Neural net predictive action at work.

File: C57KuNLWQAAvDRW.jpg (27KB, 450x346px) Image search: [Google]

27KB, 450x346px

>>59235859
Intel.

>>59235859
magic?

>>59236085
Nah, just becomes smarter as the workload increases, I can see this be very useful in Big Iron where they rerun the same shit over and over and over again.

>>59235859
Is it useful for mining crypto currency?

>>59236085

It's that "machine learning" people dismissed at the new horizon event.

CAN I RUN """"GAMES"""" 5 TIMES AND GO FASTER

File: 2dd.jpg (15KB, 300x300px) Image search: [Google]

15KB, 300x300px

>>59235859
>never turn computer off
>it becomes the most powerful machine on earth

>>59236155

No.

So I gotta grind my cpu now? Wtf

Wait so Zen's branch prediction is based on neural networks? That's pretty fucking cool.

>>59236161
Maybe some parts of the game, but this depends a lot on predicting future events, and games being interactive this will be difficult.
Rendering like the OP has a linear workload that's much easier to predict.

>>59235859
>people are actually retarded enough to fall for this bait

>graph starts at 1610 and ends at 1640
Nvidia tier.

>>59235859
So it actually works?

Source articles?

Actually interesting stuff. I didn't think it would do much. The fact it does do anything means we'll see Intel copy it later.

>>59235859
It's a small boost, only a few points but it is consistent, this will be a nice boost to things like SAP and other analysis software that run 24/7

>>59236218

A 1% difference.

It works, the qurstion is, who cares?

File: 453Di.jpg (176KB, 913x1185px) Image search: [Google]

176KB, 913x1185px

>>59235859
>tfw Skynet will emerge from AMD

>>59235869
>it gets worse
JUST.

>>59235859
IT'S LEARNING
SOON IT WILL JUDGE HUMANS TO BE INEFFICIENT AND WE WILL WAGE THE BIGGEST WAR OF ALL TIME

>>59236161
Yes, but you have to beat a game 5 times to make it jump from 110fps to 115fps.

>>59236243
>Who cares
anyone who uses the PC for everything other than gayming.

>>59236243
>It works, the qurstion is, who cares?

It can be improved in the future. Both by AMD, and by programmers once it's better understood.

The fact it's even 1% is like that fucking reactionless thruster. If you told me they'd have even a working attempt at this by 2017 I'd have laughed.

>>59236277

Finally, I can live my life as a battery playing out in a virtual world and escape the tedium.

>>59235859
Pretty much a server feature, like most of the stuff in Zen.

It's neat though, I heard AMD talking about it but I didn't think it would be so consistent.

>>59236197
with actual long workload it can escalate to saving hour or two

>>59236293

Enjoy your 1% improvement.

>>59236294

"Programmers" won't be able to access it, since it's low-level encoded at the hardware.

And even then it has limitations of what it can do because of the hardware limitations. You're not going to run the same test another six times and get another 1% improvement.

So it's not exactly physics defying like an EM Drive.

>>59236317
everyone disregarded it as marketing fluff
it actually works though, huh

>>59236331

An hour or two off of a 3 day project.

>>59236375
you can fit another 3 day project per year with that one hour cut

>>59236164

Thats my plan.

I can see where this is useful but it needs to be predictable, so obvious there will be BIOS settings to handle the sample sizes.

A good example is a 40k core setup running some kind of molecular dynamics over a very long period of time, this feature is basically free performance for a setup that needs no small amount of money to upgrade.

Once again not really a consumer feature, but it is nice knowing something like this is more than on a whitepaper.

>>59236852

What it does tell us is AMD is playing to win when it comes to these sorts of workloads. The next move they will make will be vega which - from the so shitty event last week - seems to have some (a lot?) of hardware suited to these sorts of things.

I guess this is the current evolution of AMD's HSA initiative.

