What if AMD brings Jim Keller back for another 3 years after Zen+?

What if AMD brings Jim Keller back for another 3 years after Zen+?

Does Intel have enough money to donate to Feminist Frequency to survive?
Will their new marketing campaign that declares AMD buyers sexist actually work?

Other urls found in this thread:

archive.org/details/Maximum_PC_May_2000
computermachines.org/joe/publications/pdfs/hpca2017_exascale_apu.pdf
twitter.com/SFWRedditImages

...

They will realize that Making faster single cores is getting harder and harder.

AMD seems to have made a 40% jump in IPC, but their CPUs were garbage teir to begin with.

With 10nm and 7nm being put on hold, dont expect much.

>10nm and 7nm being put on hold
they were never realistically viable.
the heat on such a small and dense area is hard as fuck to keep low

They made over a 60% IPC jump, actually.
40% was merely their goal.

I don't think 7nm is "on hold" at Intel. Just probably won't happen until 2021-2023 while AMD will be using it in 2019.

It's different kinds of 7nm actually.

Isn't 10nm the threshold at which quantum tunneling fuckery begins to happen, which makes transistor control incredibly difficult?

No, it's under 7nm on a dense process.

AMD will keep modifying the Zen core arch for a few years before they have to design a totally new high level arch. I suppose it'll depend on what intel brings out in the early 2020s. A new ISA would be a mess for the market if it didn't have 1:1 perfect X86 emulation without losing performance anywhere. Thats a pretty big task to accomplish.

10nm and 7nm process nodes aren't put on hold, not for anyone apart from intel really. Though their 10nm process is online right now. It'll likely take them a good while to start yielding significant volumes of high clocking chips, but hey, they've got 3-4 years to iron out the kinks again.

Global Foundries inherited a golden goose from IBM. Their 7nm SOI FinFET process was ready to roll. Only caveat is that they were using quad patterning so its costly, they're waiting for high power EUV machines to increase output. They start setting up the lines later in this year, and some risk production will probably take place in mid to late 2018. That isn't terribly far off. Production volume in 2019 is a realistic time frame for them.

Heat is generally addressed by lower voltages. We're really not nearing the point were self heating effect is harming performance in any way, and huge tower coolers can dissipate heat output in excess of 200w easy. Intel's dense 14nm Trigate process has stock Kaby Lake pulling 21.5w per core, and their only real issue is their TIM and the IHS shim height. The chip isn't anywhere near uncontrollable.

This meme needs to end.
A process with a given channel length will have a gate designed to sufficiently control the channel. Individual transistors are modeled to see their behavior long before they're ever put in silicon. The foundry knows things like the range of leakage based on the effective gating surface area, and channel mobility.

>the heat on such a small and dense area is hard as fuck to keep low
wow you actually have no idea what you're talking about lmao

>200w easy
lol paying high end cooling isn't cheap.
don't expect cpu speeds to be infinite. heat will continue being an issue hindering performance

Top end tower coolers are under $100, many very high performance options are around $50. It really isn't a problem.
No one is saying clocks will scale infinitely, your point is entirely fallacious.

Those intel IPC gain, is this real

>people actually keep trying to tell me that bulldozer wasn't a complete and utter fucking disaster

I don't understand these people

Bulldozer was worse than the pentium 4

>Bulldozer was worse than the pentium 4
Not even remotely close.

Eh Pentium 4 was super bad. Especially when they tried to salvage it with hyperthreading when its IPC sucked to begin with and HT just made it less efficient. lmfao whose fucking idea was that?

Pentium 4 only worked and Intel made it to 2006 to begin with because of bribes, scams, and marketing.

Bulldozer was sort of okay, it just came too late, and should have launched being as good as Piledriver was. If Piledriver came out in 2010 instead of mid 2012, AMD wouldn't have been BTFO so bad.

But really they should have simply stuck with K10, increased instructions per cycle per AGU/ALU from 3 to 4, increased cache size a little, and it would have been within 10% the IPC of Sandy Bridge and similar clocks.

So yeah, it was a huge mistake, but it shouldn't have as bad as it was.

The graph is probably not scaled per clock (IPC), otherwise the jump between Haswell and Broadwell, Broadwell and Skylake, and Skylake and Kaby Lake would be nowhere near as much. It looks like it's just showing performance of an arbitrary part of each generation at an arbitrary clock speed.

Also we've seen Ryzen single threaded benchmarks showing it matching Kaby Lake within margin of error, yet on this graph it's still decently behind. Probably because Ryzen's stock clock speeds are lower.

Never mind, just saw the graph title and that Ryzen's figure has a question mark on it.

bullshit in highly threaded tasks it fucking superseded some intel shit quite a few times

I bet you're thinking in terms of performance and not overall relative performance on the market.

