Zen
Images are sometimes not shown due to bandwidth/network limitations. Refreshing the page usually helps.
You are currently reading a thread in /g/ - Technology
You are currently reading a thread in /g/ - Technology
In this post we wait for Zen, have a seat...
>>51478008
>declares for bankruptcy
>no drivers ever
Why are you retards hyping this up? Don't you remember how hyped up Bulldozer was? You can't claim muh Jim Keller either. You have absolutely no reason to hype anymore.
>>51479122
>You can't claim muh Jim Keller either.
Why?
http://www.wikihow.com/Begin-Zen-Meditation-%28Zazen%29
>>51479136
Because he left AMD.
>>51479168
>doesn't know how Keller works.
You're actually retarded.
>>51479168
Complex ICs take years of planning and hundreds/thousands of engineers to make. Keller was a lead engineer not the only engineer. He almost certainly left a team behind to finish things up.
>>51479188
You don't know how he works either, you're just clinging on to a small hope. Are you idiots just addicted to disappointment?
okay im sure it will be soon
YOU JUST WAIT
>>51479218
Here's a little assignment for you, since you're clearly underage and uninformed.
Go to Google, find out how long Keller was with AMD for K7/K8 and report back with your findings.
I wish you luck you little fucking retard.
Buying zen if it is not another bulldozer, I don't want to guy good goy processors. If it gives half decent performance, I am in.
>>51479211
>Jim Keller will make Zen great
>He leaves
>Jim Keller wasn't the only engineer, they don't need him
okay amdfags
>>51479168
After Zen was done
>>51479262
>>51479211
>He almost certainly left a team behind to finish things up.
nice reading
Why is there always a few 14 year old Intel fanboys making a big deal out of Keller leaving?
You know it's still 11 months away? Why are you making chain threads now?
>>51479279
Because Intel are incapable of making a worthy upgrade from Sandy Bridge so Zen is the only glimmer of hope left.
>>51479278
Because of all the 14 year old AMD fanboys that made him into some misaiah
>>51479289
>underageb& btfo
back to le 9gag you go
>>51479278
They love generating these sort of reactions, and they also love indulging in their highly distorted, childish view of company politics like the autistic little shitposting mouthbreathers they are.
All you can do is report and ignore. Honestly if you want real discussion about Zen, just move over to Overclock.net.
>>51479318
>misaiah
You retarded little kids never fail to make fools of yourselves do you.
>>51479318
Name one time he fucked up and didn't lead the best architecture a company he has worked at has produced.
Protip: You can't. He saved AMD, he saved Apple, he'll do it again.
>>51479352
Stop replying to the shitposting children.
>>51479378
But you just did
He's doing exactly what he said. You fags keep putting him on a pedestal.
>>51479340
>if you want real discussion about Zen, just move over to Overclock.net
holy shit anon, thanks. Found the thread about Zen on there, so much better than the bullshit on /g/. I wonder why I even come to this place anymore, it's just full of dumb idiots writing dribble.
Steamroller is the first Bulldozer-derived chip from AMD that has been re-architectured. Piledriver was just a small fix here and there and we all know how much nicer it got.
Now imagine Steamroller:
>Compressed transistors by 30% (lower temps, less power usage)
>On top of that, lowered to 28nm in size
>Dynamically changing L2 cache that can give as much of itself to a module as it needs, IPC skyrockets when cores are given more L2 cache
>More L1/L3 cache
>2x the decode speed
>Clocks even higher than Piledriver (4ghz stock standard)
>being built by Jim Keller, the engineer that created the Athlon 64 in 2003
Kaveri is a bit off-topic but it will also be receiving Steamroller cores as well as being the first HSA-compatible chip on the planet. HSA is basically the next AMD64.
As far as we know, Steamroller will be priced the same. No more, no less.
Is your AM3+ ready? Most likely, because AMD doesn't greedily change sockets every month. A friend of mine bought an FX-4100 based machine. Guess what's going to work in his machine out of the box? Steamroller.
Is your anus prepared?
>>51479451
Nah waiting for AM4 boards desu for dat dere futureproofing
>>51479451
Steamroller had a smaller IPC uplift than Piledriver.
The 28nm Steamroller module is nearly the same size as the 32nm Bulldozer/Piledriver module due to added units and reorganization
Steamroller's cache latency massively increased compared to Piledriver
APUs do not have any L3 cache whatsoever
Kaveri APUs use Steamroller cores, Kaveri parts have been out for over a year.
