AMD Ryzen Threadripper 1900X 8-core CPU for 549 USD

Also

>64 lanes
>quad chanel
>no bullshit raid keys

File: 1491221061481.jpg (208KB, 1000x1000px) Image search: [Google]

208KB, 1000x1000px

>>61650785
It still made sense to go for a non X 1700 build IMO....
1700 went for $269. Mobo was $100 .. $369.

2 x 1700 = $538. So, i'd have to pay twice as much for a Threadripper 1900x.
Not going to get twice as much performance. I do like the higher clocks on threadripper. I'm assuming because they are using 2 dies... which is comforting because each die has its own Memory controller and 2 channels of DDR4.

So, I think a 1700 build is solid for testing out the Ryzen ecosystem.
If it checks out, I'd go straight to a 1950x for the x399.
I don't see the value at any other points.
I either need a hellavuh lot more compute and will pay the premium or I can be comfy at 1700 scale compute.

Way less cost and risk doing so.
Someone will say just buy a 1900x and x399 but i'd be paying double the price on the CPU and likely triple or quadruple the price on the mobo. Just doesn't work out for the same number of cores. I doubt 8 cores will be memory starved on dual channel DDR. So, there's that.

$369 for a mobo/cpu with all other parts being transferable to X399 is comfy

Extremely impressive clockspeeds there.

>>61650936
Two dies with 4 cores each. Lower heat density and thus can push them more.

File: 1495348928524s.jpg (3KB, 125x107px) Image search: [Google]

3KB, 125x107px

>>61650785
STOOOOOOOPPPPPPP

File: 1488693519789.png (84KB, 653x726px) Image search: [Google]

84KB, 653x726px

GLUED TOGETHER

>>61650929

You forgot the 64 pcie lanes, quad channel...

For example: I do not need a 12 or 16 core CPU, but I need multiple GPUs installed on the system and multiple NVMe disks.

>>61650785
the only one worth buying is the 16 core and 32 core, fuck off with this gay shit

Are threadrippers better for virtualization?

>>61651470
Falling for the PCIe lane meme and having no use for it....
I'm sorry but even a power user would be hard pressed to max out Dual channel and 24 PCIe Lanes .. x8/x8 + x4 M.2 and x4 chipset is more than enough. Dual GPU, M.2 NVME, and 32GB of dual channel ram.

Multiple GPUs .. more than 2? Multiple NVMe? for what exactly? No doubt, some people need this. However most don't and sending that all to 8 cores is a bit of a joke.

I don't feel most people are capable of evaluating what exactly their needs are and thus overspec because they have the money and can. It's your money. So, I have no comments to make. However, one can quickly look at the balance of things and tell one's gunning for an unbalanced system w/ bottlenecks thus can't be that serious about properly evaluating their needs.

If all you're doing is loading a shit ton of stuff into the GPU and letting it churn away, you don't need x16 much less x8 which is why miners run on 6 baby pci-e lanes and have 6 GPUs hanging off a basic bitch mobo/processor w/ 8GB of ram..

Multiple NVMe? As opposed to just getting larger capacity NVMe?
You're talking about a system that is clearly in the $2-3k range. Then you say (I have no need for a 12/16 core CPU).. LMFAO, then get a bit mining mobo that supports 6 GPUs with riser cards.

Again, I doubt people really understand their needs.

>>61650785
don't get threadripper 8 cores.
the cores will be on different dies.

File: 1501273837521.png (182KB, 326x327px) Image search: [Google]

182KB, 326x327px

>>61650785
>4ghz turbo

i-it's just as I imagined

>>61652414
Brainlets aren't thinking properly..
> Muh quad channel DDR4
> Muh cross-die hop to get to the other 2 channels
> Muh I need 64 PCI-E Lanes but will only feed it to 8 cores...

>>61650785
1950x + gtx 2080 when it comes out

Fuck yus

>>61652920
are you going to wait for 10 years to build your PC?

File: 1500139454809.jpg (75KB, 491x552px) Image search: [Google]

75KB, 491x552px

>>61652092
4 graphics cards for meme learning. Lots of RAM for a ZFS fileserver and NVME drives for caching. There are enough uses.

>200MHz XFR
MUH DICK

>>61652938
unless the call pascal 2 the 1180

>>61650785
>Call the ambulance! Intel is sick!

https://www.youtube.com/watch?v=4D8dzlgqEOo

>>61652092

>multiples GPUs? More than 2?

What is blender? What is octane? What is raytrace? What is deep learning?

>>61650785
no its not about the platform cost even if it was a 500 bucks more...
amd is simply filling the gaps so that intel couldnt find a way to launch a competitive product (HA! competitive...)in a lower price

File: tumblr_static_tumblr_static_bg48gusc0zk00kg4gwo808ccg_640.jpg (89KB, 640x360px) Image search: [Google]

89KB, 640x360px

>>61651430
oh yeah amd glue together dies
meanwhile in intel
https://www.intel.com/content/www/us/en/foundry/emib.html
An Elegant Interconnect for a More Civilized Age
Embedded Multi-die Interconnect Bridge (EMIB) is an elegant and cost-effective approach to in-package high density interconnect of heterogeneous chips.

>>61653685
Is that name seriously a reference to Star Wars?

>>61653705
you fail to see the wording

>An Elegant Interconnect for a More Civilized Age

they are so full of themselfs i bet no one at intel realised that ryzen has launched

File: 1495495069551.jpg (7KB, 200x200px) Image search: [Google]

7KB, 200x200px

>>61652977
mfw you need such a config to detect a hotdog because your algos/approach are shit-tier

>>61650785
Yes, I was hoping for a 64 PCIe lane X399 version of the 1800X.

