Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/amd-carrizo.5769/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

AMD Carrizo

Li Yisuo

Member
ISSCC 2015 AMD showed off their new Carrizo. AMD Discloses Architecture Details of High-Performance, Energy-Efficient
It is the 1st HSA1.0 fully supporting processor, which means https://en.wikipedia.org/wiki/Heterogeneous_System_Architecture era has arrived. Not everyone likes it I think. Will HSA aggressively joining mobile market arena? Or mounting up big data or squeezing into IOT or VR ...?

On chip H.265 decoder, besides integrated Southbridge, made PC/mobile device makers' life much easier. Adaptive voltage and frequency scaling refines the energy saving to every details with 10 modules on chip. It is marvelous.
Most striking part:
29% more transistors, 3.1B in total, 23% less area and 40% power reduction within the same 28nm node, with faster RVT device and high density library discussed.
What kind of magic is that? :rolleyes:

By the way, HBM was not shown with Carrizo. :confused:
 
ISSCC 2015 AMD showed off their new Carrizo. AMD Discloses Architecture Details of High-Performance, Energy-Efficient
It is the 1st HSA1.0 fully supporting processor, which means https://en.wikipedia.org/wiki/Heterogeneous_System_Architecture era has arrived. Not everyone likes it I think. Will HSA aggressively joining mobile market arena? Or mounting up big data or squeezing into IOT or VR ...?

On chip H.265 decoder, besides integrated Southbridge, made PC/mobile device makers' life much easier. Adaptive voltage and frequency scaling refines the energy saving to every details with 10 modules on chip. It is marvelous.
Most striking part:
29% more transistors, 3.1B in total, 23% less area and 40% power reduction within the same 28nm node, with faster RVT device and high density library discussed.
What kind of magic is that? :rolleyes:

By the way, HBM was not shown with Carrizo. :confused:

Carrizo is junk. It's lipstick on a pig, and AMD is misleading people on it, kind of.

It is not a lower power part, it's lower power at lower clock speeds. It's fine for the target market, but wouldn't make a particularly useful desktop chip, because it's clock speed limited.

Not that this is bad, or anything, it's a targeted product, but it's built on a brain-damaged architecture that has failed in every iteration, and will continue to with Carrizo.

Intel already copied the most important part of HSA, unified memory, and did it one better with allowing it even in virtualized situations. IBM/NVIDIA already beat both, with CAPI, and NVLink, allowing a fully blown GPU, not an iGPU, to have access in the same manner as an iGPU. Currently, GPGPU is limited either in separate memory stacks with large, discrete GPUs, or unified memory with small, integrated GPUs. Not so with OpenPOWER, and CAPI (and soon NVLink).

HBM is very expensive, so wouldn't necessarily work that well with the price points they are targeting. It's also dangerous to put it in an APU first, as the risk is much greater than using it first in a GPU, and seeing what kind of issues it creates.

Also, keeping in mind Tonga used data compression successfully to improved "bandwidth", in a sense, it seems very likely AMD will implement this in Carrizo, thereby mitigating the current bandwidth bottleneck their APUs with larger GPUs have. I think this is the right approach, at least for now. Maybe when HBM comes down in price, and as the GPU component continues to grow (by nature it will, since its parallel nature benefits more from transistors than the CPU), it will be tied to APUs. I'd even consider it inevitable.

I don't mean to pan AMD, just Carrizo. It's gotten way too much attention for a lousy part, targeting a limited segment. But, the technology they are using in it, particularly with regards to power saving, will be very useful moving forward, as it will no doubt be implemented in future products like Zen, and probably to better advantage.

I also want to be clear, I'm not talking about Carrizo-L, just Carrizo. Carrizo-L is likely to be a well balanced part, much like Jaguar/Puma were before it. Maybe less, since it will have a dual-memory controller, which is overkill, and adds expense. AMD may also decide to boost the GPU, which again would kind of make it GPU heavy like AMD has a bad tendency of doing. But, the processor is extremely efficient, and if it's got some compromises due to the platform, they will be minor in comparison, so should remain an excellent part.
 
