AMD Carrizo Pre-release thread

[ShintaiDK]

Anyhow:
Kabini a4-5000:
http://www.notebookcheck.net/Review-HP-Pavilion-15-n050sg-Notebook.106490.0.html
Beema A8-6410:
http://www.notebookcheck.net/HP-Pavilion-17z-Notebook-Review.124669.0.html

The 15 inch Kabini is much slower and consumes more power than the 17 inch Beema notebook.

While Beema is more efficient. Be careful to draw any conclusion from 2 randomly different laptops. Specially since the SoC itself is only a tiny part of the consumption. Also did you forget to check your links?
While its apples and oranges compare just due to the screen alone.
Your beema link uses 4.5-10.3W idle and 21.4-28.7W loaded. While your kabini link uses 5-7.3W idle and 17.2-21.6W loaded. Opposite than what you say. but again, completely different laptops.

AMD states the difference themselves.

 

[mrmt]

Didn't know that guess it's just the superior GF process that made Beema so much better over Kabini then. Either way SoC will shave of a few watts which is big for a mobile chip. The race is no longer about highest performance the race is all about who can offer mediocre performance at the lowest power consumption.

...at affordable costs. Carrizo becoming a SoC sure brings down the costs compared to Kaveri but not compared to Broadwell, because Broadwell is also a SoC. Expect another year of CPU market share bleed for AMD.

In the last Q&A with investors Su was pretty much confident regarding their new GPU architecture, predicting market share gains, but she refuses to say the same about Carrizo. She doesn't even talks about stable market share or even stable revenue amount, let alone market share gains.

 

[Arachnotronic]

...at affordable costs. Carrizo becoming a SoC sure brings down the costs compared to Kaveri but not compared to Broadwell, because Broadwell is also a SoC. Expect another year of CPU market share bleed for AMD.

In the last Q&A with investors Su was pretty much confident regarding their new GPU architecture, predicting market share gains, but she refuses to say the same about Carrizo. She doesn't even talks about stable market share or even stable revenue amount, let alone market share gains.

If you define SoC as "everything on the same package" then Broadwell is, but Broadwell has a 14nm CPU/GPU/media complex and then the I/O is on a separate 32nm die on the same package. So if you define SoC as "single die" then Broadwell-ULT is not one.

Intel's Atom-based Pentium/Celeron chips are single-die SoCs, though.

 

[Arachnotronic]

In the last Q&A with investors Su was pretty much confident regarding their new GPU architecture, predicting market share gains, but she refuses to say the same about Carrizo. She doesn't even talks about stable market share or even stable revenue amount, let alone market share gains.

As far as I understand it, these design wins at OEMs are decided well in advance (i.e. 6 months to a year), so AMD and Intel already know roughly speaking which designs their chips have won and not won.

Now, I think there is uncertainty in how particular designs will sell (although this can be estimated from sales of prior gen models in the same line), but this is why Intel is able to give full year guidance a year in advance, with the main uncertainty coming from the overall market rather than market share shifts.

 

[maarten12100]

While Beema is more efficient. Be careful to draw any conclusion from 2 randomly different laptops. Specially since the SoC itself is only a tiny part of the consumption. Also did you forget to check your links?
While its apples and oranges compare just due to the screen alone.
Your beema link uses 4.5-10.3W idle and 21.4-28.7W loaded. While your kabini link uses 5-7.3W idle and 17.2-21.6W loaded. Opposite than what you say. but again, completely different laptops.

AMD states the difference themselves.

That is a tad conservative but if it what AMD says it is then they are probably right. (those are all low usage tasks though)

...at affordable costs. Carrizo becoming a SoC sure brings down the costs compared to Kaveri but not compared to Broadwell, because Broadwell is also a SoC. Expect another year of CPU market share bleed for AMD.

In the last Q&A with investors Su was pretty much confident regarding their new GPU architecture, predicting market share gains, but she refuses to say the same about Carrizo. She doesn't even talks about stable market share or even stable revenue amount, let alone market share gains.

While Broadwell chips such as Core M and the early ULV chips have a smaller dies as in ~90mm^2 and up they will cost a boatload of money to produce. Despite Intel's claims anyone believing that Intel's 14nm is of comparable cost is lieing to itself. We all know that Intel's node must be quite expensive. (14nm is more like 20nm for the other foundries it seems density is pretty low for 14nm when Carrizo on 28nm approaches the same density)

The advantage for Intel is that they can keep spacing which reduces leakage while still fitting a lot of transistors. (See GK110 vs Hawaii)

If you define SoC as "everything on the same package" then Broadwell is, but Broadwell has a 14nm CPU/GPU/media complex and then the I/O is on a separate 32nm die on the same package. So if you define SoC as "single die" then Broadwell-ULT is not one.