>>61653745
>his algorithm needs 1.5-3 years of training to learn to walk even on advanced neural network hardware

>>61653747
>lane X399 version of the 1800X.
It's 1500x x2.

>>61653600
> What is an 8 core bottleneck
> What are multi-die hops for memory access
> What is, not needed x16 PCI-E to feed the GPU
> What is saying you're a professional who does such work but buys the cheapest processor in the stack for which a single GPU costs more than?
> $500 difference between the 8core and 16core processor but you go for the 8core yet are talking about 4 GPUs : $500x4 .. $2,000 and $1400 in RAM and $600 in NVME drives.

Sorry, I'm not buying it and how is everyone suddenly a professional render specialist and Deep learning guru? LOL

>>61652092
>Multiple NVMe? for what exactly?

It's cheaper and faster.

2x 512GB 960pro are 2x $279.99 = $559.98
1x 1TB 960pro is $632.10

File: 1491066625155.jpg (184KB, 800x980px) Image search: [Google]

184KB, 800x980px

>>61653775
> mfw bruteforce learning needs housefire hardware because no one stopped to think about how to fundamentally approach the problem.
> muh search and convergence algo
> Muh 2TB worth of hotdog pics on a ZFS file system
> converging converging ...

> mfw when brainlets call this AI

>>61653799
>> What is an 8 core bottleneck

>You don't need more than dual core ~ intel shill
>Wait, you need more than 8 cores ~ intel shill

Make up your damn mind.

>>61653825
silicon is the problem and considering how much graphene costs we probably wont get anything out of it for many years to come

>>61653800
NVMe seems fitting for either a main drive or a cache drive. For bulk storage, i'd expect SSD drives followed up by mechanical. It's called a memory hierarchy. Only brainlets fall for maxing out the first hierarchy of storage. They don't even do this in the enterprise...

Tbh, it seems a lot of people have no clue how to configure these enterprise grade systems. Thus, you're going to see all sorts of whacky configs, bottlenecks, and choke points therein.

>>61653832
Balance :
> 8 core, x8/x8 GPU + NVME + sata
> Dual channel

> HEDT
> 8 core x16/x16/x8/x8 + 3 NVME + sata
> Quad channel
> Multi-die hop to access PCI-E and DRAM hanging off other die

I'm guessing you brainlets didn't do any due diligence w.r.t to understanding NUMA architectures.

That non HEDT 8 core is likely going to kick the pants off your HEDT 8 core simply because all the cores are on the same die as is the DRAM/PCI-E lane access...

Enjoy the disappointment when you come to realize what NUMA is.

>>61653846
Shit inefficient algos require lots of compute power.
> But but...
> silicon is the problem

Keep waiting on that silicon Machine learning brainlet

>>61650785
>inb4 unlocking TR 1900X to 32C/64T
It's 2009 all over again!

>>61653891
oh yeah thats the problem.................

>>61653908
Not possible if TR is tied to 2 dies.
The other 2 dies literally lead to null traces on the motherboard.
There's no PCIE or DRAM traces to feed them.

>>61653914
I guess they don't teach you the fundamentals behind things anymore....

https://en.wikipedia.org/wiki/Stochastic_optimization
Brute force statistical optimization...

> Muh silicon

>>61653929
Aren't all 4 dies tied together over IF? I'd like to think the other two dies can access resources over Infinity Fabric. Doesn't give you additional DRAM channels or PCIe lanes, but does give you moar coars.

The performance/latency would be shit, but hey it's two extra dies for free.

>>61653908
It only has 2 dies and 2 spacers. The only time it will have more is if after binding 2 dies, one of them is a dud (less than 4 working cores). At best you might unlock 2 to 4 more cores.

>>61653794
There better be something in that big ass box besides a processor and a mounting harness.

>>61653882
2 quad core CPUs were a thing just a few years ago.

>>61653988
4 ? TR doesn't have 4 active dies. Only EPYC does which is why it has 128 PCI-E lanes and octa-channel.

TR has 2 dies. In order to get 64 PCI-E lanes, you need 2 active dies. Thus, you're in a proper NUMA config.
- 32 PCI-E lanes and 2 DDR channels feed one die
- 32 PCI-E lanes and 2 DDR channels feed the other die

They are connected via IF. There is extra latency to cross dies. Going multi-die 8 core vs. sticking to one die 8 core is not proper scaling.

It's comments like yours that reflect that a number of people have no clue what their buying thus can't possible be serious about (performance) and scaling.

>>61654105
They were... now you have single socket/die 8 core. Going 8 core split die w/o carefully considering what exactly your work load consists of means you're not serious about why you are doing it.

If someone told me they were going 16core HEDT threadripper to take advantage of the higher scaling, I'd understand.

>>61654156
>4 cores windows OS with 16GB
>4 cores linux VM with 16GB

Nice and comfy.

August 10 confirmed.

I would like to know what uses more than 1-2 cores outside of video editing that average user uses

>>61654990
The average user is better off just using a phone. Nobody besides gamers and power users is building their own PC.

File: 1489100922662.png (246KB, 882x758px) Image search: [Google]

246KB, 882x758px

>>61650785
I CAN'T TAKE THIS ANYMOOOOOOOOOOOORE

>>61655049
now imagine a phone with a downclocked and undervolted ryzen

>>61654990
Every thread
Every single thread

If all you need is 1 or 2 cores then buy a 40 dollar Pentium and fuck off.
YOU DO NOT BELONG HERE.