It took a very long time to get binnings high enough to push average clocks north of 3.7ghz inside of a 95w TDP, the 28nm Kaveri chips do not clock as well as 32nm Piledriver based chips.
Kaveri APUs are on socket FM2+
Kaveri is only the first partially HSA compliant APU, Carrizo is the first chip that compliant with the HSA 1.0 final spec.
Carrizo is mobile only, its desktop counter part Bristol Ridge is AM4 only.
AM3+ is a totally dead socket and has been for years
Jim Keller did not work on the Steamroller arch whatsoever.
You are grossly mistaking Kaveri for Carrizo, but you're still hilariously wrong since mobile Carrizo launched months ago.
Basically you're a complete fuck up and a huge retard. Every single thing you just stated was wrong in one way or another.
I look forward to pushing your stupid shit in even further in the future.
>>51480165
Pic also related
I was fucking depressed but the amount of delusional faggots posting here made me cheer up and laugh
Never change amdfags
>>51479451
>all these cheap-looking 3rd class whores
Shilling for AMD during the the 939 x64 era and perhaps during the AM1-2 was warranted... but this is asinine.
What do you think they'll market the cores as?
Zen is a cool name, but they might just keep it as an internal codeword.
Maybe Phenom III, Maybe FX II ?
>>51479122
Because this is the last chance we give AMD. If zen flops we move to intel for good (at least for CPUs).
>>51480829
>we
>>51480969
I love you too anon-kun
>>51480829
I hope they make good zen apus. Maybe it will kill the need for low end GPUs for good now.
>>51481302
Yeah me too. AMD can obviously make 1337 APUs (ie PS4/xbone APU) but the problem is the memory for the APUs. Maybe even DDR4 won't be enough. I hope they allow zen APUs to use HBM.
>>51481092
Install gantee
>>51480800
Zen is the codename of the core architecture. Like Star, Husky, Clawhammer, Piledriver, etc.
Summit Ridge is the platform codename. Like Zambezi, Vishera, Deneb, etc.
How they'll brand the individual SKUs is still up in the air.
Reminder that this 348mm2 die would only be 104mm2~ scaled down to 14nm, and it would draw around 50 watts.
>>51480183
can this have a comparison to core2 to haswell pinned on?
AMD pulled this off, and got it to scale down to 15w with decent performance, using generic 28nm HPP.
I have hope for the future.
>>51478008
I'm disappointed that they're dropping cmt and choosing to no longer be the budget option.
>>51483919
>I'm disappointed that they're dropping cmt
There is no point in pursuing CMT when you have other means of increasing compute power per mm2 by shrinking core size independent from process node.
>>51483963
Here are 2 processors with a similar single thread performance, one is using smt and one is using cmt.
>>51479122
JIM FUCKING KELLER
Can you not grasp how Badass he is?
>>51479262
He ALWAYS LEAVES EARLY YOU STUPID FUCKS.
He's so God Damn good that he had other shit to do so he gets them on the track and moves on
>>51479262
Why should he stay any longer than he needs to? He's not a charity worker, he got hired at samsung after his work with AMD was finished.
>>51479451
Go back to 2013 :^)
>>51484185
That could not possibly be any more irrelevant, and it sounds like you literally do not understand what either of CMT or SMT actually are. Each are approaches to multithreading that aim to increase throughput for minimal added die area. CMT sacrifices individual core components and shares resources to have more smaller cores, or at least tries to. SMT uses AGUs and registers to tag ops and work them process them on a virtualized thread.
You can't directly say one is better than the other because they are nothing more than theories, you can only compare various implementations of each.
For reference:
The FX 8150/8350 have a 315mm2 32nm die.
That Lynnfield i7 has a 296mm2 45nm die.
Bulldozer/Piledriver offers lower performance per die area, so the point you tried to make is only working against you.
A 32nm Bulldozer/Piledriver module with 2MB of L2 cache is 30.9mm2.
Two 32nm Husky cores in Llano with 2MB of L2 cache is 32mm2, but these two cores have higher throughput than a module.
AMD managed to fit two cores into a very marginally smaller area, but performance flat out regressed, their implementation of CMT did nothing for them. It matter of factly did not accomplish the singular thing it was meant to accomplish.
When it comes to performance per clock the Steamroller based Kaveri is slower than Llano in a number of metrics, and the Steamroller module is almost the exact same size as the Bulldozer/Piledriver module despite being 28nm compared to 32nm.
http://www.anandtech.com/bench/product/399?vs=1291
>>51484328
http://www.anandtech.com/bench/product/399?vs=1270
Closer comparison.
