What the hell's up with AMD's clock scaling this gen,

Is it because GCN is a notably shorter pipeline than Pascal?

>>55683728
Because that's how the architecture works.

>>55683728
Because the reference can't clock to the max clock most of the time.

Also nvidia just has shit scaling because their cards go above max clocks when not OC'ed and doesn't change much even when you do OC.

>~20% OC
>~20% higher performance

Seems to add up

>>55683728
Thats nearly a 20% OC.

Clock speed is relative.

>>55683778
18% OC
21% performance improvement

How does this make a lick of sense, show me a single fucking IC that linearly scales performance with clockspeed.

>>55683778
LOL NO, no it fucking doesn't, GPU clockspeed only raises the frequency of the core logic like ALUs and TMUs, it doesn't raise the clockspeed of the uncore which runs at much lower frequencies.

>>55683833
AMD runs all of their GPU logic off a single clock speed. If the input clock increases, the entire GPU speeds up. They have always run their GPUs like that.

>>55683865
You're telling me the memory controller and I/O run at those clocks?
Sorry, not buying it.

>>55683728
Where is this chart from?

>>55683802
Probably added memory overclock

>>55683928
http://imgur.com/a/sw5u3

File: chrome_2016-07-21_18-19-50.png (21KB, 819x151px) Image search: [Google]

21KB, 819x151px

>>55683728
nice source there

File: Screenshot from 2016-07-21 17-28-40.png (15KB, 615x232px) Image search: [Google]

15KB, 615x232px

Memory clock alone increases performance by a large margin.

http://www.hardwareluxx.de/index.php/artikel/hardware/grafikkarten/39615-amd-radeon-rx-480-im-test.html?start=25

>>55683882
GPU =/= CPU. Things are a bit more simple.

The mem controller speed is directly linked to the VRAM, which is basically L3$. When you overclock the VRAM you're also clocking up the controllers and their ability to handle I/O ops to and from the L2.
The shaders, texture, load/store, and ROPs are all running at the same frequency as well. load/store units handle acceses to/from L2$ to/from L1 data caches. The L2 is not an "uncore" it is intrinsically linked to the clock of the rest of the chip.
Whether anything runs at a divider or multiplier to the core speed is unknown but consider that cards report their texel and pixel fillrates in lockstep with core speed changes, so at the very least we know that the load/store and Render BackEnd are linked to the core clock.

File: Screenshot from 2016-07-21 17-30-49.png (18KB, 610x252px) Image search: [Google]

18KB, 610x252px

>>55684106
Also undervolting increases performance because of throttling.

As expected of AMD

>>55684139
From release to now I must say. The RX480 is a really weird card.

>>55684213
It was bit rushed is what you should be saying.

They skipped out on QA and optimization time. Suffered negative press for it.

>>55684139
>>55684213
That's likely a stock cooler. It's getting to the maximum allowed temperature and throttling. Nothing weird there, honestly.

>>55684233
Na I just think it's weird. That's what I want to say.

>>55684213
It's not really a weird card, it's just a badly bungled reference implementation. They somehow succeeded in making it power AND heat constrained at the same time. Non-reference cards are fixing these issues so they don't just benefit from higher clocks.

>>55684213
AMD has this habit of not giving you 100% of the chip out of the box, they expect you to do some work to make it shine, unlike Nvidia.

>>55684139
A 4GB AIB card with 8pin under $210-215 would be a fucking monster for the price.

>>55684296
>>55684300
Honestly, AMD knows their stock implementations are shit and don't bother with it. The 480 chip is pretty damn great, but it will take non-reference coolers and changes to make the chip achieve its potential. Much like the 200 series stock coolers were awful and the non-reference ones were and still are amazing, the 480 is just another page of the same story.

>>55684249
>>55684296
>>55684300

Naw none of that, I just think it's weird.

>>55684378
If I knew that I would not release my NDA for reviews until non-reference designs were out, honestly.

>>55684404
Which is like saying "boy, that card is really something" - which of course communicates nothing except surprise.