Better yet, call Intel, tell them they can stop making everything other than Pentiums. Turns out all that r&d was pointless, all anyone anywhere will ever need in all of human existence is 1 or 2 cores so they can just stop that as well. I'm sure they will happily oblige and watch AMD crash and burn in their futile efforts to increase core counts and lower prices and tdp while casuals, gamers, HEDT enthusiasts and professionals alike flock to buy Intel's bargain priced 800$ pentiums, while laughing at AMD'S vain efforts.

Your just so smart aren't you, you fucking moron.

>>61652463
Henkuma ;∆

I just wish it were actually useful over a 1700 for people who aren't running a business with them and consider the price a negligible investment.

>>61655508

You're*

>>61654990
Google Chrome/Chromium
they spawn a large number of processes for sandboxing purposes.
Firefox is now beginning to go multicore too.

If you listen to music while using word processing while running background programs (like BitTorrent) etc, you can also benefit from more cores.

What's the advantage of the 1900x Kikeripper compared to the Ryzen 1800x?

>>61655918
Moar memory.
Moar lanes.

>>61655918
4x the pcie lanes of a non HEDT 8c/16t (64 vs 16)
Quad channel memory

Thats really about it. Well worth it if you have need of many people based devices. I'm personally waiting on benchmarks of the 16c/32t Threadripper before I jump.

>16c all at 4.0ghz

One can dream. Wonder what kind of cooler I'll need?

File: 1349260135801.jpg (17KB, 196x196px) Image search: [Google]

17KB, 196x196px

>>61654990
cinebench

>>61654156
ok where to start

first things first, for the most part quad channel vs dual channel is mostly a meme, the difference is there on paper and some work loads, but the speed is really only beneficial on tasks that a server would be better for.

now 8 core, 4 cores per die being slow? yes, in gaming it would likely shit itself with inter die latency, however most applications where you would be considering a high end desktop as an option over the consumer don't rely on inter core communication so much, these applications are the ones that show off scaling per core better then anything else.

now pcie lanes, if you are rendering 3d, this isn't a question, if cuda accelerates it 4 980's would be 8000 cores, and cost 5-600$
4 980ti's would be 10000 cores for 1000$

now to get even close to 8000 you would need 4 1070's, and does that even work? if it does, you need to spend 1200$ on it, and to break 10000 you would be spending over 2000$, now if you can't do more then 2 way, to even approach 8000 cores you would need to spend 2400$

Now let's say your workload needed high amounts of ram, which 3d rendering can demand, thread ripper would allow for a cheap solution for 128gb of ram

then you have storage, while I would personally wait till nvme got larger solutions before I get another, if you need the space now, you can use it now, then you raid 0 them all and you effectively have a near 10gbps read speed, which is orgasmic for video work, which again, is better offloaded to the gpu then done on cpu.

>>61654140
someone deleted a thread ripper and it had 4 dies on it, if this is going to be the norm or if this was just an engineering sample we don't know.

>>61655872
>Firefox is now beginning to go multicore too
The future is multi-core.
AMD was right all along.

File: core_i9.png (571KB, 1900x972px) Image search: [Google]

571KB, 1900x972px

>>61650785
Jokes on you, I'm getting an i9!

pic related

File: 1457294956960.jpg (177KB, 801x1500px) Image search: [Google]

177KB, 801x1500px

>>61656400

>>61653794
Stay mad boxlets.

>>61650785

>$100 for 0.2ghz
>can easily be OCd

>>61656087
> quad vs dual a meme except for specific work loads
agreed

> If application doesn't require intra core communication you're fine
Many do and I question whether the large number of brainlets interested in this platform @ 8 cores are not going to be using it in cases that require this. They are already having issues in cases on Ryzen where processes end up on another CCX complex. So, for the professional who knows what this platform is and why their workload will work good on it, they're fine. For the brainlet getting this vs. ryzen 8 core because it has more PCI-E lanes and DDR slots, they're fucked and in for a surprise.

> now pcie lanes, if you are rendering 3d, this isn't a question,
Explain to me how this works. Is the CPU maxed out at the same time the GPU is for GPU bound rendering? No.. you're simply piping crap over to the GPU, letting it run with little traffic back and forth. Thus why miners use shit tier processors/8GB ram and PCIE x2 attached to USb cables to PCIE 3.0 x16 cards... you ship the task over to the GPU and run it. You don't need x16 for that. Correct me if i'm wrong.

> Now let's say your workload needed high amounts of ram, which 3d rendering can demand, thread ripper would allow for a cheap solution for 128gb of ram
Now were getting somewhere. But why would you need high amounts of ram for GPU bounded rendering? You use GPU memory no? Give me some good source sauce on what GPU rendering traffic looks like between the CPUs/GPUs and ram.

> Raid 0 nvme
Please stop w/ this bullshit

> which again, is better offloaded to the gpu then done on cpu.
Take a tip and hint from miners for GPU bounded workloads.. You don't need big ass communication channels to the GPUs. If your CPU isn't doing that much work neither do you need shit tons of RAM.

Again, correct me if i'm wrong here.. maybe i'll learning something i didn't know

Serious question, why get Ryzen when a 7700k still outperforms them in video games?

>>61656654
Non vidya related tasks.
Why do people buy platforms that are not suited for their particular use case? Brainlets spend little time researching/understanding the nuisances. They see a spec sheet and get hyped by an e-celeb and are sold.

That being said, there also is future proof. Vidya is going to get tuned past quadcores in due time.