Last edited:
Carrizo is junk
EPIC FAIL

I think this is simply the state of the art of what can be achieved at 28nm (not FDSOI).
1st HSA fully supporting processor, on chip H.265 decoder and integrated Southbridge.
29% more transistors, 3.1B in total, 23% less area and 40% power reduction within the same 28nm node: sure, this is junk!!!
View attachment 13738
 
EPIC FAIL

I think this is simply the state of the art of what can be achieved at 28nm (not FDSOI).
1st HSA fully supporting processor, on chip H.265 decoder and integrated Southbridge.
29% more transistors, 3.1B in total, 23% less area and 40% power reduction within the same 28nm node: sure, this is junk!!!
View attachment 13738

Are you serious? I'm guessing you're more of a fabrication guy than a design guy based on this.

Let me ask you, which CPU performs better, Carrizo, or Intel's now quite old Sandy Bridge? Sandy Bridge by a landslide? Which uses an older fabrication technology? Sandy Bridge? Which design was continued, and which is being killed? Sandy Bridge and Carrizo, respectively?

This is lipstick on a pig. The Bulldozer family is a failed technology, and this changes nothing. It's completely non-competitive with Broadwell. Which brings up the part you apparently are being confused by, the numbers you throw around (somewhat inaccurately), without giving the implications of them.

First, HSA is an invented name, but you'd be hard pressed to say what it does that Intel doesn't have on Broadwell. So, HSA 1.0? Who cares. I'll share something else with you, that I have to with a lot of people - GPUs generally suck at running code. No amount of coding will fix that, the scope of applications that can take advantage of it will always be limited. Think about it, if that were not so, CPUs would have been made that way, since their primary responsibility is to execute programs, GPUs is to render images. It just happens that a small percentage of workloads are more similar to that than mainstream software. So, keep the scope in perspective.

Next, Carrizo DOES NOT use less power. As a blanket statement. AMD sold out on performance, and power use at high performance, and used a higher density library. If you ran both chips at 3.7 GHz, Carrizo would use MORE power, not less. So, it's better to say that Carrizo uses less power in lower power scenarios. The distinction is important to make.

I made a lot of money on AMD, knowing when they had good technology (Jaguar, et al), then dumping when the horrible Kaveri came out. Many argued with me, because they lacked the mental strength to accept a company they liked had a lousy product. So, it failed, like Carrizo will, and AMD collapsed from inferior high-end products. Luckily, Jaguar/Puma continue to sell well, and keep them afloat, particularly in the consoles.

But, don't make the mistake of getting caught up with jargon and hype. Carrizo is based on a failed architecture, that is so wholly inferior, even the then CEO trashed it. Several times AMD has panned the architecture, in subtle ways, like stating Jaguar has the same IPC, but uses less than 1/3rd the space. Bulldozer can't be fixed, it's got to die. And it is. That's not because it's any good, it's because it sucks.

This isn't to say AMD made a mistake with Carrizo. They probably did the best they could, with what they had. But, in the end, you can't build a good product on a rotten core. But, they sold out on performance, and limited the market, and tried to do the best they could. It's not good enough to generate many sales, and in reality is completely incongruous with the architecture (which is designed to be a speed demon, not a brainiac), but it generates hope from AMDroids.

Let's hope Zen sets things right for AMD. The stuff they are learning with Carrizo, et al, should transition nicely to it, and with a decent core (we hope), the company could have some success in 2016. But, if they don't improve their horrible memory controller significantly, it's going to be difficult to be competitive. I don't know how they still have such terrible cache timings, and memory latency compared to Intel after so many years, but they do. Now that they have access to ARM IP, hopefully they'll get some help there.