>>55683728
>>55683802
Reference RX 480 probably slightly below 1200Mhz in Crysis 3 and is power limited, so clocks fluctuate.

That custom RX 480 keeps 1420Mhz steady as a rock and has more memory bandwidth

>>55684300
Atleast we know that AMD drivers usually tend to give you a good chunk more performance over time. I'd expect this card to get about extra 10-20% more as it matures in driver.

This isn't due to AMD being better, but simply Nvidia has a really good driver team thats already optimized their cards beforehand. So they simply have to make subtle changes here and there.

Meanwhile AMD releases cards with sub-optimial drivers due to driver team lacking funding for extra tests/heads/optimization time. However as time goes, you get those anyway.

As pricing is dictated by current state competitor, if the cards are close enough like RX480/1060, its always wiser to pick out AMD simply because their cards are not utilized properly and will only get better as it ages.

>>55684435
Yes this is exactly where I was going when i meant the card is weird.

>>55684411
Yea, but there's nothing they can really do about it. The only thing they can do is work closer with the non-reference card makers to have them push out their implementations faster, which is what they should've done. It's been almost a month since official release, and we're still waiting for the first ones to come out.

Makes them look like shit, and it's bad for business. I've no doubt that factory OC'd aftermarket 480s would be better than the 1060 chips we've already seen, even in DX11 games, and at a lower price point, but alas, all the reviewers have to go with is a shitty stock cooler 480, and it's just not good enough. AMD is losing again in the PR department, even though they're actually doing well in terms of technology and hardware, despite being incompetent when it really matters.

>>55684550
doesn't matter what AMD does. Half the press is on Nvidia's payroll.

>>55684550
story of their lives basically.

>Hey guis let's release an octocore chip in 2011 when nobody fucking does anything with more than 2 cores
>oh its 2014 now and core are starting to be used, better focus entirely on 4 core APUs

At least Zen didn't back away from large core commitment now, when that shit finally matters.

>>55684561
This is also true. Half the press showed games where Nvidia is clearly favored, while shying away from DX12 games or showing ones with broken implementations (Tomb Raider). Doesn't help AMD either. But that's something completely out of AMD's reach, unless they open their pockets to pay more than Nvidia does, and we all know how that would go.

>>55683728
That is roughly an increaseof 17% on the clock, which results in a slightly above 20% increase in performance.
Pretty much a reason to doubt the entire chart.
From everything I know the best possible outcome is that the increase in performance is equal to the increase of the clock speed in a theorethical scenario.
Real world showed us over decades that usually the performance increase is at best slightly below the increase in clockspeed.

This tells me either someone fucked up on the chart or shit is on any one end manipulated.

>>55684729
It includes a memory overclock, which alone nets you close to 10% without touching the GPU, and the 480 reference suffers from both thermal and power issues fixed with AIB cards.
So no, this isn't all that weird, there are niggerigged 480's at 1410MHz and a firescore higher than a Nano out there.

>>55683728
Awwww yiss. Cancled my 1060 backorder yesterday, looks like it was a good move. Good OC cards are always where the best bang for buck is.

>>55684106
Yes, but that is a 7.9% increase in performance while the the clockspeed inscreased by 15%.
This seems absolutely realistic, gaining more performance percentual than the actuall increase in clockspeed is not.
Any chart that shows something like that has to be questioned and taken with a big grain of salt.

>>55684362
>tfw 4gb was basically a bait and switch phantom

Not as bad as nvidias fuck you editions and bullshit msrp though.

>>55684781
Just make sure you get an aftermarket 480 and you'll be fine. If you're not in a hurry, I'd even wait to see which one is best, but judging from the 300 series, you can't go wrong with Sapphire, PowerColor or MSI.

>>55684813
My best regards for Sapphire, especially Toxic and Vapor-x cards.

>>55684813
I'm in no hurry. Probably going to wait until stock is stable for both and enjoy the price war.

Would skull fuck an old lady to have gotten that fury for $275 though. Deal of the year right there.