>>61656580
Thread brapper's appeal is the platform and lanes.

>>61656716
Miners solved the lane issue with understanding that nothing uses anywhere near x16 worth of lane traffic and thus went for boards with x4 lanes and risers to x16 capable cards. I'm wondering when so-called professional brainlets with GPU bound compute tasks are going to catch up to this revelation. If you're getting threadripper for moar x16 lanes to GPUs for basic bitch rendering, you're a brainlet.

>>61656943
And when the GPU local memory is maxed out and it has to constantly push and pull from main system memory what then?

File: 1499951764422.png (359KB, 1300x599px) Image search: [Google]

359KB, 1300x599px

It wasn't supposed to end like this, not like this...

>>61653794
Just drop it on the son of a bitch.

>>61653882
Skylake-X/Purley Xeon's mesh needs NUMA optimization as well, stop pretending this is somehow an AMD only problem.

>>61656654
Why are you even posting this in an HEDT thread?

File: Ryzen7_7700k_1080ti.png (2MB, 1920x1440px) Image search: [Google]

2MB, 1920x1440px

>>61656654
>Serious question, why get Ryzen when a 7700k still outperforms them in video games?
Not with 3600mhz ram, it doesn't.

>>61656623
Most applications you are going to want more pcie for over more cores is not going to hammer the cpu the same way lets say a video game does where it hammer's thread 0, and then uses some of 1-12,if you send it the task of rendering something out, it's handled on core,its why these loads can sku more toward multicore loads than single core.but I said before,you are likely going to be able to offload it to something like cuda so the pcie is beneficial there.
Not going to lie with rendering,not sure how necessary it is to have it at 16x,however I do know there are real time demos that use 48gb of vram so its possible data is piped to the gpu.
Now on another note,there are also physics engines and such,that im not sure if they accelerate with gpus,as you set the detail in them stupid high and bake the physics.Then there are programs like zbrush, something thats damn near the only program of its kind,that uses the cpu exclusively I believe, so there is merit there of having a good cpu,that and more ram that quad allows you would let you keep more data up at a time.
for 3d modeling,unless you are working for optimization, you can easily end up with 16+ textures that are all between 512 to 32k all in incompressible formats, for each and every object. that adds up fairly fast, then you also have the model itself which contains the data for every single point, and unless you optimized everything, is likely 1 million or there around polys and that itself is fairly large.
and with video editing, to load large videos hundreds of gb for 30~ minutes, fast storage is king, which is why nvme raid 0 is a thing, you likely won't have ram to house the videos and while editing, jumping around is going to happen a lot, and then not everything in the process is gpu accelerated, so it's a mixed workload, but generally the gpu does very heavy lifting here. Now streaming, I know cpu >>>>>> gpu for quality per bit, but im unsure if thats also the case for properly rendered video

>>61656943
That works when you are sending small data to the gpu for the gpu to crush then send the small result back.

If you are sending big data regularly its not an option that makes sense.

>>61658715
Outperforming ... At 5 Ghz ...

Just add the Luck + the money for cooling solutions against such a house fire, add the power consumption,

et vólia

this is why AMD Ryzen is beating Intel all the way up the arse

>>61658715
Clearly this is a GPU bottleneck. You don't do CPU comparisons on Ultra graphic settings.

>>61656654
because I have shit constantly processing shit, and when 20% of an intel cpu is used, that's 20% shaved off all max frame rates and shitting the bed on small ones,

However that 20% on intel becomes 10% on ryzen, and then running a game it never even wants to touch the core that's using 10% anyway so my gameplay is un fucked with.

>>61658813
should note that amd is also overclocked, but it effects the processor less then what it takes to get 5ghz out of intel.

I was going to make another thread for this but whatever.
The 1900x is $550 and comes with quad channel memory support and hopefully supports higher frequencies as well.
Infinity Fabric is reliant on memory frequency, it's set to half your RAM's speed.
We know ryzen is amazing at 60hz but the 7700k is still superior for 144hz games and when you get your FPS that high you generally run into memory speed as a limiting factor. Upgrading to 4k ddr4 was huge for me on my 6700k 144hz machine.

How do you /g/uys think a 4.2ghz(they said it does that stock with xfr) 8 core ryzen with quad channel memory will perform for high refresh gaming? I think it's going to outperform the 1700/1700x/1800x by a fair bit and those pci-e lanes are fukkin juicy

>>61658838
Thats right. But the 5 Ghz on intel is much harder to obtain then 3.9 Ghz on a ryzen 7. Also, the ryzen wont heat up that much, and the intel will only be able to withstand this heat when delitted.

>>61654442
must be opposite day

>>61658819
>You don't do CPU comparisons on Ultra graphic settings.
Ultra adds additional CPU overhead, dipshit. Stop with this fucking "hurr durr benchmark at 720p low" meme shit.

>>61658819
>1080 Ti
>1080p
>GPU Bottleneck
?

I think it's too early to jump on the octacore bandwagon

An overclocked i7 (even an i5) will shit all over Ryzen for another couple of years yet

Obviously "media creators" are a different case

>>61658989
>shit all over
>20 extra FPS in GTA V
>at well over 100 FPS
>at 1080p with a 1080 ti

Fuck off.

File: Ryzen7_7700k_1080ti_2.png (2MB, 1919x1439px) Image search: [Google]

2MB, 1919x1439px

>>61658989
>An overclocked i7 (even an i5) will shit all over Ryzen for another couple of years yet
R7 with 3600mhz ram basically ties a 5ghz 7700k in Triple A games.