Intel's Atom-based Pentium/Celeron chips are single-die SoCs, though.

indeed in the strictest sense of the acronym.

 

[AtenRa]

Second, the same 245mm^2 die puts AMD squarely in the same situation that has been making them bleed share for Intel in the APU segment, which is selling cheaper chips which are twice the die size compared to Intel's . Selling a SoC won't matter here, because Broadwell-U is also a SoC, so the net effect in terms of cost handicap is 0.

A 245mm2 die at 28nm in GloFo is cheaper than half or even smaller the die size at 14nm.

First of all GloFos 28nm is a half node from its 32nm and that makes it very very cheap from the start. Less R&D, same equipment etc etc.

Secondly Intels 14nm is not only more expensive in manufacturing the wafer(more stages than 28nm, double or even triple patterning etc etc) but has way lower yields now. Not to mention the gigantic R&D cost to develop and equip the Fabs.
And last, Broadwell is not a single Chip design, it uses a 32nm PCH chip and thus making it even more expensive.

So Carrizo even with a 245mm2 die it is cheaper than any 14nm SKUs Intel produces.

 

[mrmt]

We all know that Intel's node must be quite expensive. (14nm is more like 20nm for the other foundries it seems density is pretty low for 14nm when Carrizo on 28nm approaches the same density)

Sorry, but I don't share your set of assumptions that 14nm costs boatloads to make. Sure, it is being a slower ramp up than 22nm, but the costs of the new node by Intel financial statements are still following Moore's law, e.g., transistors are getting cheaper.

And I don't know where to start regarding the rest of your assumptions. GLF 28nm isn't not even close to Intel 14nm, maybe close to Intel 22nm but with much more inferior electrical properties.

Let's wait until Q214 to see what AMD will be forecasting to Carrizo.

 

[el etro]

It seems that 2015 will be a long year for AMD.

Maybe they are retreating from as many markets as possible to save R&D money for powerup Zen and K12, and they need to end the bulldozer cycle as soon as possible. Very much years of investment for nothing but a heavy direction change(from Big High Performance Very High clocked cores to High efficiency mobile-targeted cores) starting in Trinty/Richland.

AMD can do well and stay hugely relevant in the GPU universe, but they need to realize(and i know they realized it) that the AMD's core and the heart is the CPU processors, so they desperately need a competitive CPU arch to get really back into the business.

AMD needs to have more notebook with its APUs selling in the market, is very hard to found Kaveri products in the market but their mobile Kaveri is their most competitive product they have in many years. Selling directly reference designs(like Nvidia does with their Shields at the ARM land) could be a good thing.

 

[mrmt]

AMD needs to have more notebook with its APUs selling in the market, is very hard to found Kaveri products in the market but their mobile Kaveri is their most competitive product they have in many years. Selling directly reference designs(like Nvidia does with their Shields at the ARM land) could be a good thing.

Kaveri is extremely uncompetitive. I can't really fathom what kind assumptions the product managers had in mind when they were building the product scope for their APU series.

 

[maarten12100]

Sorry, but I don't share your set of assumptions that 14nm costs boatloads to make. Sure, it is being a slower ramp up than 22nm, but the costs of the new node by Intel financial statements are still following Moore's law, e.g., transistors are getting cheaper.

Moore's law is just an observation of IC progress over time it isn't bound to any true physics laws. Intel's statements about how the node does are quite broad , rumour has that they even reduced the complexity of early products to get them to tap out OKish to begin with.

Lower cost per transistor doesn't solve all the problems because you would want to put more transistors in every new product. If the transistor budget doubles the cost/transistor has to become half to keep an equal total cost.

Intel's marketing slides are really vague about this without concrete numbers and only logarithmic scales.
Let's try to work with Intel's slides anyway and assume best case scenario for Intel:

It's hard to read but when I interpret this log scale I come to:
Intel's 22nm node costs 0.15 unit of currency/transistor
Intel's 14nm node costs 0.09 unit of currency/transistor

This would make the 14nm node for double the transistor budget 0.15/0.09 = ~1.67x more expensive. This is with absolute best scenario numbers it means that Intel could according to their marketing save money if they were to use the same number of transistors. (those claims are probably not true TSMC has more money poured and they know very well how things work)

Going by Intel's other picture in that slide which is unit of currency/surface area:
For a die of the same size there will be a cost increase. Also note how the line even on that logaritmic scale is much steeper than previous nodes.

