INTEL'S MESH FABRIC IS ACTUALLY FUCKING WORSE THAN THEIR

File: Screenshot_20170620_113836.png (19KB, 729x218px) Image search: [Google]

19KB, 729x218px

Meanwhile here's Ryzen the CCX is insanely fast alone(will be something else when a CCX is 6 cores or more)
And looking at core to core latency on Skylake makes me feel better about it.

Tom's didn't test 1800X, unfortunately

File: JUST krzanich.png (1MB, 1002x1480px) Image search: [Google]

1MB, 1002x1480px

>J-JUST WAIT™ FOR B-BIOS UPDATE GOYIM

>>60987620
Manufacturers should totally dump anything below 3000Mhz of RAM.

>>60987620
IF at 2:1 should make CCX latency irrelevant.

>>60987680
/thread
I've read several times already that Threadripper and Epyc have IF running at 1:1 or even 2:1.
Is there anything to it?

>dildograph.png

>>60987697
It's been rumored since April and it's a fact debug motherboards had 2:1 setting for IF.

Also there's simply not enough bandwidth to feed 8 channel memories using Ryzen's IF speed.


Fun trivia: IF scales to 500GB/s, this is made for GPUs

>>60987570
Their mesh fabric it's a shitload of bing busses, what did they expect?

>>60987620
>40ns core to core latency

Single CCX (Ryzen3/APU) are gonna be monsters.

>>60987723
>Fun trivia: IF scales to 500GB/s, this is made for GPUs
Nvidia in trouble soon.

>>60987570
>Bigger, more separate cores
>Take 10% more nanoseconds than previous generation
>Offers more than 10% faster and better overall performance

I think you're retarded, buddy. It's like saying a 2500lb 14 second car is better than a 3200 pounds 11 second car. It's autistic on the purest form to think so.

>>60987723
>Fun trivia: IF scales to 500GB/s, this is made for GPUs

inb4 AMD spends the next 2 years lowering bandwidth requirements so it can fit 4 dies on a video card.

>>60987750
Bandwidth has always taken a backseat to latency in CPUs, kill yourself.
Bandwidth is only a problem if you're starved in the first place, which in 98% of situations you're not, latency is always the bottleneck.

>>60987680
>>60987697
>>60987723
what is IF?

>>60987755
It's rumored that Navi will bring MCM'd GPUs
Vega's tiled rendering and primitives culling should lower bandwidth requirements enough to fit at least 3 biggish dies per package, but that would be way too insane

>>60987570
That's because ringbus doesn't have much issues under 12 cores, while it scales to absolute shit when you go for something like 20 cores.

File: 1484809854879.jpg (20KB, 316x395px) Image search: [Google]

20KB, 316x395px

>>60987723
>Fun trivia: IF scales to 500GB/s, this is made for GPUs
What the fuck

>>60987783
Infinity Fabric

>>60987896
I heard out could be the work of an antisemitic man dubbed the "Jew Killer"

>>60987570
>no net benefit with a greater energy use

ahahahaAHHahHahAhAHAHAHAHAHAHAHA

i fuckin called this like two days ago

>>60987794
When is Navi again? 2018? 2020?

>>60987620
>the fabric in Ryzen/Threadrip/Naples is so underutilized memory standards won't catch up for generations

AhahahaAHAHAHAHAHAHA

ITS LIKE THE '98-2093 IS HAPPENING ALL OVER AGAIN AHAHAHAHAHAHAHAHAHA

This is too much holy shit I cant stop laughing

>>60987955
Q4 2018 or Q1 2019

>>60987980
Damn it. If I can fucking find it I'm going to get another polaris card and crossfire it while waiting for Navi.

>>60988000
>Vega not even out
>Wait for Navi《tm》

>>60988020
Not interested in Vega. I want Navi.

>>60987957
> '98-2093

yeah cant wait to see a 100 years of amd reign

File: 1497921066971.png (148KB, 602x389px) Image search: [Google]

148KB, 602x389px

>>60987570
lol.

>>60989158
Gross

>Intel's new cache hierarchy is similar to Bulldozer
>surprised it's shit

>>60987896
Wendell talked it about it in part 1 of the L1T ryzen videos, APUs will literally just have a GPU connected through IF as standard. Just another standard interconnect like a core cluster, afaik.

>>60987664

2666 is the fastest non-overclocked RAM.

>>60987570
>WORSE
You do realize that bandwidth increased by ~25%? If you're just judging by latency then the same can be said of pretty much every successive generation of DDR RAM.

>>60987620
>no bandwidth stats
you failed op

>>60989873
LGA-4094 HBM APU fucking when?