>MUH BATTLENAWCKZ!
The test was done with a 1080 Ti

>>61659006

i5 7400 outperforms all the Ryzens GTA V 4k

>>61653799

For God sake, are you retard or something?

File: 14958035970qy2jlakgl_7_1.png (164KB, 630x372px) Image search: [Google]

164KB, 630x372px

>>61659053
There is no difference at 4k, you're completely GPU bottlenecked at that point.

>>61653799
Epyc was already tested and kicked serious ass, and TR is basically the same, so I'm afraid you'll have to put your FUD to rest. http://www.anandtech.com/show/11544/intel-skylake-ep-vs-amd-epyc-7000-cpu-battle-of-the-decade

>>61658977
Yes, every GPU can become a bottleneck on every resolution if the CPUs is strong enough.

>>61659127
>Yes, every GPU can become a bottleneck on every resolution if the CPUs is strong enough.

No need for a 7700k then, since even at 1080p on a 1080 Ti it's GPU bottlenecked anyway, and only a fucking retard would buy a 1080 Ti for 1080p...

>>61659159
>No need for a 7700k
Yes, you don't need a fucking i7 for video games. Get a 4 thread CPU like a Ryzen 3 instead and spend more on the GPU instead
>only a fucking retard would buy a 1080 Ti for 1080p
You need one if you want those FPS numbers on Ultra.

The server version of this is going to be a huge winner for bioinformatics / HPC applications.

>dat I/O throughput

>>61658819
>1080ti
The quality of your posts is bottlenecked by your IQ.

>>61659306
The 1080Ti is not nearly as fast as you believe. Next generations midrange cards will be as fast as the 1080Ti just like how the 1070 is as fast as the 980Ti.

>>61651470
>You forgot the 64 pcie lanes, quad channel...

>For example: I do not need a 12 or 16 core CPU, but I need multiple GPUs installed on the system and multiple NVMe disks.
This is why I considered ryzens as complete shit. I'm glad that they actually remidied thier offerings, and now amd has viable alternatives over intel's hedt, xeons.

>>61652092
>Falling for the PCIe lane meme
>No doubt, some people need this.
If people need it then it's not just a fucking meme you dipshit.
>However most don't
This is for the people that need/want it.
Most people could get by with just a pentium today for what they actually do on a pc shit maybe even just a c2d, or a mobile cpu. This cpu is meant for the people that need the extra features, and it's not too shabby for those that do not.

I'll be buying an ehhpick if all the threadriper mobos are faggy gamer trash

>>61658778
> demos that use 48gb of vram so its possible data is piped to the gpu.
So, given that no GPU or combo therein of 8GB variants have that capacity then you're saying it is paged in in blocks most likely (big chunks). Still, instead of one big pre-load what is this 1 or 3 loads? that's trivial in the grand scheme of the compute time spent on the GPU and its trivial w.r.t to PCI-E x4,x8,x16 bandwidth .. x4 = 3.94GB/s. So, literally seconds to transfer the data chunk.. minutes and hours to render...

> Extended commentary
My issue with this is.. Has anyone stopped to do an honest and detailed review/detailing about how to configure such a system properly to avoid bottlenecks. I did some searching and I found someone asking the obvious question on a blender forum : Why not use a mining motherboard for GPU bound rendering? Meanwhile I keep running into people yapping about more PCIE lanes... Do you amateur video guys know how to spec a non-bottle necked compute stack for your tasks or do you just throw money at random components and guess performance therein? RAID SSD arrays exists for a reason. It's far cheaper and more effective to run a distributed storage array around cheaper SSD drives than to be goofing around with NVME motherboard raid solutions.

I can undertand the professional video/3d modeling professional.. It just seems like every random jackass is a 3d modeler/video editor all of a sudden and claim to be doing it at 4k/8k and the upper reaches of the industry standard.

Color me skeptical as fuck. I'm going to do more research on this as I'm curious about the data/compute access patterns of such task. Thanks for the detail you provided. It will serve as very good pointers.

>>61658799
Big data regularly.. what workflow task does this? Furthermore PCIe x4 is 3.89GB/s. You have 8GB of Vram. WTF does x16 do for you? nothing.
SSDs pull ~500MB/s. So, you're nowhere near even saturating this if you're pulling assets off a drive. I seriously want to understand this. So, if you have pointers to someone actually detailing the technical aspects of these workflows I'd appreciate it. No offense but it seems like you guys are pulling shit out your ass and actually have no idea how the data access/flows work.

>>61653794
Fucking boxlets, when will they learn?

>>61658849
Use your brain.
1900x is composed of two dies.
The L1/L2/L3 of a 1900x is the same as an 1800x single die.

The difference is :
The 1900x has 4 cores on each die. Thus, you have to cross the die + infinity fabric to communicate to a process/thread.

4 channel DDR is a meme for most workflows. There are tons of real-world performance tests that show 4 channel DDR doesn't change performance vs Dual channel.

Furthermore, you're not grasping that there are 2 channels per Die in a 1900x. Thus, if you're accessing memory on a particular stick and a particular core needs it on another die, you have to fetch and send it across from another die.

So, for most brainlet tasks, it's going to be worse than a 1700.

> those pci-e lanes are fukkin juicy
You likely wont have a use for them and again the PCI-E lanes are split 32/32 across two dies.

> Mfw brainlets have no understanding of true NUMA architectures and tradeoffs therein.

>>61659667
Nobody is spending $1000 on a CPU without considering if it's worth it for them.

File: 1500793659657.png (7KB, 224x239px) Image search: [Google]

7KB, 224x239px

>>61659667
Lol you okay buddy?

