AMD Carrizo Pre-release thread

[Enigmoid]

Why are you comparing to a hybrid crossfire kaveri notebook.

Yeah low is 1024 x 726 for TR but that really is all that is playable. Apples to apples. And further again, neither can play at native so a stronger gpu really doesn't do anything.

19W graphics are much less bottlenecked than the desktop simply because they run at such low speeds that you don't need as much bandwidth to power the low performance. Either way, increasing the bandwidth increases the effective load so if TDP limited performance can only rise to full TDP.

Why are we still using the Yoga pro? You are the same person who argued months ago about eliminating some of the temash and kabini notebooks due to high power usage and crappy optimization. Not to mention that the yoga pro is rocking a power hungry screen and Notebookcheck gets their max values with everything cranked up.

If you actually read how notebookcheck does its rating its subjective, based on product and class (the competition). Which is why I don't pay attention to it nor have I ever attempted to conclude anything from it.

 

[geoxile]

People keep saying the Yoga's 3200 x 1800 panel is power hungry but I've yet to see it proven beyond a shadow of a doubt. Does anyone have specific power measurements or a spec sheet? Because it's supposed to be manufactured using IGZO, which grants significantly reduced power consumption over traditional amorphous silicon

 

[maarten12100]

Why are you comparing to a hybrid crossfire kaveri notebook.

Yeah low is 1024 x 726 for TR but that really is all that is playable. Apples to apples. And further again, neither can play at native so a stronger gpu really doesn't do anything.

19W graphics are much less bottlenecked than the desktop simply because they run at such low speeds that you don't need as much bandwidth to power the low performance. Either way, increasing the bandwidth increases the effective load so if TDP limited performance can only rise to full TDP.

Why are we still using the Yoga pro? You are the same person who argued months ago about eliminating some of the temash and kabini notebooks due to high power usage and crappy optimization. Not to mention that the yoga pro is rocking a power hungry screen and Notebookcheck gets their max values with everything cranked up.

If you actually read how notebookcheck does its rating its subjective, based on product and class (the competition). Which is why I don't pay attention to it nor have I ever attempted to conclude anything from it.

I was comparing to that to show that while something is twice as powerful in terms of graphics performance it basically receives the same score on that subject. 66-68%

It is bandwidth bottlenecked why do you think low end solutions from Nvidia come with GDDR5. Higher latency higher bandwidth GDDR5 removes congestions and increases the efficiency over DDR3 the energy consumption per unit of performance will go down.

Nobody uses the Yogo pro unless they are complete fools. It is a crappy made products with a crappy made chip. "a power hungry screen"
proof please :|
I'm quite sure pushing those pixels consumes more than the actual screen backlight and tcon combined.

People keep saying the Yoga's 3200 x 1800 panel is power hungry but I've yet to see it proven beyond a shadow of a doubt. Does anyone have specific power measurements or a spec sheet? Because it's supposed to be manufactured using IGZO, which grants significantly reduced power consumption over traditional amorphous silicon

yeah

 

[maarten12100]

https://www.youtube.com/watch?v=I6zX2IqBI7A

 

[mrmt]

http://ir.amd.com/phoenix.zhtml?c=74093&p=irol-newsArticle&ID=1991472

2015 AMD Mobile Roadmap Adds "Carrizo" and "Carrizo-L" SoCs to APU Lineup
SINGAPORE -- (Marketwired) -- 11/20/14 -- AMD (NYSE: AMD) today at its Future of Compute event announced the addition of its first high performance system-on-a-chip (SoC), codenamed "Carrizo", and a mainstream SoC codenamed "Carrizo-L" as part of the company's 2015 AMD Mobile APU family roadmap.

(...)

The flagship "Carrizo" processor will integrate the new x86 CPU core codenamed "Excavator" with next generation AMD Radeon™ graphics in the world's first Heterogeneous Systems Architecture (HSA) 1.0 compliant SoC. The "Carrizo-L" SoC integrates the CPU codenamed "Puma+" with AMD Radeon™ R-Series GCN GPUs and is intended for mainstream configurations

In a video linked in the announcement (https://www.youtube.com/watch?v=I6zX2IqBI7A&feature=youtu.be) AMD John Byrne is saying that both Carrizo and Carrizo-L share the same platform and same packaging. I wonder this is the reason for Carrizo being mobile-only.