>>55684783
Memory is never going to scale with the clock increase unless the card is severely bandwidth limited like APUs are, so this is expected behavior.

7.5% from the memory plus a 12% clock increase (1266 to 1420) = 20.5%
It falls in line assuming core clock to perf is linear WITH increased bandwidth

>>55684362
>$210-$215
>$10-$15 margin
>cost of af production alone adds additional $10-$15
>no profit margin

Are you stupid? At best, I'd see the 4GB AfM cards being $230

>>55684843
No? Because there are AIB cards that cost the same as reference cards out there.
And no I'm not talking about the Fools Edition cashgrab Nvidia does.

>>55685012
You're right. Now that I re-think this through. Nvidia subsidized the cards to AIB. Those AIB cut their profit short inorder to not lose to AMD.

I suspect AMD subsidized their cards too. So profits would still be there if they price it at $215. However we don't know how much discount the AIBs get. Maybe it depends on number of cards AIB partners order. So if AIB are small volume, their discounts are lower. If discounts are lower, pricing at $215 might put them in negative or close to no profit.

>>55685012
>>55685082
AMD sells the chips and reference boards, nothing else. This is why 3rd party (please stop referring to 3rd party as "AIB") manufacturers can sell their own card for the same or nearly the same price if they want to, they don't need to buy a board AMD also needs to profit off of, it cuts out a middleman.

The chip is going to priced double what it costs to make + expected per-unit R&D margin.
Boards are honestly dirt cheap, you have a $10 PCB and a smattering of parts that bought in volume all cost about $12-15 per board (resistors are pennies each - MOSFETS are quarters, then IC logic etc.etc.), and then about $15 in VRAM. On top of that you have, again, expected per-unit R&D margin.

If the final margin of a $240 product is any less than 40% from any vendor I would be totally surprised.

>>55684835
>Fury for 275
Holy shit, have you missed out on that one

>>55683728
Because 200MHz OC is almost 20% on the RX480 while 200MHz OC is less than 10% on Pascal.

>>55684574
The problem was not that it had 8 cores and most of the software was using only 2.
The problem was that the cores were shit. Phenom II had better single core performance than bulldozer.

File: Screenshot_2.jpg (163KB, 782x379px) Image search: [Google]

163KB, 782x379px

>>55685835
Sigh. No, actually it did not. Bulldozer cores were a slight IPC improvement over Phenom II right from the get-go.

I know passmark is shit, blah blah blah, but it does show raw, relative performance in a certain task (completing the passmark benchmark). Thus for my narrow argument it is valid.

The reason big Phenoms outperformed the 8150 back when it came out were many, most of which related to the fact that CMT core architecture was not something most software handled very well. First of all, sharing an FPU between core modules meant that FPU-based operations often got shat on due to delays in freeing up that resource. Second, operating systems of the day didn't unpark Bulldozer cores well, so they got more latency there. Finally, AMD fucked up how those cores accessed their cache, adding latency there.

As you can see in the picture, by the time we get to the FX-8350, which is a refresh of the same Bulldozer core architecture, many of those problems were mitigated, and the core's performance was clearly better than K10.

Which is not to say I don't have a massive boner for all K10/LLano chips to this day.

>>55684839
Let us assume you really could simply add the performance incease of the chip and vram clock together like that, in that case it would still only be a best case of 19.5% (7.5+12=19.5) and not 20.5%.

>>55685302
Did not read the thread did we.
This falls into the theorethically impossible area.

>>55686034
>Bulldozer vs Phenom II
>http://cpu.userbenchmark.com/Compare/AMD-FX-8150-vs-AMD-Phenom-II-X6-1100T/2006vs2004

IPC increase was mainly due to higher default clock.

3.3Ghz vs 3.6 Ghz
Thats 10%. The IPC improvement between 8150 and 1100t is roughly 7%.

Per clock basis, they're roughly same.