>>61659106
>http://www.anandtech.com/show/11544/intel-skylake-ep-vs-amd-epyc-7000-cpu-battle-of-the-decade

Kicked serious enterprise workflow task Ass...
Not brainlet vidya games nor brainlet configured HEDT setups for video editing.
All of the benchmark tests were done on server workloads which are nothing more than scaled indepdent process execution across large data pools.

Are you running a database/enterprise web server brainlet?
These are not vidya games.

>>61653685
>An Elegant Interconnect for a More Civilized Age

Holy... Fucking... Shit...

I thought this nigger was making a joke, but that's literally in their marketing materials. They have shills posing on this board about glued together CPU's while they are frantically attempting to put something similar together so that they can hope to compete.

>>61659746
>FrankenCPUs
http://www.pcworld.com/article/3185875/hardware/future-intel-cpus-could-be-cobbled-together-using-different-parts.html

>>61659254
It's quite clear that most of the brainlets harping about this platform know fuck all about configuring a balanced system or what is bottle necking their performance. I'm almost tempted to cancel my pre-order so I can wait a week or two and buy a threadripper setup off a dumbass who thinks it will increase their FPS only to find it plummet due to NUMA access downsides.

>>61659689
kek.

>>61659719
Market research. Pushing the envelope gets you to the heart of the matter quicker. Results so far are that a slew of brainlets are going to be buying this platform .. That understanding will be quite valuable in the times ahead.

>>61659765

Intel is trotting this shit out there acting like AMD doesn't exist. I can't wait to see some Vega APUs if they utilize IF.

>>61659346
u

>>61659746
>AEIMCA
Are they about to pull the patriot card?

>>61659558
it was useing 4 titans if I remember correctly.

now on raid ssds, to match a nvme, and lets just use samsung and the 3200 and any other ssd and their 550mb claim as fact to make it easier, you would need 6 ssds

you can either buy a 1tb ssd for 250$ get 6 of them for nvme matching, and spend 1500$ to match one nvme in speed, and get 6tb, or you could get 6 500mb nvme drives, and raid them for a theoretical 19,200 mbps read

Now I know that there is a limit to useful ness here, and likely other bottlenecks, but if you don't require more than 3tb for your videos, this seems like the better option.

well, there are youtubers, or people who aspire to be one, or people who make short movies of questionable quality. cell phones are able to take 4k video, not good 4k video, but 4k video and a pro quality 4k camera costs 3000$ so its well within an amateur's budget. as for 3d modeling, there are a fuck load of artists who use 2d programs, just liek there are a fuck load who want to use 3d, and I will tell you now, dropping 2-10k for art tools that will let you make unlimited projects for the duration of their life is honestly a bargain. in the realm of 3d, 4-32k is kind of the norm, with many mods for games dropping in 8k textures because when you lick a wall, you don't want to see pixels, same with making the model, usually when it comes to final render time for a movie, a studio will trim textures to the correct size that even at the most zoomed in they get on screen is never showing pixels because if they didn't do this a movies workload would be multiple petabytes larger than it currently is.

as for 3d modelers and what's better, for the most part a pro has a workstation system, and until relatively recently, gpu renders were not viable, some of the industry standard ones still don't use gpu acceleration, but their use case is largely render farms, not someone's home builds.

>>61659351
well AMD was focused on budget and in order to reduce the cost and therefore selling prices of the chips and the motherboard they had to strip anything extra than regular consumer dipshits would never use.

Be real man. Most normies want a vidya gaymer build or a machine to edit youtube videos.

I give AMD credit. They make a cheap bore bones platform for noemire consumers thats cheaper than intel for not all that far behind performance and then they come out with a workstation/server platform with all the trimmings for power users that destroys the overpriced and locked down garbage intel offers in the same category.

>>61652977
>ZFS fileserver
>doesn't use ultra low power parts like Xeon D
KEK

>>61658852
>but it effects the processor less then what it takes to get 5ghz out of intel

well aware, but just saying both are overclocks, and both are fairly heavy ones

amd relies more on silicon lottery to get there, intel relies more on being willing to void the warranty along with a bit of silicon lottery and a great cooler.

>>61653600
>he doesn't use clusters for deep learning

>>61650785
I've been hoping for this, 8C should really be enough for me, but having enough PCIe lanes for x16/x16 dual-GPU is nice at 4K, it can make a difference in certain games or engines (like UE4). 1920X at $800 was a bit too pricey considering I had no need for the cores, but $550 I'm willing to part with if it performs well.

>>61658715
BUT MUH 5GHz

>>61660106
>using xeon dicks
>instead of real horsepower
>being a cachelet

>>61659803
This shouldn't be a revelation. I'd love to know exactly how many morons bought into intels' HEDT over the years strictly for gaymen.
I watched this video from paulshardware where he was upgrading his main PC and he did benchmarks on a bunch of CPUs and even though the 7700/7740x was the best he still said "i'll probably go with the 7900x because it feels right". It's just like the people that buy titans.
People love to throw money at shit just because it's expensive.

>>61660150
>just slap cluster on top lol
best way to turn your problem into several distributed ones

>>61659667
>The 1900x has 4 cores on each die. Thus, you have to cross the die + infinity fabric to communicate to a process/thread.
Do you have any evidence for how large the latency penalty is for crossing dies vs. crossing CCXs? I'm curious.

>>61660395

>Solving problems using Regular Expressions:
>Now you have two problems.