>>60988188
Same here, I can't justify (in my head) upgrading my PC yet, so I'll keep saving my dollars for a couple more years.
Current PC:
i7-3820
GTX-780Ti
32GB DDR3-1866MHz CL10
Win10Pro (this eats up 2.4GB of RAM by itself)
It still runs everything JUST well enough yet to postpone my wanting to upgrade, as such I wait for Zen2 and Navi / Volta

>>60989977
That would destroy both intel and nvidia

>>60987570
>mesh fabric is actually worse
Why does this company still exist?

File: ehp-node-551x385.gif (173KB, 551x385px) Image search: [Google]

173KB, 551x385px

>>60989977

Pic related is an early exascale APU design.

32 CPU cores with literally hundreds of GPU cores in the middle all connected with IF.

This has all completely changed since Vega. HGCC specifically reducing the amount of HBM needed, and SSG acting as a huge active volatile memory. We are now at a point where only one or two industries could tax out this amount of data throughput.

Now they just need money for their software stack.

>>60988188
Vega is pretty much a single high-end chip, but will there be mid-low end Navi chips? I never bothered with AMD GPUs before only because nVidia had better fanless/silent cards.

>>60990145
Holy fucking shit that's insane.
How about something more realistic, 1-2 CCX and some 1050Ti-tier GPU?
Actually, can any desktop OS handle multi-tiered RAM? Like 16GB of HBM shared between CPU and GPU and 2-4 channels of DDR?

>>60987742
don't be too quick to say that, NVLINK does up to 150Gbyte/s currently
https://en.wikipedia.org/wiki/NVLink

File: intel future.png (20KB, 628x516px) Image search: [Google]

20KB, 628x516px

You guys ready for that 18 core part?

File: 1489340459137.png (195KB, 498x332px) Image search: [Google]

195KB, 498x332px

>>60990296
Not even Intel is ready for it. >>60982659

>>60987570

It is a trade-off. It is meant for cramming more cores onto a monolith die. It doesn't work well with low-core count chips which is why Intel never bother with it until they went beyond 10-cores.

>>60990347
What's the advantage? The yields are still shit.

>>60990250
Nvidia without competence maybe in 2025 put NVlink in consumer GPU.

>>60990362

In theory, cache cohesion is much easier to manage with a monolith die. You also don't have to worry about glue logic between different dies.

Intel gambled on their fab tech to overcome the difficulites in having massive dies but it appears to have backfire massively which is why they are already shifting to multi-die chip designs for their future Enterprise/HEDT platforms. It will take a couple of years for such chips to come to the market though.

>>60990469
intel's new arch will show up after zen support ends, who knows if amd is cooking up something for that time too

File: 1494967406388.jpg (57KB, 900x505px) Image search: [Google]

57KB, 900x505px

>>60987570

gotta make room for that RFID botnet chip

>>60990386
i dont think that people realise what amd is trying to do with vega and ryzen...
if the rumors are true and vega plus ryzen is the first true hsa platform we will see insane shit

>>60990558

Not without the software support. It will always be two separate systems.

>>60990528
>which is why they are already shifting to multi-die chip designs for their future
Source? First I've heard of this.
>who knows if amd is cooking up something for that time too
Of course, AMD isn't sitting on their ass. They are already working on the successor to Zen.

>>60990711
you seem to have misquoted, friend, I wasn't the one talking about intel multi die stuff

>>60990711
Meant to quote >>60990469

>>60989873
Wait.. does that mean the APU's GPU won't be connected with PCIe?

>>60990550
apparently it's not connected and can't do shit, it's left over from being scrapped xeons.

>>60990558
Yes,HSA is insane, offload heavy computing on GPU with low latency, Signal processing,matrix multiplication.

>>60990228
>can any desktop OS handle multi-tiered RAM?
Firmware does that, all the OS sees is a shared memory pool.

>>60990807
Suuuuuuuuuuure.

>>60990296
My fire insurance is ready.

>>60990250
What NVlink is point to point PCIe on steroids, it's not a fabric for gluing stuff outside and inside the chips.

>>60990469
AMD says scaling is near liner up to 64 cores on two sockets, apparently all those NUMA nodes aren't a problem.

But look at it this way, Intel's latencies increase with core count, AMD's 32 core chips are 4 different chip doing their own thing, assuming a task is not extremely heavy on inter thread communication, and pretty much anything running on that many threads is not, 4 clusters of cores with their own dedicated cache is gonna do it much faster and much more efficient.

As for the inter thread communication part, microcode could ensure a core is dedicated to such a task (like pulling cache data from other cores) so other cores can do their job without hopping through IF.
Anyhow, I don't see a problem with this issue, most workloads that scale to the moon aren't heavy on inter thread communication and multiple pieces of software will delegated to their own die or CCX, and a CCX alone has latencies on par with a 7700k, which is clocked significantly higher, no Broadwell-E or Skylake-X can touch those latencies.