The steeper line -> higher slope coefficient on a log scale means that the cost per mm^2 has increased dramatically. Even on a normal scale a higher slope coefficient would indicate bigger over generation gains (in cost) but on a log scale this means it's a huge deal.

And I don't know where to start regarding the rest of your assumptions. GLF 28nm isn't not even close to Intel 14nm, maybe close to Intel 22nm but with much more inferior electrical properties.

You can't possibly be serious.

Here we go density wise:

ISSCC Tips Hot Circuit Designs
AMD Kaveri GF 28nm SHP 2.41B 245 mm2 9.837M
AMD Richland GF 32nm SOI 1.30B 246 mm2 5.285M
AMD Llano GF 32nm SOI 1.178B 228 mm2 5.166M

AMD Carrizo GF 28nm SHP? 3.1B 245mm^2 -> 12.6M/mm^2

Intel Core Sandy Bridge quad core Intel 32nm 1B 216mm^2 -> 4.6M/mm^2
Intel Core Ivy Bridge quad core Intel 22nm 1.4B 160mm^2 -> 8.8M/mm^2
Intel Core Haswell quad core GT2 Intel 22nm 1.4B 177mm^2 -> 7.9 M/mm^2
Intel's Core M Broadwell Intel "14"nm 1.3B 82mm^2 -> 15.8M/mm^2

Intel density is super duper low compared to competition up until 14nm hit but even there it is struggling as it comes out only slightly ahead of a node that is 2 full node shrinks behind.

Intel's nodes have superior electrical characteristics according to Intel versus the competition. Which is at least partly true yet still from time to time a node that is 2 shrinks behind matches Intel's efficiency. (ARM, MIPS companies and AMD have all produced similair and sometimes better efficiency designs despite the node disadvantage they faced)

Let's wait until Q214 to see what AMD will be forecasting to Carrizo.

Waiting for actual numbers is never a bad idea.

 

[Arachnotronic]

maarten12100

Don't confuse the intrinsic density capabilities of a given process with design choices that drive the density of a given chip.

The only meaningful way to compare the density of two processes is to look at identical circuits built in the two processes. SRAM cell sizes are a good way to do that.

 

[el etro]

Kaveri is extremely uncompetitive. I can't really fathom what kind assumptions the product managers had in mind when they were building the product scope for their APU series.

Competitive, i say, much better than former competitors. Does well compared to Trinty or Richland.

 

[mrmt]

You can't possibly be serious.

Here we go density wise:

AMD Carrizo GF 28nm SHP? 3.1B 245mm^2 -> 12.6M/mm^2

Intel Core Sandy Bridge quad core Intel 32nm 1B 216mm^2 -> 4.6M/mm^2
Intel Core Ivy Bridge quad core Intel 22nm 1.4B 160mm^2 -> 8.8M/mm^2
Intel Core Haswell quad core GT2 Intel 22nm 1.4B 177mm^2 -> 7.9 M/mm^2
Intel's Core M Broadwell Intel "14"nm 1.3B 82mm^2 -> 15.8M/mm^2

You can't compare two designs built on two different processes and expect to get the actual density advantage of one node over the other with it, because design affects *a lot* the transistor density you can get on a given design. In Carrizo's case AMD indeed built a very dense design, but one that reaches pathetically low clocks and is unsuitable for desktop/workstation duties. Broadwell on the other hand will scale from tablets to mission critical servers, the iGPU will have a narrower scope but far larger than Carrizo.

 

[mrmt]

Competitive, i say, much better than former competitors. Does well compared to Trinty or Richland.

It doesn't. Trinity had a much narrower power consumption gap compared to Sandy Bridge and a much larger GPU advantage compared to what Kaveri has against Haswell or Carrizo will have against Broadwell and Skylake. Kaveri got a very cold reception, both by OEMs that didn't bother too much with it, or by the technical press, that went great lengths to find a use beyond HTPC for it. AMD competitive position has been consistently eroding in the last 4 years, no wonder their market share has been consistently crashing.

 

[NTMBK]

Broadwell on the other hand will scale from tablets to mission critical servers, the iGPU will have a narrower scope but far larger than Carrizo.

Remember that Core M Broadwell is built on a different process variant to all other Broadwell- I expect it to be both more power efficient and denser.

 

[Arachnotronic]

Remember that Core M Broadwell is built on a different process variant to all other Broadwell- I expect it to be both more power efficient and denser.