I'm hopeful, but also doubtful for Zen. Carrizo is just hopeless. You'll either realize it now, or suffer a lot of pain as events unfold that show just how poorly it sells. Zen is where the hope is, but that's a year away.
 
Last edited:
Are you serious? I'm guessing you're more of a fabrication guy than a design guy based on this.
Do you know the design details of AMD APUs? How, please tell us?

Let me ask you, which CPU performs better, Carrizo, or Intel's now quite old Sandy Bridge? Sandy Bridge by a landslide? Which uses an older fabrication technology? Sandy Bridge? Which design was continued, and which is being killed? Sandy Bridge and Carrizo, respectively?
Sandy Bridge is probably the last good Intel innovative CPU. Do you agree? If not, we have nothing more to talk about. That said, SB was 32nm and CRZ is 28nm, both planar nodes: it sounds like a fair comparison.

This is lipstick on a pig. The Bulldozer family is a failed technology, and this changes nothing. It's completely non-competitive with Broadwell. Which brings up the part you apparently are being confused by, the numbers you throw around (somewhat inaccurately), without giving the implications of them.
Why the hell should we compare Carrizo with Broadwell. Do you have any idea about the cost per chip difference between them? Broadwell simply sucks: such a great process and such a poor performance jump. Wait, no jump at all. Just less power at best.

First, HSA is an invented name, but you'd be hard pressed to say what it does that Intel doesn't have on Broadwell. So, HSA 1.0? Who cares. I'll share something else with you, that I have to with a lot of people - GPUs generally suck at running code. No amount of coding will fix that, the scope of applications that can take advantage of it will always be limited. Think about it, if that were not so, CPUs would have been made that way, since their primary responsibility is to execute programs, GPUs is to render images. It just happens that a small percentage of workloads are more similar to that than mainstream software. So, keep the scope in perspective.
HSA is great, and all the others will just copy it sooner or later.

Next, Carrizo DOES NOT use less power. As a blanket statement. AMD sold out on performance, and power use at high performance, and used a higher density library. If you ran both chips at 3.7 GHz, Carrizo would use MORE power, not less. So, it's better to say that Carrizo uses less power in lower power scenarios. The distinction is important to make.
It is not correct. The 40% power reduction is at the same (even slightly higher) performance. Sad but true, that's instead exactly what Intel is doing: see the Core-M chip. Just a castrated version of Haswell (at 14nm).

I made a lot of money on AMD, knowing when they had good technology (Jaguar, et al), then dumping when the horrible Kaveri came out. Many argued with me, because they lacked the mental strength to accept a company they liked had a lousy product. So, it failed, like Carrizo will, and AMD collapsed from inferior high-end products. Luckily, Jaguar/Puma continue to sell well, and keep them afloat, particularly in the consoles.
I do not care about investments. So I just wish you all the best.

But, don't make the mistake of getting caught up with jargon and hype. Carrizo is based on a failed architecture, that is so wholly inferior, even the then CEO trashed it. Several times AMD has panned the architecture, in subtle ways, like stating Jaguar has the same IPC, but uses less than 1/3rd the space. Bulldozer can't be fixed, it's got to die. And it is. That's not because it's any good, it's because it sucks.
Once more, that's your opinion. You simply compare it with the I7. The market segment and the price are completely different.

This isn't to say AMD made a mistake with Carrizo. They probably did the best they could, with what they had. But, in the end, you can't build a good product on a rotten core. But, they sold out on performance, and limited the market, and tried to do the best they could. It's not good enough to generate many sales, and in reality is completely incongruous with the architecture (which is designed to be a speed demon, not a brainiac), but it generates hope from AMDroids.
Carrizo is just a mobile chip, and to me a very good value for the money you spend. Think about the crazy expensive Intel IrisPro GPU: Carrizo can easily draw circles on it.