Note that the Llano part does not have any turbo whatsoever.
The Kaveri chip here can turbo to 3.3ghz, 3.7ghz, or 3.8ghz depending on how many cores are being used. It has a 200mhz higher base clock than the Llano chip as well.
Multi core scaling is worse with Steamroller than with the K10.5 Husky cores in Llano.
AMD's CMT was an utter waste of money and R&D time.
>>51478008
the thread can't wait that long
>>51483212
>only 20% CPU cores
dropped
>>51484390
>>51484328
Don't forget: http://www.tomshardware.com/reviews/a10-5800k-a8-5600k-a6-5400k,3224-11.html
When the Trinity A10 5800k launched with a 3.8ghz and 4.2ghz turbo it was flat out slower than the 2.9ghz Llano A8 3850 in FPU heavy workloads.
>>51484507
It won in integer-based workloads though.
[spoiler]But still has a lower IPC :^)[/spoiler]
>>51485235
>It won in integer-based workloads though.
Yeah only do to its clock speed.
>But still has a lower IPC :^)
Exactly.
If they wanted performance they would have been better off continuing to develop the K10 core. The only saving grace of the last 5 years for AMD is that they developed an ungodly amount of power saving tech that they can apply to future arch and drastically improve perf/watt. Engineers have spent the last half decade of their lives studying a train wreck in motion. Thankfully they've learned from it, and AMD now has management making sound decisions.
>>51480165
I am not that guy but
>Steamroller had a smaller IPC uplift than Piledriver.
While true it wasnt ment to
>The 28nm Steamroller module is nearly the same size as the 32nm Bulldozer/Piledriver module due to added units and reorganization
Not true at all,
>Steamroller's cache latency massively increased compared to Piledriver
Where the fuck did you get this?
>APUs do not have any L3 cache whatsoever
>Kaveri APUs use Steamroller cores, Kaveri parts have been out for over a year.
>It took a very long time to get binnings high enough to push average clocks north of 3.7ghz inside of a 95w TDP, the 28nm Kaveri chips do not clock as well as 32nm Piledriver based chips.
The clocks were not low based on tdp. The 32nm soi process simply allowed better clocks.
>Kaveri APUs are on socket FM2+
>Kaveri is only the first partially HSA compliant APU, Carrizo is the first chip that compliant with the HSA 1.0 final spec.
>Carrizo is mobile only, its desktop counter part Bristol Ridge is AM4 only.
Why say 2 things that contradicts?
>AM3+ is a totally dead socket and has been for years
True
>Jim Keller did not work on the Steamroller arch whatsoever.
>You are grossly mistaking Kaveri for Carrizo, but you're still hilariously wrong since mobile Carrizo launched months ago.
>Basically you're a complete fuck up and a huge retard. Every single thing you just stated was wrong in one way or another.
He may be retarded but you aren't much better.
>>51484328
You are retarded.
>>51485302
You're quite literally a shit eating retard.
>>The 28nm Steamroller module is nearly the same size as the 32nm Bulldozer/Piledriver module due to added units and reorganization
>Not true at all,
The 28nm Steamroller module minus L2 is just under 18mm2. The 32nm Bulldozer/Piledriver module is 18mm2.
Despite the die shrink they are almost the same size.
The Excavator module decreased die area by about 24%, and the result is a module of 14.48mm2.
You not knowing this yet choosing to open your stupid fuck mouth is absolutely pathetic. You're a loud mouthed retard dipshit trying to argue with no ground to stand on.
>Where the fuck did you get this?
Dozens of reviews. AIDA64 results. L2 cache latency specifically took a massive hit.
>Why say 2 things that contradicts?
None of these statements contradict, you dumb fucking child.
When Kaveri was released only the provisional HSA spec existed. Carrizo is the first and only APU compliant with the final 1.0 HSA spec. Even AMD's marketing material reflects this.
The Steamroller design was finalized before Jim Keller began any real work at AMD. During their FAD earlier this year Lisa Su said herself that Keller first jumped on Excavator for Carrizo.
>>51485319
>I can't handle 100% empirical facts
Cute.
![snapshot_03.53_[2014.12.24_00.22.18].png](https://i.imgur.com/teyE4bE.jpg "snapshot_03.53_[2014.12.24_00.22.18].png")
>>51485235
you smelly /v/ermin, please die
When will the new zen apu come out? and do you think it will be better than the intel iris anons?
