Counting the success of Kabini & Temash

[bullzz]

@siliconwars
"What AMD should be doing is licensing their GPU like Nvidia is doing, however"

and who wud be using it? nvidia optimised their GPU for phones n tablets. so they can expect others to use it. AMD themselves are yet to fully jump into tablets/phones. soc makers will have better ROI in licensing nvidia or IMG since they have a proven record in these devices

 

[monstercameron]

@siliconwars
"What AMD should be doing is licensing their GPU like Nvidia is doing, however"

and who wud be using it? nvidia optimised their GPU for phones n tablets. so they can expect others to use it. AMD themselves are yet to fully jump into tablets/phones. soc makers will have better ROI in licensing nvidia or IMG since they have a proven record in these devices

maybe they dont have any recent wins in the mobile/tablet space[temash?] but these are the guys who brought and sold imageon, which is the basis of adreno...though this has been a while.

 

[pablo87]

I would tend to agree with you about 17/19W Richlands but A10-5745M is not that case.
On CPU side is faster than A6-5200 and it's GPU is a different class.
128 SP vs 384 SP (600 Mhz vs 533/626 Mhz) plus dual channel IMC means huge advantage for Richland.

http://www.cpubenchmark.net/cpu.php?cpu=AMD+A6-5200+APU&id=1975
http://www.cpubenchmark.net/cpu.php?cpu=AMD+A10-5745M+APU&id=1962
http://www.notebookcheck.net/AMD-Radeon-HD-8400.93720.0.html
http://www.notebookcheck.net/AMD-Radeon-HD-8610G.93719.0.html

At some point, you have to stop arguing and start asking questions. If Kabini had been optimized solely for TSMC, given DC IMC and 384 SP, it would prob. be half the die size of Rich, not to mention the integrated FCH. Packaging incl., its almost a 3 for 1 deal.

 

[AtenRa]

You say that but then one just has to look at the Temash Acer Aspire V5 with the 17W Celeron one that costs $50 less (last I saw)..

You have to understand that what you pay in that category is not performance but battery life, weight and features. With the Celeron you get a 6-cell battery (6.5h) with a weight of 3.26lb no USB3 and Win 7.
The Temash with half the battery cell size you have 5h battery life with a weight of 3.04lb one USB3 port and Win 8. Also the Temash has AES something the Intel Celeron lucks, not to mention way better iGPU performance and features than the Celeron. That’s why it cost more.

Then you have mobile Richland which has 17W, 19W, and 25W variants. I don't know how much they cost compared to Kabini but if they performance is worse why would AMD even be bothering?

Again, the 15W Kabini is a SoC when the Trinity/Richland also needs one more chip that adds to the size and weight of the device and the power consumption. The 25W QuadCore Kabini must be faster than the dual core 19W Trinity in CPU performance at almost the same power consumption (19W + Northbridge Chip). Those two are targeting different customers, the Kabini is going for low power when Trinity is about Performance (especially in iGPU).

 

[AtenRa]

Is the die size of Silvermont even known at this stage? I would have thought that it's smaller than Kabini.

Sorry i was talking against 32nm ATOMs.
32nm ATOM die size is close to 80-90mm, now the 22nm ATOM will have double the Core count and bigger iGPU + it will intergrade the northbridge. I will say it will be close to ~100mm2.

 

[SiliconWars]

At some point, you have to stop arguing and start asking questions. If Kabini had been optimized solely for TSMC, given DC IMC and 384 SP, it would prob. be half the die size of Rich, not to mention the integrated FCH. Packaging incl., its almost a 3 for 1 deal.

Richland is 246mm2, Kabini is ~110mm2 so it's not far from half the size as it is. The 65-nm Hudson FCH for Brazos was ~28mm2 so it'll barely even register on 28nm, let's say it'll be 100mm2 without the FCH.

An extra 256 SP's (I'd wager 128 SP's take up a big part of the Kabini die), 128-bit bus, double the cache and everything Piledriver has over Jaguar on top of that is going to take it closer to 200mm2. Then of course 28nm is smaller than 32nm anyway.

It would be be more like 1 for 1. Kabini is not particularly small for what it is, Richland is not particularly big.

 

[rainy]

True but at the cost of a blown tdp budget. Notebook check has the 17w a6-4455m going over 40w in a Samsung notebook. Pile driver apus will exceed their tdp while jaguar will be relatively close.

I want to make a small correction: those single module Richlands at 17/35W TDP don't make sense at all to me.
As 19W model (A8-5545M) is a dual module part, it is OK in my opinion.
Btw, it's CPU Turbo mode is rather amazing - close to 60 percent (1,7/2,7 GHz).

As we are talking about TDP: Richland have that called TSTC (Temperature Smart Turbo Core).