 

[inf64]

Carrizo will come to FM2+ socket also.

 

[NTMBK]

Wait, Carrizo-L is Puma+?! Whaaaaat.

Also, extremely interesting that the post only discusses the "2015 Mobile APUs" and says that Carrizo will be coming to "laptop and All-in-One systems". Is it really still coming to FM2+, I wonder?

 

[Abwx]

Digitimes were right when they wrote that Carrizo would replace both Kaveri and Beema in the mobile segment, it s just that they didnt specify that there was two chips, hence the skepticism at the time.

Wait, Carrizo-L is Puma+?! Whaaaaat.

Also, extremely interesting that the post only discusses the "2015 Mobile APUs" and says that Carrizo will be coming to "laptop and All-in-One systems". Is it really still coming to FM2+, I wonder?

The Excavator equipped version will undoubtly make its way in FM2+, mobile versions can use whatever other socket or BGA.

 

[NTMBK]

The Excavator equipped version will undoubtly make its way in FM2+, mobile versions can use whatever other socket or BGA.

Why? If it is limited to a 65W TDP, and cannot have either DDR4 or HBM in the FM2+ socket, what is the real benefit over the existing Kaveri chip?

 

[inf64]

New model number

 

[NTMBK]

Slides up at http://www.hardware.fr/news/13971/amd-carrizo-carrizo-l-mi-2015.html

Both are on FP4, so dual channel for Carrizo-L is confirmed. Should help it compete against the Bay Trail Celeron/Pentium parts.

 

[ShintaiDK]

And no 20nm on the roadmap.

But its good they address the memory bandwidth issue for the small cores.

However besides the FCH integration, the chip looks like a step back.

 

[NTMBK]

And no 20nm on the roadmap.

But its good they address the memory bandwidth issue for the small cores.

20nm is confirmed for later in 2015 though- it was on the leaked Embedded roadmap that showed up a few weeks ago.

 

[Abwx]

Why? If it is limited to a 65W TDP, and cannot have either DDR4 or HBM in the FM2+ socket, what is the real benefit over the existing Kaveri chip?

Better CPU perf at 65-45W, theses are obviously the max TDP sought by OEMs, Kaveri is already essentialy a 65W chip even in its 95W variants, hope we ll get harvested dies using half functional memory controler, would be a nice upgrade for the AM1 plateforms, although the logic is that Beema will be used preferably for such a purpose.

 

[ShintaiDK]

20nm is confirmed for later in 2015 though- it was on the leaked Embedded roadmap that showed up a few weeks ago.

I assume you refer to the G series Puma or A57 based chips. However they seem to be embedded only. The R series are still 28nm for example. And it seems all non embedded are 28nm.

 

[NTMBK]

I assume you refer to the G series Puma or A57 based chips. However they seem to be embedded only. The R series are still 28nm for example. And it seems all non embedded are 28nm.

They are also confirmed for non-embedded platforms, with Android explicitly listed as a target for the A57 part in other presentations:

And let's face it, AMD's embedded business is hardly big enough to justify making two whole new SoCs just for that one market They'll reuse it in consumer, just like the previous G-series parts. (In fact I'd argue that the G-series is a reuse of consumer parts, not the other way around.) Given how AMD is scrabbling for market share right now, they'll try to get as many sales as they can out of their R&D costs.

 

[Erenhardt]

Better CPU perf at 65-45W, theses are obviously the max TDP sought by OEMs, Kaveri is already essentialy a 65W chip even in its 95W variants, hope we ll get harvested dies using half functional memory controler, would be a nice upgrade for the AM1 plateforms, although the logic is that Beema will be used preferably for such a purpose.

Yea... They got me on that 95W. I had blamed the tower cooler for not properly transfering heat from my kaveri.
The reported temps were way off, showing its overheating, but the radiator was not getting hot and the air blowing through fins wasn't heating up when stress testing. After testing the cooler with boiling water to see if heatpipes work, I blamed the mounting for not pushing cooler base down to the IHS. I tried to polish and flatten the base of cooler - which was already pretty flat. Nothing worked.