OC'd bulldozers reached 500W.
By themselves.

youve fucking seen some fanfic from a couple sites

funny how those "leaks" show it doing far better than even AMD is claiming

totally legit. totally

>1/3rd speed cache

Actually they don't. AMD's benchmarks showed it on par with Broadwell (Blender) or a bit ahead of Broadwell (Handbrake). Kaby Lake only has ~3% higher IPC than Broadwell so that means it's around the same level. The leaked benchmarks show it ahead in some benchmarks, on par in others, and below in others. Seems about right to me.

RYZEN HYPE TRAIN COMING THROUGH!!
Next station: Auschwitz

archive.org/details/Maximum_PC_May_2000

My benchmark brings all the jews to the yard,
And there like,
Its better than yours,
Damn right its better than yours...

holy shit all the information is on the simple-as-fuck graph.

It's single core Cinebench R15 scores normalized to 4GHZ.

>What if AMD brings Jim Keller back for another 3 years after Zen+?

he's working at tesla now

>He's working at tesla for now
ftfy

Keller works on project basis, company bring him in with a target. He delivers a product that hits said targets, then leaves for a new project.

Yeah but this is the thing, I don't understand how after all that people people still fell for it buying 970s and 980s.

People with 5870x2 and 6990 and shit were able to play all the newest games still for 5+ games at 60fps at least. 7970 still going strong.
Peoples' NVidia cards are made obsolete so quick and they just rush out to buy a new one anyway?

Now this time around with Pascal it does seem different. It does seem NVidia is finally offering decent compete performance, and the 1060 is good value. But still. How do people ever trust NVidia to keep buying their shit with how often they burn their customers?

Elon will bribe him to stay with a pay package that AMD can't match.

DELET

While I think that Zen will be good, this benchmark is complete bullshit.
What the fuck did wccf even do to get such skewed results?

There isn't a "threshold" for quantum effects you cuckeroos, it's just generally more noticable at smaller scales but the orbit of planets is also affected by quantum effect

>They will realize that Making faster single cores is getting harder and harder.
manycore has never been tried

>high power EUV
EUV was supposed to roll-out 14 years ago. i'm really sick of this overhyped meme technology

Is made in China not as bad as Designed in Israel?

> What if AMD brings Jim Keller back for another 3 years after Zen+?
What if Intel will hire Jim Keller? They have the money to do it.

Its just assembled in China, where the die is placed on the package. The die is made in the US.

What the literal fuck? 5870 is VLIV arch not GCN, and it's not even supported today. You're better off with housefire 480 or dual housefire 480s today. And it doesn't have performance either, my GTX 460 can't even do 30 FPS in TW3 at min graphics. The 5870 might be able to do 30 assuming you don't use Win 10. It's a far cry from 60

I really doubt Keller does this stuff for the money. The dude doesn't just have the ability to reshape entire markets with everything he touches, he's done it many times in the past. I think of him like John Carmack, an autistic savant who just does whatever he finds neat and moves on when he gets bored. And considering how many stupid bullshit projects Elon is trying to half-assedly get going, I doubt Tesla will offer much to keep Keller engaged.

Fermi saw performance regressions past 2013, I wouldn't install any recent drivers on it. Kepler is seeing the same. Just the other day I gave a buddy with a 670 drivers from 2014 and his Fps in CSGO jumped from 100 to 180, which is important with his 144hz monitor.

I said 5+ year on the 5870x2. That card is 8 years old now.

But my 6 year old 7970 IS GCN and still works amazing.

>We will reach Bremermann's Limit in your lifetime
>Tech economy collapses as there are no more upgrades left to do
>Past this point you will never, EVER upgrade your PC again
>If you can't run Crysis @ 1048576K, 2985984 FPS then too bad

guys, APUs are coming with fullforce.

computermachines.org/joe/publications/pdfs/hpca2017_exascale_apu.pdf

>320 CU's at 1GHz

>100,000 high-performance computing nodes
>16TFLOPs per node

Welcome to exascale, user. We're actually trending towards exascale computing and beyond consuming the majority of all power humanity produces. Its actually a legitimate concern that it'll grow to take 99% of our power production. Its why perf/watt improvements are so vital.

Though it is somewhat hopeful that they see it as feasible to create a 16 core CPU with a 16 DP TFLOPs GPU, and keep it under 200w in the near term future.

10TFLOPs at 200w? The 1080 is "only" 8.9 and this is all 1 die.

What if 4 APUs is the new SLI/CF but actually good? 64cores, 124 threads, and 40 TFLOPS.

oh fuck it's actually 16TFLOPS

Are they assuming this would require 7nm?

Why the fuck then Keller is not in Kiketel?

16 TFLOPS double precision, it is 32 in single. The GP100 die used in the Tesla P100 is only 9.5TFLOPS single precision at 300w.
There are also 16 CPU cores fitting in that 200w power envelope.

Do try reading a bit harder.

Intel:
>How can we charge people 10% more money each year for a 3% IPC improvement, half the manufacturing quality, and 15% increased temps?
AMD:
>How can we put 3 RX480s and a 8c/16t CPU on a single CPU sized die, with 8 of them on a board using less than 200W each

Oh holy shit you're right it's only single precision for the 1080.