>>51485400
Their first Zen based APU is named Raven Ridge, and its due out Q1 2017. Just a few months after the Summit Ridge CPUs launch.
>>51485412
is it revealed which socket?
>>51485435
AM4, same as Summit Ridge.
AMD is only going with one main socket from now on.
>>51484328
>A 32nm Bulldozer/Piledriver module with 2MB of L2 cache is 30.9mm2.
>Two 32nm Husky cores in Llano with 2MB of L2 cache is 32mm2, but these two cores have higher throughput than a module.
How did they ever think this could work?
I mean people are sitting around designing a CPU and a guy says "how about we make one huge CMT module that takes up as much space as two entire cores" How could anyone even think that's a good idea?
>>51485436
thats great
i might actually buy amd in the future because of that
>>51485459
The original Bulldozer design as was first presented to the public at a shareholder meeting was a totally different design from what released in 2011. It was a larger module, and AMD was absolutely explicit that it was aimed at being their highest *serial* throughput architecture ever in addition to being their highest parallel throughput CPU.
They thought they could do it all with this one arch. They may have been able to accomplish that if management actually dedicated the required time and money instead of twiddling their thumbs. It cannot be stressed enough how much poor management can sabotage a company.
The most skilled and visionary engineers in the world will be utterly crippled by upper management if they don't understand what is actually required to deliver what they're asking for.
Is AMD the company in dilbert?
Is the official unveiling today or something?
Will Summit Ridge have an integrated gpu?
>>51485578
Summit Ridge is CPU only.
>>51479220
Well, thats something alright
>>51485497
How could they have made it work? Did it have more FPU hardware originally? If FPU was double or even bigger I guess it could switch between sharing it for multithreaded situations or have 1 thread with huge FPU ability. Even then though, it was so focused on high clock speeds it could never reach that it may not have changed much.
>>51485374
no
>>51485657
Off the top of my head I remember the integer cores being wider. Of course this meant more die area, and I'm sure some genius eventually decided to make them smaller to try and salvage any possible CMT benefit and reduce power consumption. Keep in mind that server parts are targeting 2-3ghz(lower power per core) since they're dealing with much more parallel workloads. Desktops not so much.
I have no idea what they ever planned for the FlexFPU aside from it not being used much, which they were wrong about in a ton of cases. It might have been a pure gamble that they took and lost on. Lots of server workloads re primarily integer bound, and GPUs are used more than ever in enterprise/HPC, but larger vector ops are still heavily used in HPC as well. The FlexFPU was a poor choice no matter how you look at it.
To bad they finally noticed that employees were posting interesting stuff in their linkedin profiles and stopped that. There wont be much info on zen until they want to...
>Successfully implemented Zen(znver1) hybrid branch prediction algorithm. Algorithm works by keeping smith counter to choose between stable quantum bogosort and PigeonRank algorithms.
https://www.linkedin.com/in/steve-havlir-714b337
>>51485915
> keeping smith counter to choose between stable quantum bogosort and PigeonRank algorithms.
wut
>>51485915
Oh holy fucking tits. I cannot wait until wccftech or some other shitbait site posts an article about this.
Pure comedy gold.
Everyone at AMD knows that people datamine linkdin profiles. They continuously put bullshit in their profiles to throw people off and make inside jokes. This is one of those cases.
>>51485930
It's bullshit, obviously. But earlier you found interesting info in some of the employees profiles before AMD officialy released it. Like tape out info, what exactly they designed (HBM MC for example, before AMD released info in Fiji) and so on
>>51485939
Never noticed that earlier. There was some interesting info in some of those profiles though, i remember "Leaked" details on excavator, Hawaii and Fiji from there
>>51485930
AMD is building a quantum computer to rank the status of every pigeon in the world in accordance of utility for basic remote cerebral functionality.
tldr world domination via army of pigeons
Can any mayor hardware company buy AMD? Like what happened with Nokia
>>51485971
Samsung will buy them in 2018
>>51485969
Imagine swarms of pidgeons attacking server farms that use Intel/nvidia/IBM/Fujitsu hardware
>>51485971
If AMD wanted to be bought, then yes. They have a treasure trove of IP that could be applied all over the place, not to mention the value of their operating business, and staff.
They're not looking to be acquired by anyone though. Their long term debt is nothing in the grand scheme of things. They could pay off that $2,000,000,000 in a single good year.