File: 1459884500883.jpg (21KB, 259x300px) Image search: [Google]

21KB, 259x300px

>>61660451
>hack together some Perl code
>can't understand any of it two weeks later

>>61653685
>heterogeneous chips
so now intel will play a Frankenstein to be able to compete?

>>61661368
Needs to be complete with lightning and JUSTnich yelling "It's alive!"

>>61660578
just make a ProblemFactory in Java, that always helps

>>61656623
>Correct me if i'm wrong.
No idea if 3D rendering requires this, but there are GPU workloads which require GPU-to-GPU communication and plenty of bandwidth. Take a wild guess what bus is used for that if you're not using some NVIDIA HPC cards with NVLink? Just because PCIe bus bandwidth is irrelevant in mining doesn't make it true for compute universally. Fuck, you don't even need compute, just CF and SLI at high enough resolution need the bandwidth to run games.

>>61661501

Don't you mean an Abstract Problem Factory?

File: 87676.png (37KB, 650x300px) Image search: [Google]

37KB, 650x300px

>>61659053
With horrible stuttering thanks to shit frametimes

File: 87733.png (35KB, 650x300px) Image search: [Google]

35KB, 650x300px

>>61658989
The 7700k can't even maintain 30fps in Civ 6

>>61660412
STH did some tests on naples and there are five distinct latency zones on a 2p system. So a TR is going to have three.
>local l3
>adjacent l3
>die jump l3
The die jump was around 20% more than the adjacent snoop, not much but it's there.

>>61660090
I just want some real barebones shit though, a single zepplin die with exverything exposed would make an immense mini server and probably a lot of other things. am4 was almost a mistake!

Ryzen is good but right now I will most performance across the majority of titles with an overclocked 7700k

also I only build mITX

>>61660055
This is the kind of detail I was looking for desu.
I sincerely thank you for spending the time necessary to write this out and detail this for me. Sounds fair enough. Upon reflection, this is moreso indicative of a higher common standard that is being forged that has come to value the hardware costs associated with such ventures.

Also, thinking more deeply.. True enough : content is now and forever distributed as opposed to centralized thus there are way more content producers doing high production level media.

*shakes head in approval.
Now i'm off to understand the data flows of such workflows a little bit more. Thank you for this bread again anon.

>>61662399

They didn't overclock the 7700k in those benchmarks

>>61660412
What this anon said right here :
>>61662599
Source sauce :
https://www.servethehome.com/amd-epyc-infinity-fabric-latency-ddr4-2400-v-2666-a-snapshot/
Chart porn details a significant range of latency numbers.
You shoud be looking at the Yellow to green colored fields within a column only as there are only 2 dies and one socket for threadripper.

For the lazy ...
@DDR4 2666 :
> 80ns idle latency for accessing DDR connected to a die
> 137ns idle latency for accessing DDR connected to another die

Memory bandwidth for access from DDR connected to die :
> 36346 MB/s
Memory bandwidth for access of DDR connected to other die :
> 19103 MB/s

Each die appears and behaves as a numa node
> Essentially a 2nd socket CPU with latency and bandwidth reductions therein

For workloads with a lot of inter-process/thread communication, an 8 core (2 die) threadripper is not going to beat an 1800x single die. This is just common sense but now you have the undeniable proof.

Like i said, threadripper is moreso if you need the increased core count to scale with the increased DDR / PCI-E lanes. Putting an 8core on threadripper will be unbalanced for certain workloads. Thus, you're going to really have to know your workload and computer architecture. Otherwise, you're spending in the blind.

>>61662388
>7700K is better than the 7740X
How did Intel fuck up X299 SO BADLY?

>>61663479
So just as an example, this would probably do better for serving requests of some kind rather than multithreaded rendering yes?

>>61663651
Not that anon, but rendering doesn't tend to be so incoherent in what it wants from processors. It's probably one of it's strong points. Systems which demand all sorts of random snooping and unpredictable instructions aren't all that common, video games and in certain situations databases are some examples.

>>61663178
They also didn't want to risk burning their setup too.

>>61663651
Yeah, you have to do some leg work here.
I'm actually doing that now because I'm curious. But yeah, check out various forums and such related to your particular workflow and i'm sure it's been discussed in detail.

I gave a confirmed point to reflect on.
This anon added a little bit more detail:
>>61663814
So, it's time to hit google and research this a bit to get your system spec'd out right.

File: ping-amd.png (159KB, 602x477px) Image search: [Google]

159KB, 602x477px

>>61663479
>You shoud be looking at the Yellow to green colored fields within a column only as there are only 2 dies and one socket for threadripper.
Okay, but how does that square with testing showing similar ~140ns idle latency for accessing cores on the other CCX on an 1800x?

>>61664290
because the infinity fabric on the consumer ryzens is running at half RAM speed, while on TR and EPYC it's running at full speed

>>61664290
>>61664708
and that paired with quad/octa channel RAM on TR/EPYC, from the dual channel on Ryzen also does things to the speed/latency

>>61664290
This latency between cores is the reason Intel CPUs with bingbus and only four cores does better in games right?
And the reason Intel's CPUs with funnymesh like the Skylake-X stuff also generally has a lower average.

>>61664768
intel has some 10% IPC advantage, as well as higher clockspeeds

>>61664818
Is this in certain loads or absolutely?
Did anyone ever do a review with a 7700K and a 1800X, both at say 4GHz?
You know something like single core and 4 core 8 threads?