>>60990120
Because they sit on more money than some countries have in their entire budget

This multicore nonsense has to stop.
Why not make a behemoth single threaded cpu.
With a fucking insanely large instruction set, for every single cases, like "add 7 int with 7 other ints and multiply those 3 other number by register z"
One hell of a compiler work, and a lot of inefficiencies. But compare that to using 1 core out of 10 most of the time.

>>60992320
They already tried that, it was called Itanic, and it failed miserably.

>>60992320
You're asking for VLIW. VLIW is always good on paper, but ultimately, never fucking works. So nech yourself, pal.

>>60992344
Then, would it kinda work with grouping two cores together? And limit it at that.
Spawn same context for both cores, run a instruction on both.
I guess it would need a controller ship though.

>>60992424
So, Bulldozer? It also don't fucking work.

>>60990114
>tfw when the most powerful graphics solution on the market is integrated

>>60990807
hehehe stupid goyim

File: 1492248557744.png (177KB, 465x607px) Image search: [Google]

177KB, 465x607px

>>60992320
>>60992424
>boy, everyone is stupid except me
>those retarded engineers surely never thought of that, i'm a genius

>>60992764
>let's try VLIW, it'll surely work this time!

>>60992784
>>60992764
>>60992375
VLIW worked like bliss on radeon HD 4000-6000 series.
Educate me.

>>60992806
Ok
HD 6000s were GPUs. We are talking about CPUs. While we are at it, those are old GPUs, been there, done that. There are better ways to do things.

Fuck Intel

Buy 1700

>>60992806
>like bliss
The drivers were ASS due to VLIW. Also GPGPU stuff is near impossible on VLIW.

>>60990807
Nobody believes you.

>>60992942
GPGPU worked, but that's not what we buy graphics cards for.
There was no Opencl at the time, so it was just an AMD API, so that didn't help.

File: IMG_1238.jpg (2MB, 4032x3024px) Image search: [Google]

2MB, 4032x3024px

Investor here. Just sold etsy and looking for a new stock. Is AMD our guy?

>>60992997
>GPGPU worked, but that's not what we buy graphics cards for.
>Graphics cards are only for playing games

>>60993012
Yeah, unless Intel is brewing something really new, Zen2 is gonna literally make them the new AMD.
I kinda remember them outing a bunch of engineers last year though.

>>60993056
Intel's new arch is at least 4 years away. Remember Zen started in 2012.

File: 88a.png (48KB, 200x200px) Image search: [Google]

48KB, 200x200px

>>60990807

>>60993012
>>60993056
>>60993078
After looking at AMD financial information I can't justify an investment. Earnings are bad and the book value and free cash is extremely low. Sure they may have an advantage over Intel now, but that takes time to reflect in earnings per share.

I'll wait until after Q2 earnings to take a better look. I assume their console chip sales went up because of E3.

>>60993336
Who the fuck cares you nigger cattle?

>>60993336
The logic is sound and we are in no place to tell you what to do with you money, but;

>AMD are selling GPUs as fast as they can be made thanks to coin miners, great burst of income
>EPYC launches in hours, already 2 of the 7+1 confirmed to be using EPYC (Amazon and Microsoft) with Dropbox showing interest at Financial analyst day
>3 more months of Ryzen sales to show
If you leave it until Q2 the stock price may well double, if not more considering how volatile AMD is of a stock. But it's comfortable at $10, it's not going any lower than that unless Ryzens start bursting into flames.

>>60987570
>Intel has 83ns

>>60987620
>AMD has 111ns

LMAO AMD IS FUCKING FINISHED HOLY SHIT BROS

>>60993353
He's telling the truth, good companies with good products are not always companies worth investing in.

>>60993336
Unless you understand what AMD is producing, from a technical side and have good knowledge on the market trends, you shouldn't be investing in AMD, it's too volatile, but it's potential is pretty amazing with the ability to increase its revenue by 5-7x in the next 4 years.

>>60993451
>intel has 83 ns between all cores
>amd has 111 ns between specific cores

File: AMD.png (59KB, 832x537px) Image search: [Google]

59KB, 832x537px

AMD IS FUCKING FINISHED!

THE STOCKS ARE PLUMMETING AND PEOPLE ARE BAILING OUT!

ABANDON SHIP RIGHT NOW!

File: OdRDs4u.jpg (48KB, 510x543px) Image search: [Google]

48KB, 510x543px

>>60993583
did intel shut down the shilling class for the summer?

I love AMD's recent trend of sandbagging.

Remember those Vega numbers against a Titan XP? Turns out they were not using pro drivers, but regular ones.