Let's hope Zen sets things right for AMD. The stuff they are learning with Carrizo, et al, should transition nicely to it, and with a decent core (we hope), the company could have some success in 2016. But, if they don't improve their horrible memory controller significantly, it's going to be difficult to be competitive. I don't know how they still have such terrible cache timings, and memory latency compared to Intel after so many years, but they do. Now that they have access to ARM IP, hopefully they'll get some help there.
I bet so.

I'm hopeful, but also doubtful for Zen. Carrizo is just hopeless. You'll either realize it now, or suffer a lot of pain as events unfold that show just how poorly it sells. Zen is where the hope is, but that's a year away.
Once more, I disagree. Let's wait and see.
 
Do you know the design details of AMD APUs? How, please tell us?


Sandy Bridge is probably the last good Intel innovative CPU. Do you agree? If not, we have nothing more to talk about. That said, SB was 32nm and CRZ is 28nm, both planar nodes: it sounds like a fair comparison.


Why the hell should we compare Carrizo with Broadwell. Do you have any idea about the cost per chip difference between them? Broadwell simply sucks: such a great process and such a poor performance jump. Wait, no jump at all. Just less power at best.


HSA is great, and all the others will just copy it sooner or later.


It is not correct. The 40% power reduction is at the same (even slightly higher) performance. Sad but true, that's instead exactly what Intel is doing: see the Core-M chip. Just a castrated version of Haswell (at 14nm).


I do not care about investments. So I just wish you all the best.


Once more, that's your opinion. You simply compare it with the I7. The market segment and the price are completely different.


Carrizo is just a mobile chip, and to me a very good value for the money you spend. Think about the crazy expensive Intel IrisPro GPU: Carrizo can easily draw circles on it.


I bet so.


Once more, I disagree. Let's wait and see.

The design stuff is easy to find no the net. Do a search, but I like Realworldtech the best. I mean really, that's kind of a silly question, anyone can find this stuff. Here's a good link on BD -> AMD's Bulldozer Microarchitecture

Your view of "good" is simplistic. Broadwell is vastly superior to Sandy Bridge in the same way you point out Carrizo is improved, except Broadwell is better at everything (albeit, not to the same extent we expected). Ivy Bridge was a nice a shrink, Haswell looked great on paper, and fooled me when it came out. Based on the design, I was expecting much better multi-threaded performance, but in reality, it didn't help that much. Broadwell is too narrowly focused, but it does move Intel lower into tablets. But, that market is iffy. Carrizo is pure junk, by comparison, and can't replace the predecessor because it actually performs worse at high clock speeds.

You obviously don't understand what Carrizo is. I'm not talking about IPC, I'm talking about clock speeds. Carrizo CAN NOT clock as high, and is a lower performance part because of it. If you use them both at 25 watts, Carrizo will perform better, if you look at 125 watts, or 95 watts, Steamroller would. Of, if you want to look at 3.7 GHz base, for example, you'd need much more power for Carrizo than Kaveri. If you don't know this, there's nothing left to talk about until you learn more about the processor. Even AMD has published this. Actually read it. That's why it's not going to be on the desktop, it performs worse, and is targeted at narrow watt ranges.

You have the market segment stuff completely backwards. AMD can't market it there, so you have the tail wagging the dog. AMD's BD line is quite large, not small, so from that perspective it IS the same segment. But, you have it all backwards. They don't sell it for as much, because it's wholly inferior and no one would buy it. Thus, the inferiority I'm telling you about is exactly the cause it's in a different segment, and exactly why Intel can charge more. It's not because AMD's chips are small, they're huge. But, they have to segment them where people will actually buy them. Of course, I'm not referring to Jaguar/Puma, just the big junk.

HSA is just a name. It didn't help Kaveri, and saying it is fully HSA 1.0 now isn't going to change anything. But, sadly, once again Intel has already copied the unified memory architecture, so it's not even an advantage for AMD. So, yes, it's good to have, but not a big deal by any means. The problem is, once again Intel failed to innovate, but didn't fail to copy before AMD gained any advantage. It's easy to hate Intel, not only for their lack of innovation, but also because nonsense like contra revenue destroys the merits of technology. It's disgusting.