>http://cpu.userbenchmark.com/Compare/AMD-FX-8350-vs-AMD-Phenom-II-X6-1100T/1489vs2004

Even the 1100T vs 8350 is roughly similar story. However with the older 8150, it was 7% performance for 10% clock increase. With 8350, its 21% clock for 22%.

This isn't limited to AMD either. If you look at Intel's side, its roughly same.

>http://cpu.userbenchmark.com/Compare/Intel-Core-i7-6700K-vs-Intel-Core-i7-2700K/3502vs1985

6700k 4Ghz vs 2700k 3.5Ghz. Difference in clock is about 14%, performance is about 20%

Its the reason why OC'd Sandybridges still perform the same as latest new intel shit.

>>55686245
>Its the reason why OC'd Sandybridges still perform the same as latest new intel shit.
There's a 10-15% improvement in Skylake over Sandy Bridge though. Also less power.

>>55683728
>>55684447
So anyone care to explain how XFX Black Edition works at 1328Mhz if actual reference can't hold it's stock clock at 1266 while using the same heatsink?

File: 41694.png (29KB, 550x420px) Image search: [Google]

29KB, 550x420px

>>55686034

>>55686458
During testing it peaked at 1328 for a nano second so they slapped that for marketing. It's a known fact that RX480 can't reach 1266 without giving it extra power.

>>55686458
>what is mv per clock

>>55686522
Ignore shitposting /v/ children who spam passmark garbage here

>>55686423
6700k TDP is 91w.
2700k TDP is 95w.

4Ghz stock vs 3.5 stock.

6700k overclocks to about 4.7 on max.
2700k overclocks to about 4.8 on max easily.

That 10-15% is now 0%. I'm being generous and giving a 20% performance. So now the 4.7 skylake vs 4.8 sandy difference is reduced to 2-3%.

However that doesn't mean its the same. The newer skylake obviously supports better memory, DDR4 which has been known to be faster in certain tasks and certain games.

Is there any ranking for which aftermarket cards are the best? Both 480 and 1060

>>55684729
reference rx480 with 100Mhz OC gives you about 10% boost in performance, so this seems logical to be 20%.

>>55686756
>AMD
Sapphire usually is the best.
>nVidia
Asus or EVGA can't decide.

>>55686756
Check warranty. Whoever has better warranty usually is the better company. EVGA has a good reputation on nVidia side of things. They used to have lifetime warranty, but they removed it. However its still applicable if you owned an older EVGA card that had it. If those broke down due to natural use, they will replace it with current gen similar performing/price card. Carrying over the life warranty.

>>55686164
The benchmark numers are so low, that 1% could be margin of error.

>>55686458
because it's not power limited like the reference model.
Slap a decent cooler on the reference 480 and raise the power limit and it stays at 1266 all day and goes to 1350Mhz OC'd.

File: facepalm.jpg (90KB, 1440x1080px) Image search: [Google]

90KB, 1440x1080px

>>55686245
>IPC increase was mainly due to higher default clock.
>Instruction Per Clock increase was due to higher clock
Please stop using terms you don't understand.

>>55687149
So you're saying 1 Ghz 6700k has same IPC as 4Ghz 6700k?

If clock speed isn't a thing, then this should be true right?

File: helpp.jpg (198KB, 720x960px) Image search: [Google]

198KB, 720x960px

GOOD JOB MR AMD

YOU DID IT

YOU ARE HERE

>>55687204
Holy shit you're actually retarded.

They are the same die, the same architecture. IPC is exactly the same.
IPC has nothing to do with clock speed.

>>55687256

Different anon here, but isn't instructions per second the important metric here as that factors in both IPC and clockspeed right?

File: doubleFacepalm.jpg (67KB, 750x600px) Image search: [Google]

67KB, 750x600px

>>55687204
IPC is Instructions PER CLOCK. Clock speed inherently has no effect on it. You're thinking of total performance, which is IPC multiplied by the clock speed.

You're too dense for words.

>>55687286
No one uses this as a metric.

A micro arch has a finite potential for throughput, that is its IPC with a given type of instruction.

