AMD CPU speculation... and expert conjecture

[jdwii]

Broadwell seems to throttle around 62c vs the 80c of Haswell.

http://www.phoronix.com/scan.php?page=article&item=9way-nehalem-broadwell&num=2

See the chart at the bottom for LAME MP3 Encoding v3.99.3.

That's not good at all, if the desktop version even gets released i hope they don't throttle it at that temp. 80C is fine for me but 62C that's Piledriver-Bulldozer max temps.

 

[noob2222]

^^

Thats the downside of thermal density. Its not as simple of what exact temperature problems arise, but the amount of time it takes to move heat from the source to the heatsink. The smaller the object in question, the less time you have. Our general perception is "how hot is it" while on the inside its a much faster rise and fall.

Juan, your defeating yourself. 50% less area from 14nm to 7 nm, increase clocks by 15% and only lower voltage by a mere 5% ... just think about that. Hint: its more than double thermal density.

 

[juanrga]

Juan, your defeating yourself. 50% less area from 14nm to 7 nm, increase clocks by 15% and only lower voltage by a mere 5% ... just think about that. Hint: its more than double thermal density.

First, 14nm --> 7nm is not "50% less area". Your math fails.
Second, power is proportional to area, less area implies less power.
Third, 5% lower voltage is only from silicon scaling. I mentioned there are additional voltage reduction from using NTV techniques. I gave you one link to a paper on the topic...

 

[juanrga]

http://wccftech.com/amd-godavri-apu-kaveri-refresh-coming-june-2015-carrizo-desktop-apu-arrive-1h-2016/

Some relevant quotes from his AMD source.

Godavari change will be even less exciting or groundbreaking than Trinity (to) Richland was.

This coincides with what my source said me about Kaveri refresh and I quoted some pages ago.

AMD actually had a desktop Carrizo planned, something that was later canceled due to management’s decision. It was originally designed to be put onto the FM2+ socket and FP4 packaged and relieve the Kaveri, being a true successor to the older APU. However, to cut costs and following the profit motive (something rather understandable) they decided to reiterate Kaveri with a refresh named Godavri.

[...]

Interestingly he also mentioned that Carrizo still might end up on the desktop segment and AMD is currently debating this very question. If they do decide to go forward with a Carrizo Desktop, it won’t arrive till H1 2016.

Thus my early leak of Carrizo for the desktop was based in real data. Some people even said that the roadmap was fake.

So basically, 2015 and 2016 will still be host to the 28nm process, even though I am pretty sure all of us have started to hate it by now.

This explains why AMD's Kumar said that 28nm would be used for 2016 products. Some analysts said that Kumar misspoke. But now it is confirmed that he didn't and that AMD will rely on 28nm node for a long period.

 

[noob2222]

Sorry, just quick typing instead of focusing on miniscule details. Your right, 14nm to 5nm is way more than a 50% reduction.

Yes, transistor power is relational to the area, its not however linear.

Lets look at that NTV ...

Generally, NTV is an ideal fit for HPC workloads and works very well for graphics, but not general purpose CPUs.

Try again

 

[juanrga]

Sorry, just quick typing instead of focusing on miniscule details. Your right, 14nm to 5nm is way more than a 50% reduction.

Yes, transistor power is relational to the area, its not however linear.

Lets look at that NTV ...

Generally, NTV is an ideal fit for HPC workloads and works very well for graphics, but not general purpose CPUs.

Try again

The discussion was "14nm to 7 nm" not "14nm to 5nm".

Power is linear with area.

Sorry but that quote from Kanter is not correct. The technique can be used in general purpose CPUs. There are working prototypes of mobile CPUs! If you have some serious question PM me.

 

[juanrga]

Some new data about Carrizo

The IPC gain is significantly lower than expected, but the performance gain falls within the expected. The second slide shows what I commented before about Excavator doesn't scaling up well in frequency. Above the normalized frequency of "1" the use of HDL libraries makes that Excavator hits a frequency 'wall' compared to Kaveri. This is the explanation of why Carrizos above 35W were canceled.

 

[etayorius]

Some new data about Carrizo

The IPC gain is significantly lower than expected, but the performance gain falls within the expected. The second slide shows what I commented before about Excavator doesn't scaling up well in frequency. Above the normalized frequency of "1" the use of HDL libraries makes that Excavator hits a frequency 'wall' compared to Kaveri. This is the explanation of why Carrizos above 35W were canceled.

No wonder why they cancelled Desktop... 5% IPC gains is the best they could do to maintain efficiency.

So, how much faster was Richland compared to trinity and Kaveri from richland? if these are the best AMD can do, it will take many MANY generations to match Haswell.

Zen has zero chance of even competing with Haswell, let alone Skylake that will at least be 10% faster than Haswell.

 

[juanrga]

Some more slides

 

[esrever]

No wonder why they cancelled Desktop... 5% IPC gains is the best they could do to maintain efficiency.

So, how much faster was Richland compared to trinity and Kaveri from richland? if these are the best AMD can do, it will take many MANY generations to match Haswell.