Broadwell isn't a bad chip, but it's easy to see why people are disappointed by it. I am as well. Intel is being very myopic, and worrying too much about power. AMD had no choice, there's no way they're getting performance out of the BD line, it's horrible. Intel has no idea how to make a low-end chip, Atom is a disaster and a terrible design. But, it keeps AMD out with contra revenue. But, in terms of performance per watt, at low wattages, it's a good improvement. Also, unlike AMD, they didn't have to sell two chips when they came out with Haswell (same process as IB, so I'm using it as an example). Because AMD hurt the potential performance with Carrizo, they have to keep Kaveri, which didn't improve performance over the Piledriver processors. So, AMD has three lines of big processors, from different generations, because they can't improve performance uniformly. In each case, the max clock speed dropped. Haswell didn't have that problem, and yet still had better performance at lower wattages than IB. So, you pan Intel for something, but then go easily on AMD for doing something far worse. Three generations of the processors, and yet the oldest is still the performance part, due to higher clock speeds. That's OK, because, well, AMD said it's for mobile, but it's not OK for Intel, even though they improve performance in mobile dramatically, but only improve performance on the desktop slightly. That's worse, right? AMDroids always use weird logic.


Until Carrizo is out, it's silly to say what it can and can't do compared to Intel's Iris Pro. Obviously, AMD's GPU technology is more efficient, but if you haven't noticed, memory is becoming the bottleneck on APUs, and Intel's memory controller is way better. So, we'll have to see, but iGPU performance obviously matters little, since AMD has had a edge for a long time, and it hasn't correlated into sales. Look at the last several years. AMD had better iGPUs, and suffered badly in market share losses. But, again, you don't compare the cost from a retail perspective, but from a creation perspective to judge the success of a design. AMD can't sell their chips for much, because they suck. That's not a positive, so why are you using it as one? If you want to talk about how tiny a core is, and use that as a basis, I'll go along with that. Telling me a large chip sells inexpensively is exactly the point you don't want to make.

You're right, let's wait and see. Most of the time, it's easy to see, and it is this time too, Carrizo should continue to languish, but you never know until you know. Particularly on a product we haven't seen benchmarked, but given AMD's guidance about performance, I'd be astounded if it performed radically different in IPC, and if it doesn't, the performance improvement will be too small to even compete with Intel's superior design. And yes, I do compare them, because AMD's chip is large, and isn't cheap to make. Much larger than Core M, about double the size. It's also significantly larger than Haswell, if you want to move up a node for Intel, so let's not try to pretend they aren't competitors, because you don't like the comparison. It has to compete with those chips, or sell at very low margins. Or more like not sell at very low margins, since they can't move chips even with low pricing. And that should tell you the chip isn't any good. But, let's wait and see, you're quite right. Anything can happen, as always.
 
Last edited:
@ta152h,
Thanks for your reply. May I know CAPI+NVlink uses ESSL + CUDA, how does it work?
In general, HSA is supported by Linux kernel, we can find some code on git. OpenCL-MPI vs I-MPI+CUDA, I support mkl and I-MPI and also cuda, but they won't work together.
Regarding share memory, HSA is more than that in term of paralleling.
To render a 3d image, GPU is fast but not perfect for all the situations, APU is a reasonable choice.
MKL is marvelous with Pardiso, for large scale CFD computation, Intel has huge advantage. But HSA certainly has its position.

Do you know why they integrate Southbridge? Thanks.

Process:
Such density power gain within same node is very difficult. Never happened.

Eda:
Magic high density lib also quite interesting, can we use it for other nodes?

Design:
They play voltage and frequency together at every detail part. Worth studying.
 
Last edited:
Back
Top