>>55687256
You're retarded. If you walked 2 miles every hour, thats 2 miles per hour. If you ran 10 miles per hour, thats 10 miles per hour.

If your CPU can do 10 works @ 1 Ghz and 20 works @ 2 Ghz, thats twice the IPC.

>>55687289
Rereading your shit, I see where your dumbass shit logic comes from. Its not instructions per clock, but cycle. Cycle is a arbitrary time. Faster you run, faster the cycle.

>>55687333
Damn you're dumb

File: smuggeramada.jpg (233KB, 512x483px) Image search: [Google]

233KB, 512x483px

Retards buy the GTX 1060 for "The Way Its Meant To Be Played" and because they saw a benchmark where it does better than a stock 480.

Everyone else buys a 480 and overclocks it to BTFO the 1060 in every way possible.

All is right in the world.

File: tripleFacepalm.jpg (31KB, 636x367px) Image search: [Google]

31KB, 636x367px

>>55687333
>Cycle is a arbitrary time. Faster you run, faster the cycle.
Stop this, the internet is running out of facepalms.

>>55687333
Its clock cycle, you tech illiterate retard.
2 (X) instructions processed per clock cycle.

>>55687368
>480
>overclock
kek

>>55687373
>>55687353
>>55687380
https://en.wikipedia.org/wiki/Instructions_per_cycle

Go troll somewhere else.

>>55687409
>tech illiterate retard links a wiki page that explicitly proves himself wrong

Top tier bait

>>55687373
so, how fast is a cycle?

>>55687464

About half as fast as a piece of string is long.

>this car's wheels have a certain number of revolutions theyt can do per second.
>another quantity is the number of meters the car can travel per revolution of the wheels. let's call that the IPC.
>the meters per revolution increased mainly due to the increased number of revolutions per second.

Lol wut?

>>55687464
four gigahertz to the power of minus one

>>55687333
>>55687409
>>55687464
You need to redo secondary school physics...

File: finalFacepalm.jpg (31KB, 499x399px) Image search: [Google]

31KB, 499x399px

>>55687409
>>55687464
That's it, I'm officially out of facepalms. Congratulations, you broke the internet with your stupid.

Hol up, what about the AIB 1060's though? Surely they'll make a good tri-fan one. Might be hard to compete in the RAM department though.

If NVIDIA had any brains they would lower the price of the 1080 to 500/550 and 1070 to 300/325 and just crush AMD this cycle. They would literally never recover.

Just some thoughts my .02
I'm going Nitro 480 anyway, first time fagmen build.
>Also using GPU for gaming.

>>55687832
>just crush AMD this cycle
I have a suspicion that AMD and NVidia are more buddy-buddy than you think. They were done for price fixing not too long ago. I also doubt that NVidia want to be alone as the only GPU manufacturer.

>>55687871
>I also doubt that NVidia want to be alone as the only GPU manufacturer.
Aren't there a lot of anti monopoly laws out there that would literally fuck nVidia to hell and back if they did that?

>>55687333
>>>55687256
>You're retarded. If you walked 2 miles every hour, thats 2 miles per hour. If you ran 10 miles per hour, thats 10 miles per hour.
>If your CPU can do 10 works @ 1 Ghz and 20 works @ 2 Ghz, thats twice the IPC.

Hory sheet! it would have been twice the ipc if you had actually changed something, the top one, sure that is OK logic, but the second one is like saying "if you walk 5 miles in 1 hour, and 10 miles in 2 hours, then that means that you walk twice as fast when you walked 10 miles"
Damn you're stupid...

>>55687333
Holy fuck, you can't be THIS fucking retarded.

>>55687333
Do you know what constant and variable are?

So how much is the Nitro gonna cost?

>>55683802
Maybe the new cooling solution allows it to not throttle when OC'd? Also, I don't see your point.
The stock clock Sapphire card runs at 59.3, a 18% OC yields a 8% increase in performance. Or is the Sapphire one already OC'd?