>>60992997
CUDA was there already as well, it was introduced on 8 series which was DX10 thing, it was even before HD3000

>>60993583
>latency between CCXs is higher than within one of them
WOOOOOW, so fucking new information

>>60993613
What numbers? They told us precisely fuck about Vega.

>>60993583
To be more informative, you should see about the same latency along every 32 threads of TR.
And way higher latency between two EPYC CPUs.

>>60993636
Those Catia, Creo and Solidworks ones.

>>60993652
You're not supposed to compare gaymen and workstation GPUs.

>>60993684
Follow the conversation.

>>60993705
So you mean they're selling people workstation cards with no workstation drivers ready and i'm not even talking software profiles? That's dumb.

>>60993728
Vega FE is under the "pro" brand, it's just that those benches were ran with gaymer drivers with no profiles or optimizations for pro workloads, obviously it has its own pro drivers but it can also run gaymer drivers.

>>60993728
nothing is being sold dumbass

>>60993853
>you can preorder Vega FE card right now
?????

>>60993886
and because you preordered it means that the drivers can't be fixed/updated why?

>>60993886
do you have an actual card on your hands?

>>60992806
>Educate me.
VLIW GPU:
thousands of cores
+
most intensive workloads are already well designed in parallel
=
no problem when half of cores are cycling noops during non-parallel tasks

VLIW CPU:
few cores
+
most tasks are hard to parallelise (it works in specialized cases, but the market for that is way to small to justify such investment)
=
most of the time the "very long" part of VLIW is underutilized, or not used at all; advantage is so negligible that you're better off with stronger CISC/RISC

>>60993619
ATi Stream existed before CUDA, and in fact OpenCL support existed through most of the life of Terascale

>>60993597
Different companies and political parties are discovering that ham-fisted shit tier shilling actually hurts them a lot more than no shilling at all.

>>60993500
>pretty amazing with the ability to increase its revenue by 5-7x in the next 4 years.
I see no proof of this at all.

>>60990093
For CPUs I'm waiting to see what their APUs look like. I've gotten spoiled by having integrated graphics on the side helping troubleshoot any dedicated GPU problems.

>>60990228
Navi is going to be what Zen+ is to Zen. I'm expecting another Evergreen out of Raja since this is going to be entirely his baby.

>>60994116
>EPYC smashes Intel in all but 4 socket market
>Intel will not lower prices to compete because they want to maintain 60% margins
>Intel's biggest earner is the server market
Even if AMD only secure 5-10%, that's hundreds of millions in income.

>>60994116
You see no proof of this at all? Where do you think Intel gets most of its money from?
x86 is a $20 billion market, quarterly.
AMD doesn't even get $1 billion of that at the moment.

>>60994173
>hundreds of millions in income.
That would only be a revenue increase of 10-20%.
>>60994189

>AMD doesn't even get $1 billion of that at the moment.
Last year AMD had total revenue of 1 billion. That's it. Of course not all of that is x86.
Now in the x86 space it isn't just Intel vs AMD. The big players like IBM are not going to have revenue slashed from AMD.

I agree with you that they will increase revenue. But 5-7x just isn't realistic. Maybe as the guy above said, 20%.

>>60993012
Eww

>>60993012
Well, you're late. AMD was $2 a few years ago. Definitely not worth buying now.

>>60994173
AMD spent billions on R&D and manufacturing of Ryzen.

>>60995089
Lol no.
Half of Zen's base come from bulldozer power IP and jaguar.
In the first place AMD never had such money and its R&D spending over the years was around $270 million per year in the last 5 years.

>>60995082
>Well, you're late. AMD was $2 a few years ago.

This, really. I was going to when it had bottomed out and the first murmurs of Zen actually being a competent product began. Potential family crisis (which ended up being nothing) required I sit on my investable funds just in case.

Such is life.

https://www.servethehome.com/amd-epyc-7601-dual-socket-early-power-consumption-observations/

File: Praise Kel.png (400KB, 816x1020px) Image search: [Google]

400KB, 816x1020px

>>60987570
"Final form? This is not my final form.

I haven't even started yet."

- Jim Keller, C.S.W.

File: amadaexciteULTRA.png (210KB, 422x537px) Image search: [Google]

210KB, 422x537px

>>60996074
>With the level of power/ performance of the new systems, you can essentially replace four Intel Xeon E5-2600 (V1) servers with a single dual socket EPYC node and get more performance (in most cases) in a single node that uses half the power. That is absolutely stellar.
>The AMD EPYC platform is still seeing major updates to BIOS for power and performance which is why we are calling these preliminary results. At the same time, we are already seeing some impressive figures.

It's over, Intel can just go home now.

>>60995082
What if it climbs up to the 50 dollar range in 4 years? It should still be ok for long term.

>>60990550
holy shit wait

does this mean

does this mean, i can use my CPU to make wireless payments to intel so i can have RAID?