File: 7a4.png (341KB, 600x387px) Image search: [Google]

341KB, 600x387px

>>59235859
>1610 to 1640
>30 units
>max delta of 15 between runs
>less than 1% change

>>59235869
>820 to 920
>100 units
>max delta of 5 between runs
>less than 1% change

margin of error

File: AMD%20Zen%20December%202016%20Update_Final%20For%20Distribution-page-018.jpg (467KB, 3999x2250px) Image search: [Google]

467KB, 3999x2250px

>>59236852
Wonder if it's held in some dedicated SRAM or in the L3?

>>59237042
Looks like it's in the load registers, so the data is literally bite sized.

>>59236161
the longer you let the game run for, the better performance you get, can't get that from the retarded reviewers who bench 30 seconds for gaming for each game

>>59238376

>who bench 30 seconds for gaming for each game

Amusingly enough thats almsot entirely the inverse of Nvidia's gpu boost - it will clock really fucking high in those 30 second runs reviewers do because the card hasn't heated up yet which (to a degree) artifically inflates numbers.

>>59238512
It's a shitty practice but everyone does this crap, nobody mentions it because muhpayrolls

>>59238376
actually AI or repeating post effects might get a boost or things like day-night cycle for scripted characters
I know it's margin of error increase but it would be there

>>59238534
>nobody mentions it because muhpayrolls
>gotta get review ready before NDA lifts or less monies

this is false for hardware though and specially youtube, since everyone has their own isolated audiences that would watch video anyway

>>59236277
Can't wait for Intel Luddites to declare it dangerous and assert that you should buy a 7700k because it won't try to take over the world.

>>59235859
how long do you have to keep running it before it can become your wife?

>>59239355

If I use it to crunch seti@home for a couple years does this mean I will get an orion slave girl?

>>59235859
That Nvidia GraphWorks™ though...

But yeah, that prediction thing seems to help a little, which is cool.

Great for people who need to do the exact same thing several times, i.e. nobody

>>59239904
>says someone who probably shitposts all day

>>59239909
But each of my posts are unique.
Great CPU for brap posters

>>59239929
>each of my posts are unique
lol

>>59235859
>>59235869

>using a cpu to train a neural net instead of gpu

Pajeet, go home

>>59240239

Huang, pls.

File: memory-bandwidth.png (37KB, 763x662px) Image search: [Google]

37KB, 763x662px

>>59240334
White American male here.

Pic for proof of my claim, now gtfo OP

>>59235859
>1% improvement

INTEL BTFO IT'S OVERRRRR

>>59235859

These are deltas of under 1%

This is margin of error

>>59240472
anyone who uses a cpu for neural net training has a budget of a ThinkPad for their gaymen rig

>>59240447

>4 year old data

File: 1404579891208.jpg (99KB, 676x682px) Image search: [Google]

99KB, 676x682px

>>59240485

>retried 6 times
>consistent increase
>margin of error

The chances of that happening randomly are astronomically small

>>59240488
Nobody is talking about using CPUs for neural net training. They're talking about the (claimed) neural net branch prediction inside Ryzen CPUs.

How does this work? Is this the scheduler learning?

>>59240239
Nobody is talking about that you spaz

>>59236986
You know nothing about statistics

File: luk-vrbel-deep-convolutional-neural-networks-43-638.jpg (45KB, 638x479px) Image search: [Google]

45KB, 638x479px

>>59240518
Are you fucking kidding me? Nice bait troll.

>implying in the past 4 years cpu's have gotten 200x faster than gpu's at matrix math

I'll let the machine learning community know.

>>59240485
You also know nothing about statistics

>>59240590
If you think the topic at hand is the performance of neural net training on the CPU and not the neural net branch predictor in the CPU you're a fucking retard and need to leave.

>>59235859
>>59235869
1636/1619 = ~1.01050030883261272390
893/888 = ~1.00563063063063063063

>>59236164
>mfw AMD makes the only CPUs that level up

>>59240876
>another person who knows nothing about statistics

>>59235859
>>59235869

There's no reason to buy Intel now with their obsolete trash CPUs.