Not sure how much I buy this given that Core M's die and Core i3/i5/i7 have the same die size...same transistor count...and same pretty much everything else.

 

[AtenRa]

Not sure how much I buy this given that Core M's die and Core i3/i5/i7 have the same die size...same transistor count...and same pretty much everything else.

No they dont, Core i3/5/7 have more cache. We cannot compere them directly to see the process differences.

 

[Enigmoid]

Here we go density wise:

AMD Carrizo GF 28nm SHP? 3.1B 245mm^2 -> 12.6M/mm^2

Intel Core Sandy Bridge quad core Intel 32nm 1B 216mm^2 -> 4.6M/mm^2
Intel Core Ivy Bridge quad core Intel 22nm 1.4B 160mm^2 -> 8.8M/mm^2
Intel Core Haswell quad core GT2 Intel 22nm 1.4B 177mm^2 -> 7.9 M/mm^2
Intel's Core M Broadwell Intel "14"nm 1.3B 82mm^2 -> 15.8M/mm^2

Intel density is super duper low compared to competition up until 14nm hit but even there it is struggling as it comes out only slightly ahead of a node that is 2 full node shrinks behind.

Intel's nodes have superior electrical characteristics according to Intel versus the competition. Which is at least partly true yet still from time to time a node that is 2 shrinks behind matches Intel's efficiency. (ARM, MIPS companies and AMD have all produced similair and sometimes better efficiency designs despite the node disadvantage they faced)

Waiting for actual numbers is never a bad idea.

Intel's CPU design is for high performance, NOT high density.

See all the dead space under the igp in the 3770k?

All that dead space on BW-U GT3

All that dead space on the GT3 die at the bottom where memory I/O is (thats dead space, not I/O). And on the top around the gpu and cpu cores.

 

[mikk]

No they dont, Core i3/5/7 have more cache. We cannot compere them directly to see the process differences.

You are wrong, they have both 4MB L3 Cache and both are 82 mm² big.

 

[maarten12100]

You can't compare two designs built on two different processes and expect to get the actual density advantage of one node over the other with it, because design affects *a lot* the transistor density you can get on a given design. In Carrizo's case AMD indeed built a very dense design, but one that reaches pathetically low clocks and is unsuitable for desktop/workstation duties. Broadwell on the other hand will scale from tablets to mission critical servers, the iGPU will have a narrower scope but far larger than Carrizo.

3,5GHz is pathetically low

It works for gpus almost linearly the comparison is pretty valid. Though I agree not entirely considering how bad AMD's core design currently is. A good wide core design on these nodes would prove very favorable towards AMD. Zen could become really good.

Intel's CPU design is for high performance, NOT high density.

See all the dead space under the igp in the 3770k?

A wide core but needless to say a wide core will stomp all over Intel's design. Some 120 million startup blew away Intel's Haswell IPC already. More transistors poured into a wider core rather than a longer core payed off. (though no x86 backward compatibility tardiness they had to deal with)

Yeah It's probably more memory IO. Not really sure though.

 

[mrmt]

3,5GHz is pathetically low

It works for gpus almost linearly the comparison is pretty valid. Though I agree not entirely considering how bad AMD's core design currently is. A good wide core design on these nodes would prove very favorable towards AMD. Zen could become really good.


A wide core but needless to say a wide core will stomp all over Intel's design. Some 120 million startup blew away Intel's Haswell IPC already. More transistors poured into a wider core rather than a longer core payed off. (though no x86 backward compatibility tardiness they had to deal with)

Yeah It's probably more memory IO. Not really sure though.

Carrizo won't achieve sustainable 3.5 Ghz.

 

[monstercameron]

Carrizo won't achieve sustainable 3.5 Ghz.

Source? Else don't post things you know will spread falsely.

 

[NTMBK]

Carrizo won't achieve sustainable 3.5 Ghz.

Please don't post guesses as if they were fact.

(Though I agree with your guess.)

 

[jpiniero]

3.5 Ghz seems reasonable since 35W Kaveri can turbo to that level. But it might be lower, yeah.

 

[maarten12100]

Carrizo won't achieve sustainable 3.5 Ghz.

Sustainable I don't know but there is this:

I see no reason for it not to go up to 3.5GHz. I'm quite positive Carrizo won't even take a performance hit just like with Kaveri I think there will be an upper range for efficiency of 3GHz but at the same time if you push enough voltage and keep the die cool 5GHz could be doable. (not talking extreme cooling just water cooling or decent air)