>>51485977
Asia is already to dangerous, no way the US would allow that. Before they go bankrupt they'll throw them a bone with some multi billion supercomputer deal
>>51479262
Keller is a hardware mercenary. He just designed Zen architecture, now is time to produce prototypes, his job is done
IS THIS GOING TO SAVE AMD /g/?
I'm going to buy AMD stocks
I'm going to buy AMD socks
I'm going to buy AMD sticks
Reminder that Intel fags will relentlessly spam rigged CineBench "benchmarks"
>>51486285
I WILL WEAR ONE AMD SOCK
AND TWO INTEL SOCKS
I'm going to buy AMD dicks
>>51486448
rofl, that killed me
its like the nvidia 980 and r9 290x benchmark
1fps difference
>>51485364
You are retarded. I don't have time to argue with someone this retarded.
I WANT TO BUY AMD
>>51488924
Sounds like someone presented facts and you got BTFO, anon.
>>51485364
This post is off on one point.
Actual die area for Steamroller is 18.61mm^2. The cache is denser but the module grew in total size.
Anyone have the full version of this image? Saw it on one of the threads a couple days ago so I know a good version exists
>>51492428
Dumping more "get hyped" images
>>51492442
>>51492428
>secure boot
>crypto processor
sounds shady to me, i hope it won't go full retared botnet like can't install linux or intel management anti theft spying on hardware level.
>>51492463
>>51492479
Original version didn't have that if I remember correctly, think someone tacked that on
Though this is their rumored full soc
>>51492442
This is completely fake.
>>51492463
This is also completely fake. Worse yet its an edit of someone else's fake.
>>51492428
Its from Fuddzilla, and it looks like they made it themselves.
>>51492442
>>51492428
>>51492463
Seriously people. It's not hard to stop with the fake slides, FAD is well past and we have the real slides to hand:
https://community.amd.com/community/amd-corporate/blog/2015/05/12/amd-financial-analyst-day-2015
http://ir.amd.com/phoenix.zhtml?c=74093&p=irol-analystday
Not hard to tell people what is confirmed, rather than constantly spreading misinformation via slides that have been proven to be fake multiple times now.
>>51492593
>>51492673
I have new leaked slides for you guys.
>>51492708
HOLY FUGG
SO I CAN RULLY FUG ME POOKAHOO!
>>51492673
>>51492593
Yeah but the fake leaked slides get me more hyped
Ffs look at the real slides
It literally is an autism awareness chip
>>51492752
He's not lying
>>51492752
The slides were meant for investors. Not tech reviewers, and they are targeted at their intelligence level and understanding as a result.
Getting you hyped off of misinformation will lead to nothing but disappointment.
But we have been told that Zen has met the 40% IPC increase as expected. Now all we need to know, are some solid numbers, TDP, Power Draw. Hell, something, anything! Die size even, so we can start speculating what GPU they could squeeze on there when the APUs start rolling. AMD are the best at hyping products, and they don't even need to say anything.
There is so much potential wrapped up in that company, but years of poor management has left it in a financial mess. All they need is for Zen to be "Ok" compared to skylake, and that will restore faith and cashflow in the company, and they can get their shit in order.
If it is the rebirth of the Athlon 64, then it could be the deathblow for Intel's public image, as they will surely bribe OEMs again, and the EU will most certainly not take 8 years to reach a decision, ending in a US$1,000,000,000 fine this time around.
>>51492851
>There is so much potential wrapped up in that company
everybody who could have made it happen left ages ago
they are now employed at samsung, apple, qualcomm, intel etc
>>51492868
They have Su.
They have Koduri.
Some of the biggest names in their respective fields. It will not take much more than cashflow to rehire some of their old CPU teams from Apple and other companies.
Hell, given that Zen is somewhat a success, they could bring Keller back for more work.
The one thing AMD does not lack, is potential.
>>51479451
the short one.
>>51492868
They were the first to hbm
>>51492939
They created HBM.
>>51492939
They co-invented the damned thing, I should hope they were the first to use it.
It would be interesting to know just what IP they hold on HBM.
>>51492908
hiring a few rockstar names won't change reality unfortnuately
>>51479446
Look for the user Looncraz on overclock.net and check his posts about Zen. I know fuck all about the inner workings of a CPU, but I'm pretty sure this guy knows his shit.
>>51492971
They likely have tons considering pic related is from 2010~
>>51492996
Fiji die and package circa 2015 for comparison
>>51492975
No, but it gives them something to build off of.