But then, I decided to ignore temperature readings and overclocked it. How happy I was to feel the warmth coming through the exhaust part of the cooler... It finally made the pins warm to the level fx6300 does...

I got the cooler for almost free (a rabate). Now I have and overkill cooling that doesn't increase fan speed from idle...

I don't like those low TDP parts... You never know if its working or not. I wouldn't notice if that kaveri is runing cryptomining virus mining $ for someone 24/7

 

[Shehriazad]

Why? If it is limited to a 65W TDP, and cannot have either DDR4 or HBM in the FM2+ socket, what is the real benefit over the existing Kaveri chip?

65W TDP is literally the only benefit you will see on the CPU side (gaming performance of the CPU may take yet another hit down due to lower MHZ and even smaller cache)

On the GPU side...the architecture and delta compression might make up for the bottleneck...but that's about it. (A GPU can only take you so far if it's limited by DDR3)

 

[Erenhardt]

65W TDP is literally the only benefit you will see on the CPU side (gaming performance of the CPU may take yet another hit down due to lower MHZ and even smaller cache)

Hold on a minute...
... how did word "another" made its way to your sentence.

 

[NTMBK]

(gaming performance of the CPU may take yet another hit down due to lower MHZ and even smaller cache)

Cutting the cache size also improves latency. There's a reason why Haswell only uses 256KB L2 caches. (Though this is of course then backed up with 2-8MB of L3.)

 

[Shehriazad]

Hold on a minute...
... how did word "another" made its way to your sentence.

if you compare Kaveri and Richland head to head...in a lot of gaming benchmarks it is either equal or LOWER than Richland (higher IPC does not manage to make up for the lower clock as well as it would have to). It also doesn't overclock as well generally...unless you get some super cherry pick.

Cutting the cache size also improves latency. There's a reason why Haswell only uses 256KB L2 caches. (Though this is of course then backed up with 2-8MB of L3.)

That is the whole problem. If you reduce L2 cache...fine you improve the latency there...but L3 is a thing that doesn't exist on APUs.
They might as well entirely remove the L2 cache then...no cache = no latency, right? Oh...wait ;P

 

[Erenhardt]

if you compare Kaveri and Richland head to head...in a lot of gaming benchmarks it is either equal or LOWER than Richland (higher IPC does not manage to make up for the lower clock as well as it would have to). It also doesn't overclock as well generally...unless you get some super cherry pick.

Oh I did.
http://www.guru3d.com/articles_pages/amd_a10_7850k_apu_review,16.html

 

[NTMBK]

That is the whole problem. If you reduce L2 cache...fine you improve the latency there...but L3 is a thing that doesn't exist on APUs.
They might as well entirely remove the L2 cache then...no cache = no latency, right? Oh...wait ;P

Well yes, clearly there is a balance to be struck Just wanted to make the point that there's more to cache performance than just capacity. Alternatively you could have an 8GB cache... then whoops, the latency is roughly the same as going to main memory

 

[Shehriazad]

Oh I did.
*Snip*

This benchmark is a terrible pick and I'm not sure if you picked it on purpose to make it seem like Kaveri is better than it is.


Check the hardware they used.

6790K...not 6800K.

At first you might think "but hey, that's just 100MHZ"...wrong.

The 6790K in that benchmark was powered by 1866mhz ram because that is what it officially supports while the 7850K was powered by 2133 also:

The 6800K supports 2133mhz Ram AND is clocked 100mhz higher than the 6790K

The 7850K would have lost vs a 6800K in a CPU performance test. (Which also means 760K > 860K in raw CPU performance


I'm not sure if you're trying to spread misinformation on purpose or if you actually didn't research before posting this. There is a performance difference between 6790K and 6800K...enough to invalidate your entire post as it actually beats Kaveri in CPU related performance.

Kaveri still has reasons why its superior overall (Namely iGPU and PCI-E 3.0)...but still...don't try to bullshit a bullshitter

 

[geoxile]

Carrizo-L is 28nm? What the heck...

Well I hope the 20nm Skybridge APUs aren't cancelled, perhaps we'll see Puma+ on 20nm in March 2015