test benches and I don't mean in regards to the AMD benchmarks.
I have the exact same RAM and much better read/write/copy and latency. Like ~15% better across the board on an 8700k.

40gb/s sounds about right for 2666mt/s