Su is an engineer, and now the CEO. She knows what needs to be done to push a quality chip out the door, and we can bet that she's seeing that Zen gets all the attention it deserves. And more.
Koduri is nothing short of a genius, having him onboard, and Zen on the way, and HBM here, we can bet he'll do some freaky shit for their APUs.
The shear fact that Rory Reed is gone should be enough reason for their previous CPU teams to come flooding back if nothing else.
Su has held to her word so far.
She's given the slackers and the marketting teams the sack, she's got investors interested again.
>>51479262
protip: amd will let the indians tweak this new core for the next 10 years. just like k10 and the current intel cores.
>>51493056
A team has been working on Zen+ since the beginning of the year. Before Zen+ reaches market a team will already be working on the successor to that.
They're going to be really aggressive with upgrades, keeping all those technicians busy.
![23619-1280[1].jpg](https://i.imgur.com/cTesnA1.jpg "23619-1280[1].jpg")
>>51493076
>Cutting off Dobberpuhl
forwhatpurpose.jpg
>>51493104
Because we prefer to stare at shrek
>>51492851
I've been playing with some figures trying to guess what die sizes Raven Ridge and Summit Ridge will have.
If the Excavator module scaled directly down to 14nm It would be about 4.4mm2. If a single Zen core is 40% larger than this entire module then it would be right around 6mm2 in silicon.
Factor in uncore, L2, L3, memory PHY, IO, and all the rest and I'd venture a guess of Summit Ridge being somewhere around 125mm2 to 150mm2 at most.
The Raven Ridge APU will likely be a bit smaller, I'd say right around 100mm2. The PS4's APU above has 20CU, 8 Jaguar cores with some interconnect fabric between the clusters, a 256bit memory PHY, and misc other stuff. That all scaled down to 14nm would be about 105mm2.
4 Zen cores with L2 and an IGP with 20CU should fit inside of 100mm2, and the IGP may not even be nearly that big so the die could be substantially smaller still.
>>51493626
Heres to hoping they have at least one HBM module on package.
A single HBM1 module can provide up to 128GB/s bandwidth.
A single HBM2 module can provide up to 256GB/s bandwidth.
>>51494092
It would be pretty stupid not to try to incorporate hbm
>>51494303
They could use it as a L4 cache, like jewtel did on their Iris Pro chips
>>51494486
Being off die its going to have latency far too high for that, and APUs aren't going to have an L3 to begin with so any L4 is out of the question to begin with.
If they use HBM on their APUs its likely going to be dedicated VRAM, and a work space for compute with preemptive context switching.
>>51494579
Feel free to call me an idiot if I'm wrong but didn't amd place their hbm off die for their fury cards without increasing latency significantly?
>>51494644
VRAM is never on-die.
>>51494644
HBM is being used as VRAM on the Fury cards, it does not function the same was a CPU cache.
CPU caches use varying times of SRAM cells, and they are absurdly fast. No type of DRAM compares, especially not when you drastically increase wire length.
>>51494677
>>51494677
Ok, I seem not to have known the difference between dram and sram
>>51494677
>>51494675
Looks like this patent suggests they will use hbm on all apus as L3 cache
http://semiaccurate.com/forums/showthread.php?t=8125
Why only 4GB of HBM for Fury X
wont that shit fucking bottleneck or sum shit.
>>51479451
Who are the dick craving cumbucket to the left?
>>51495152
No
At those speeds you don't need to have that much
>>51495152
Ha ha no
>>51495114
>jeff rigby
Hoo boy.
AMD has patents for many things, that does not mean they use all of them, or that any of their latest patents are even related to upcoming products.
Latency wise HBM is so slow that it would incur a huge performance penalty if the CPU tried to use it as an L3 cache.
DRAM and SRAM are magnitudes apart.
>>51495295
How high IS the latency on HBM, anyway?
>>51495371
256 GB/s+
>>51495407
That's bandwidth, not latency.
>>51495371
Its comparable to GDDR5, its just regular DRAM dies all attached together sitting atop a die of control logic.
L1 can have a latency around 1ns, or even faster. L2 typically is significantly slower, could be 10ns~ and L3 is slower still.
A DRAM slide in either GDDR5 or HBM can have a raw latency of over 40ns, and the memory controller is going to address this is an entirely different way from its internal caches so they're worlds apart.
Its worth noting that SRAM also has absurdly high bandwidths in addition to its ridiculously low latency.
>>51495407
Go back to /v/
>>51495503
No you