Zen has zero chance of even competing with Haswell, let alone Skylake that will at least be 10% faster than Haswell.

Excavator is not designed for more CPU performance. The core is smaller and has even less cache than kaveri. Cutting the L2 down just to fit more power adjustment hardware. The power consumption of these cores are most efficient at 5-10W. IPC doesn't matter at that level if they could use less power and hit higher frequencies in a power constrained envelope. Increasing IPC would mean a bigger core and that won't be economical. AMD never planned for excavator desktop for the looks of it. The design is purely to get as much performance with as few transistors as they can. At the mobile thermal constrained conditions it looks like carrizo is a big win for what AMD has to work with. Maybe a 20-30% CPU performance increase and a major GPU performance increase while using less power will be great for a laptop.

Anyways, Richland was 0% faster than trinity in desktop. It only had power gating enabled and memory optimizations.

If Zen is designed for the desktop, it will be a completely different core and different design philosophies. If you are talking about single core performance, probably not going to reach haswell but multi core performance will be pretty easy.

 

[juanrga]

Excavator is not designed for more CPU performance. The core is smaller and has even less cache than kaveri. Cutting the L2 down just to fit more power adjustment hardware.

Cutting the L2 down to one half also saves power.

AMD never planned for excavator desktop for the looks of it.

Really?

What happens is that the original performance-optimized Excavator that would be used in FX and Opteron CPUs was canceled some time ago and replaced by the power-optimized Excavator used in Carrizo. Note that, as mentioned above, AMD is actually discussing if will release Excavator on the desktop in 2016 or not.

Anyways, Richland was 0% faster than trinity in desktop. It only had power gating enabled and memory optimizations.

A fully functional Resonant Clock Mesh did help Richland to hit higher frequencies. Aka it was faster than Trinity.

 

[esrever]

I am not going to argue with your Jaun. You can post all you like. You can pull all the stuff out you like. Stop replying to my posts.

 

[jdwii]

No not when you keep talking about things that Amd clearly says otherwise^^^^^^^^^^

You don't get to say whatever you want without anyone questioning you this isn't a dictatorship forum. Just like i can't post anything i want. Juan or anyone else has every right to reply to any post they see fit with respect.

Amd intended to have Excavator on the desktop its already documented

 

[de5_Roy]

let's witness warmandcasualchattingonparrottech.com rumor
AMD’s x86 Nolan APU will be Fabricated on the 28nm Process not 20nm
http://wccftech.com/amds-x86-nolan-apu-fabricated-28nm-process-20nm/
"basically, what we can really confirm is that a future amd semiconductor device will use a fabrication process in the future. our first ever exclusive!"

 

[Cazalan]

So looking at all the Carrizo slides they're saying 5% IPC and 10% faster clocks at the same watts. ~15.5% combined. Not sure if that is just base clocks or also turbo.

FX-7600P (2.7/3.6) to equivalent FX-86xx (~3.1/~4.1) or (~3.1/~3.6)

Same die size means they've cut the cost and total platform power by removing the need for the additional PCH. That price reduction could get more OEMs on board. We'll see.

 

[juanrga]

So looking at all the Carrizo slides they're saying 5% IPC and 10% faster clocks at the same watts. ~15.5% combined. Not sure if that is just base clocks or also turbo.

FX-7600P (2.7/3.6) to equivalent FX-86xx (~3.1/~4.1) or (~3.1/~3.6)

Same die size means they've cut the cost and total platform power by removing the need for the additional PCH. That price reduction could get more OEMs on board. We'll see.

Look at the crossover point in the frequency/power slide. Carrizo is optimized for low power. The low 15W models will have enough room for frequency gains, the top 35W models will not. The top 35W Carrizo SoC will be probably clocked at 2.8GHz (base) and 3.4GHz (turbo). However sustainable turbo must be about 2.9GHz, which is less than the 3.6GHz turbo of the FX-7600P.

 

[juanrga]

Broadwell seems to throttle around 62c vs the 80c of Haswell.

http://www.phoronix.com/scan.php?page=article&item=9way-nehalem-broadwell&num=2

See the chart at the bottom for LAME MP3 Encoding v3.99.3.

^^

Thats the downside of thermal density. Its not as simple of what exact temperature problems arise, but the amount of time it takes to move heat from the source to the heatsink. The smaller the object in question, the less time you have. Our general perception is "how hot is it" while on the inside its a much faster rise and fall.

Broadwell works as expected even in seriously constrained thermal envelopes:

The Intel NUC5i5RYK provided us with the opportunity to take a look at what Broadwell-U can deliver when coupled with a motherboard providing premium features. The migration from 22nm to 14nm has allowed for higher base clocks while maintaining the same power envelop.

http://www.anandtech.com/show/8986/intel-nuc5i5ryk-review-a-broadwellu-ucffpc-for-enthusiasts/8

 

[noob2222]

Sorry, just quick typing instead of focusing on miniscule details. Your right, 14nm to 5nm is way more than a 50% reduction.