>>59235859
"Neural net prediction" is a simple perceptron, just like in every CPU from the past 15 years.
Don't get me wrong, Ryzen's branch prediction looks really really solid compared to Intel, but it's definitely not some unheard of black magic in chip development.

>>59235859
I wonder how far it can go? 2% 5%?

>>59241472

>really really solid compared to Intel,

Brave words to post on /g/ given /g/ believes that Intel will always - without fail - be ahead of AMD. Even now /g/ believes P4 bested athlon.

>>59235859
1.1% increase? That's about 1% more than Skylake to Kabylake.

>graph axis is from 1610 to 1640

>>59241522
Let them, I don't care what some lardass whose argument is nothing but >MUH GAYMEN on a Chilean strawhat weaving textboard thinks.
It's solely his fault for making poor choices and ending up with a slower, more expensive product.

Even if you argue with only >MUH GAYMEN, it doesn't make sense.
Gaming alone is by far not the only thing a person does at the same time.
They will have multiple programs open in the background, be it web browsers, IM clients, voice comms etc. all chugging away on those precious 4 i5/7 cores, while a 6/8 core chip can easily offload these tasks on unused cores.
This is actually the reason I bought a 4930k several years ago.

>>59241522
>/g/
Try /v/.

File: 1448917281249.jpg (39KB, 297x296px) Image search: [Google]

39KB, 297x296px

>>59241697

Is there a difference?

>>59241716
Supposedly so.

Can someone ELI5?

So now that Intel's 8-core housefires are obsolete will Ryzen's 4 and 6-core variants make i5 obsolete?

>>59241884
Don't really give a shit about poorfag CPUs

>>59240527
There's a 1 in 64 chance of that happening randomly. Not what I'd consider "astronomically small".

>>59241959
if you only look at the straight math, sure, but this is the cpu starting weaker, and getting more and more powerful as you go on doing your task, with a step up each time. the number is quite a bit more then 1 in 64 look at the intel one, that's what everyone expected to see, not a ramp up over 6 tests.

>>59241680
during haswell some people benched games on i7's and 6 and 8 cores in real world case use scenarios, sure, the 6/8 were slower, but god damn that load balancing, that's something you just don't see in normal, fresh boot nothing running benchmarks.

>>59242215
This is rarely mentioned, true, people seem to forget you do have a OS running in the background with some 2000+ threads spawned, most not doing anything, but all it takes is one haywire to lower your framerate, and Windows 10 being Windows 10 that's more likely, and hardly anyone but reviewers run games on a sterile Windows install.

>>59241680
You made sense until this part
>They will have multiple programs open in the background, be it web browsers, IM clients, voice comms etc. all chugging away on those precious 4 i5/7 cores
I'm sure all those extremely costly processes which cause 0-2% total CPU load while in a low power state at 800MHz are going to have an absolutely catastrophic impact on game performance when the processor operates at full speed at 4.7GHz or beyond.

Better double up on dem cores, get that $500 1800X, wouldn't want Steam to have 0.01% CPU usage in the background while a game is running!

>>59242270
It's called context switching, you're playing and suddenly the CPU has to flush caches to do whatever the fuck's going in the background, be it a AV for normalfags, some dynamic browser tab, IM client, or svchost and telemetry suddenly getting a fit.

>>59235859
IT HUNGERS FOR KNOWLEDGE
HUMANITY WILL FACE AN EXISTENTIAL CRISIS VERY SOON


AMD HAS DOOMED US ALL

>>59235859
We have to go deeper

>>59242828
Thanks AMD I'm literally never buying from you guys again and a overclocked 1700 looked really tempting.

>>59235859
Graphs are nvidia tier. Presented clearly though Ryzens performance even with launch hiccups is very good. this particular aspect is pretty interesting though

>>59236108
had't even thought of it in this regard.. yeah naples is going to crush it in enterprise

>>59242857
Of course you're not buying because by the time you need a new CPU these things will have already taken over humanity and started using us as manual labor to maintain their machine kingdom.