>>61664888
I don't have the graphs, but yes, these tests happened, and the 1800X with 4c/8t @4GHz was ~10% behind the 7700k @ 4GHz

>>61657225
>>61656500
>>61656400
what is this meme, the threadrippers literally have 30% higher tdp.
i mean, jesus fucking christ, 180W based tdp before any overclocking did you even read the release specs. This thing is a crime a against humanity

File: TDP.png (14KB, 678x181px) Image search: [Google]

14KB, 678x181px

>>61665085
>muh on paper TDP

>>61664983
I shall have to search for that.
On several different kinds of loads?

>>61665116
video transcoding and a synthetic prime number calculation if my memory doesn't fail, but Ithink they tested it with more things too

>>61665114
>threadripper totally won't have paper TDP
i'm building PCs for 20 years now and AMD has always been the king of electric heating solutions

>>61665193
>Netburst

>>61665175
Thanks guy. I do care about the video stuff.
Infact I have no idea why no reviewer tests with Sony Movie Studio/Magix Begas.
That's what all the fucking normies use.

>>61665193
That image is real tested figures, you retarded kike.
Threadripper pulls less power underload while having higher performance. Intel's Xeons can't compete.

>>61663099
>mini-itx
>a cpu that can't be properly cooled even on big cases with watercooling
have fun

>>61663161
I remember your writing style. Still worried about the dead traces on threadripper mobos?

>>61659031
dude delete this stop the antijudaism

File: aHR0cDovL21lZGlhLmJlc3RvZm1pY3JvLmNvbS9MLzEvNjg1MDQ1L29yaWdpbmFsLzEyLVRlbXBlcmF0dXJlLURpZmZlcmVuY2VzLnBuZw==.png (21KB, 711x533px) Image search: [Google]

21KB, 711x533px

>>61665085
Protip: The TDP isn't the reason Skylake-X is a fucking housefire.

http://www.tomshardware.com/reviews/-intel-skylake-x-overclocking-thermal-issues,5117.html

>>61659031
Hah. That switcheroo in the Overlook lows.

>>61665436
true, they've also done a miracle in finding something worse than air for distributing heat from cores to cap

>>61652092

Nigger I'm waiting on this. I have 4x 960pro 1TB drives waiting and two Vega cards that I've been saving up for. Can't wait to cram all that sexy stuff together with 64gb ram and a 1950x

>>61655797

Kekd. He is right tho. You are just an idiot.

Also BF1 runs better on 8core VS 4core.

>>61658819

I run all my games at ultra 1440p. I need to know how everything performs at 1440p. I don't give a Shit If a cpu does 500fps in cs go 720p.

>>61664738
the on die latency is all ram speed based, not ram throughput, you could have octa channel memory and the cpu would still be bottlenecked by the speed of the ram, not the max throughput.

>>61652092
640k should be enough for anyone!

>>61664818
the ipc advantage is debatable, with proper completing amd should beat out intel, as they get 5 things per cycle opposed to intel's 4, but intel does have a massive legacy catalogue, will have a large catalogue going forward, and does have a clock speed advantage till zen 2, which is suppose to be 5ghz (keep in mind, the current process targets 3ghz but rzen is 3.2-3.8 base clocks, if 5ghz has that same room zen 2 could push 6ghz)

>>61666185
Decode is not everyting.
But yes, Zen is 5-wide compared to Core being 4-wide.

Anyone have a comparison of the $550 Ryzen vs a similar performance intel chip? Is intel offering the same sort of thing in 8 cores but for $1000?

>>61653794
is that a fucking CRT

>>61667229
the 8 core is basically a 1800x with higher base clocks and 64 lanes, quad channel RAM

Too much talk over nothing

Gaymers don't need moar cores

The split die will have drawbacks

File: Mysterious Merchant.gif (4KB, 452x523px) Image search: [Google]

4KB, 452x523px

>>61667541

>>61667558
64 lanes

You gonna run four 300 watt Vega's over two kilowatt PSU? Mmmmkaaay

File: ssg.jpg (537KB, 1920x1080px) Image search: [Google]

537KB, 1920x1080px

>>61667588
4 of these bad boys make the ultimate workstation.

>>61667598
>8TB GPU memory
epyc

>>61667659
>literally THE ONLY product in the world that is able to hand raw 8k playback
Hilarious.

>>61667679
b-but muh hardware key from RED

>>61667706
RED itself shills SSG.

>>61650785
0.2 ghz over base ryzen
Same shit that cant past 4.0 ghz
You think x399 boards are gonna be cheap? And a cooler designed for this bizarre chip?

Just because AMD designed a cpu with 2 dies disabled so you can pay a lot more on useless pins

>>61667706
RED is an expensive, proprietary nightmare of a company

>>61653600
Its okay we know you are a neet and dont do any of this

does having more cores do either of these:

speed up extraction time of compressed files
increase speed of video/audio muxing

>>61667918
Yes and yes.

File: intel ceo JUST.jpg (119KB, 1000x1000px) Image search: [Google]

119KB, 1000x1000px

>>61667771
Give it up, Brian. Nobody's paying $1k for a 10 core that has worse performance and can't OC w/o burning down your house.

>>61667541
>>The split die will have drawbacks
Examples? So far as I can tell, the latency between dies appears to be similar to the latency between CCX's on Ryzen.

>>61668488
It's +30ns more when going off-die.

>>61668501
Where are you seeing that?

>>61663479 shows ~80ns on die and ~140ns across dies for Threadripper, while >>61664290 shows ~40ns intra-CCX and ~140ns inter-CCX. This would appear to make inter-CCX hops identical in latency to inter-die hops, unless I'm misunderstanding something?

>>61668592
>one bench is from STH aka THE server dudes
>the other is from fucking pcper
You know what to trust.