Yes, transistor power is relational to the area, its not however linear.

Lets look at that NTV ...

Generally, NTV is an ideal fit for HPC workloads and works very well for graphics, but not general purpose CPUs.

Try again

The discussion was "14nm to 7 nm" not "14nm to 5nm".

Power is linear with area.

Sorry but that quote from Kanter is not correct. The technique can be used in general purpose CPUs. There are working prototypes of mobile CPUs! If you have some serious question PM me.

juan, you wont answer questions because you can't so instead you focus on "OMFG A TYPO, YOU MUST BE RETARED"

power is not linear. there is this unknown to juan argument phenomenomonemom called leakage, wich doesn't exist when dealing with theorytical hypothesis of the world in its perfect flawless form.

http://www.realworldtech.com/near-threshold-voltage/4/

ya ... ntv ... double transistor count for ultra low frequency operational power refunction.

Spoiler

I included some more typos so you can avoid the topic all together

broadwell works as expected, just had to cancel desktop and quad cores and operating > 3 ghz

 

[blackkstar]

Broadwell seems to throttle around 62c vs the 80c of Haswell.

http://www.phoronix.com/scan.php?page=article&item=9way-nehalem-broadwell&num=2

See the chart at the bottom for LAME MP3 Encoding v3.99.3.

lol, this review is gold.

Broadwell throttles way earlier than anything else
It uses less power!

Obviously, it's throttling itself. Chips use less power when throttling.

Go to raw performance numbers.
http://www.phoronix.com/scan.php?page=article&item=9way-nehalem-broadwell&num=7

Broadwell losing to Haswell in raw performance the majority of the time. It's similar architecture to Haswell except on 14nm instead of 22nm. Broadwell even losing to Ivy Bridge in Macbook Retina.

b-b-b-but check out these marketing slides and this academic paper, 4.3ghz base clock chips from Intel are coming as soon as they keep shrinking their nodes! Intel 14nm is the greatest thing ever way better than 22nm AMD IS FINISHED HAHAHAHAHAHHAHAHA. IF INTEL CAN KEEP THIS UP BY 7NM WE WILL FINALLY HAVE THOSE 10GHZ NETBURST NEHALEMS JUST LIKE THE OLD ACADEMIC PAPERS AND MARKETING SLIDES SAID! SAMSUNG 14NM CANT TOUCH THIS!

I feel sorry some of you invest so much time into this stuff only to be so wrong all the time.

 

[jdwii]

http://www.anandtech.com/show/8986/intel-nuc5i5ryk-review-a-broadwellu-ucffpc-for-enthusiasts/7

This review claims it hit 80C or so.

 

[griptwister]

We have some actual "CPU" news guys. I don't know about you guys, but if this is real, I'm excited!

http://www.kitguru.net/components/cpu/anton-shilov/amds-zen-could-sport-intels-skylake-technologies-new-512-bit-fpu-rumours/

 

[faye__kane]

AMD IS FINISHED HAHAHAHAHAHHAHAHA.

And we'll all be paying a grand for mid-range processors ha ha ha. And you can forget about Intel dumping a fortune into furious R&D when they're the only fab on the block ha ha ha.

I ain't laughin'.

--faye kane ♀ girl brain

 

[faye__kane]

We have some actual "CPU" news guys. I don't know about you guys, but if this is real, I'm excited!

yEAH, i'M EXCITED TOO goddam capslock!! Does this mean that AMD changed their mind about rolling over for Intel and only making dinky little A-series processors with built-in video? Does it mean that the 9590 won't be the last good AMD CPU?

Because if not, I'm prepared to keep my wonderful 9590 for the next ten years if the only other choice is hand over money to those bums at Intel.

Did you know that even after being caught at it, their C compiler STILL turns off CPU optimizations for AMD chips? All they did was post a disclaimer at the bottom of the download page. And tellingly, the disclaimer isn't text, but a graphic image of text, so you can't copy and paste it if you're outraged.

How many developers¹ don't know this and use that compiler? How many carefully-written applications by innocent, trusting programmers are being damaged on purpose by Intel so that they can hurt AMD at the expense of the users?

That doesn't just say something about Intel; it says something about crapitalism.

-- faye kane ♀ girl brain
__________
¹Ballmer killed that word for me. It's hard for me to say it now without visualizing that incompetent frat-boy marketing bozo jumping and screaming.

 

[faye__kane]

We have some actual "CPU" news guys. I don't know about you guys, but if this is real, I'm excited! http://www.kitguru.net/components/cpu/anton-shilov/amds-zen-could-sport-intels-skylake-technologies-new-512-bit-fpu-rumours/

Cool, but W(ho)TF needs 512 bit-wide arithmetic? What's the app? Can't be multimedia; that's already done in parallel on the GPU. And there can't be enough demand for it in science to warrant engineering an entire new FPU just for that.

-flk

 

[faye__kane]

> FX 8350 @ 5.124ghz

Blackstar, are you running an 8350 at 5.1 GHz continually, or is that a peak performance thing? If the former, what's the under-load temp?

-faith