>>59236184
Feel like this was pointed towards why frametime consistency was strong in cpu bound gaming loads. Am i wrong? Is there any particular part of the architecture credited for that?

>>59242923
That's mostly due to more cores, load balancing.
Pegging 4 cores at 100% is more jittery for the OS than 8 mores at 70%, since the OS has to pull extra cycles that have nothing to do with the game from somewhere, leading to drops.

Am I the only one who noticed the actual performance increase is 1% but they scale the graph to look much bigger

>>59243002
Yes ur wery smarts here have a dongle

>>59235869
>>59235859

>Same 1% different
>Scale graph differently to make it appear more drastic

File: terminated.jpg (345KB, 925x955px) Image search: [Google]

345KB, 925x955px

>>59236266
OH GOD

File: ▄█▀ █▬█ █ █ █ █ █ █ ▀█▀.jpg (38KB, 450x495px) Image search: [Google]

38KB, 450x495px

>>59236358
>you guys need to get the 7600k not the 7400 for the 0.1% gaming improvement
>DON'T BUY AMD THOUGH IT GETS A FREE 1% IMPROVEMENT IN APPLICATION THAT MATTERS

>>59241759
OP is a masterbaiter.

File: NFS group types.jpg (509KB, 1550x1197px) Image search: [Google]

509KB, 1550x1197px

>>59243182

File: 1428197856887.jpg (59KB, 519x644px) Image search: [Google]

59KB, 519x644px

>>59242276
>His file system doesn't have native gzip-9 support
>He doesn't load average 2/3s writing 2GB/s

>>59235859
>tfw too intelligent to buy dumbtel

>>59243964
>cinebench matters
Kek

>>59236108
rip intel's server market

>>59241759
Branch prediction, but with perceptrons. Because perceptrons are in artificial neural networks, AMD gets to toss in all the machine learning buzzwords when describing their branch predictor.

>>59235859
Skynet is upon us. T-Thanks AMD

>>59235859
neural nets are notoriously bad at predictions. This is why good prediction software does not use neural nets.

>>59239929
*bbbbrrrrrrrrrrrrrrrrrpppp
but seriously, your words might be different, but your computer repeats many tasks to get your words to the website.

>>59235859
It continues to rise steadily to the end of graph. Why didnt they do more passes to see how much performance it can gain?

>>59247061
Probably just to prove that it works.

https://forums.anandtech.com/threads/ryzen-strictly-technical.2500572/page-9#post-38776310

This is certainly interesting - looks like something a windows patch could solve.

>>59235859
I-Is this real?

>>59241959
Chances are exponential for each roll, so in reality, by the 6th time the chances are 1 in 10 billion, because on each consecutive roll it doesn't only have a chance of being equal to or less than the last result, but also equal to or less than all previous results.

File: 1431229361740.png (52KB, 346x360px) Image search: [Google]

52KB, 346x360px

>>59241716

>swapped panels

Never post that image again. It ruins the timing of the delivery.

>>59242276
There is much more going on than that, and it just eats away performance due to the CPU having to flush caches for other applications etc.
While it doesn't seem much on its own, it just adds up to a whole lot of lost performance on a 4 core chip.

>>59235859
BOTNET

>>59244148
I like ProStreet.
Undercover was shite though.

>>59242276
>Better double up on dem cores, get that $500 1800X, wo
You can have a 8 core 16 thread CPU for $325.

>>59250533

With insane power effeciency as well.

File: aXm3661xjU.jpg (84KB, 650x650px) Image search: [Google]

84KB, 650x650px

>>59235859

That tickles me.

When I get bucks to blow on a replacement for my 3770K I'll go with Zen, even if just to have something unique among my relatives and friends.

The crazy multicore performance will prove useful, too.

captcha: impaired holmshill

>>59244148
I actually liked Shift a lot.