Richland's new temperature and clock management technology. TSTC primarily involves the use of 17 temperature sensors (five on each CPU and seven on the GPU) combined with a package sensor that allows the internal software to dynamically change the clock speed of both the CPUs and GPU based on load to enable the best possible performance whilst staying within the chip's thermal limits. For example, if the user is playing a GPU intensive game, the technology will reduce the CPU's clock speed which will then reduce its temperature and enable the GPU to run at a higher clock rate.

Though AMD's previous generation of Trinity chips also featured a similar system, it did not account for bottlenecks since the algorithm simply granted the CPU or GPU whatever power it demanded. Richland's algorithm, on the other hand, takes this into account and only supplies additional power if it will result in performance improvements or a more efficient operation.

https://en.wikipedia.org/wiki/List_of_AMD_APU_microprocessors#.22Richland.22_.282013.2C_32_nm.29_2
http://www.tomshardware.com/news/Richland-APU-AMD,21318.html

 

[mrmt]

It would be be more like 1 for 1. Kabini is not particularly small for what it is, Richland is not particularly big.

Richland CPU part alone is almost the size of SNB 2C complete die and both are 32nm. I very much doubt that Jaguar would be that bad.

Even if AMD could get 10-15% better costs, that would do wonders for their gross margins.

 

[rainy]

At some point, you have to stop arguing and start asking questions. If Kabini had been optimized solely for TSMC, given DC IMC and 384 SP, it would prob. be half the die size of Rich, not to mention the integrated FCH. Packaging incl., its almost a 3 for 1 deal.

Sorry but it doesn't work that way: if you would add Kabini 4 CU more, replace 64-bit MC with 128-bit version and increase L2 to 4 MB it would be much bigger as it is - probably close to 200mm2 as it was mentioned already by SiliconWars.

If you don't remember Brazos is about 75mm2 - at 28nm it should be circa 50, which means Kabini is over twice bigger.

One more thing: Jaguar cores can't reach 2,5 GHz clock (at least at 28nm) like Piledriver and they don't have Turbo mode (maybe in Kabini 2.0).

Btw, Richland at 28nm could be about 200mm2 too.

 

[SiliconWars]

Richland CPU part alone is almost the size of SNB 2C complete die and both are 32nm. I very much doubt that Jaguar would be that bad.

It's only about 10% bigger than the SB 4C die, and that's with far more area given over to graphics (half the die instead of what, 1/5th or 1/4 on SB?).

 

[zlatan]

What AMD should be doing is licensing their GPU like Nvidia is doing, however.

With the current GCN design (x86-64 TLB support) they can't do this.

http://download.intel.com/pressroom/legal/AMD_settlement_agreement.pdf

2.5.1. Neither Intel nor AMD shall either request of, suggest to or cause any public procement agent or authority to issue or adopt a procement specification for Computer Product that excludes the qualification of a Computer Product empoying the other Party's microprocessor on the basis of a function, industry standard specification or benchmark that is immaterial to the intended use of of the product.

They have to create a GCN for ARMv8 ISA.

 

[strata8]

If you don't remember Brazos is about 75mm2 - at 28nm it should be circa 50, which means Kabini is over twice bigger.

Brazos was 75mm2 + 26mm2 for the FCH, or 101mm2 in total.

 

[mrmt]

It's only about 10% bigger than the SB 4C die, and that's with far more area given over to graphics (half the die instead of what, 1/5th or 1/4 on SB?).

I wasn't talking about the GPU parts, but CPU parts only, because this is the crux of AMD problem on the big core market: Richland >>CPU<< alone is almost as big as an entire GPU/GPU Intel 2C die, meaning that even at 32nm AMD would have some kind of area handicap because of bulldozer. I think we can agree that bulldozer isn't very area efficient, can we?

And I doubt that a low clock heavy synthesis design like jaguar would be THAT bad in a die area comparison.

 

[rainy]

Brazos was 75mm2 + 26mm2 for the FCH, or 101mm2 in total.

OK, I forgot about FCH but it wasn't made at 65nm?

 

[krumme]

Richland CPU part alone is almost the size of SNB 2C complete die and both are 32nm. I very much doubt that Jaguar would be that bad.

Even if AMD could get 10-15% better costs, that would do wonders for their gross margins.

What matters as you said earlier is ROI on investment.
As a TSMC 28nm wafer is what 2000usd now, a kabini is under 10usd per pcx. for marginal production cost. It doesnt matter if it was 80mm2 or 130mm2 - the cost is not here. What matters is scalability to other products eg. consoles and lowering design cost - and easyness for the OEM. I guess its build very object like to be easy to port to different designs or processes. That cost something - and thats space. The console deal already shows this was a good investment, but it was an initial risk.

The problem for Richland is not its size its the performance and the huge design cost. For a notebook user, who is the customers who is not sattisfied with a 25w dirt cheap kabini implementation and go for the cheap Richland?
Its a damn small segment, and AMD can only get marginally more for Richland than Kabini. I dont think Kaveri will change that. Its just to plain expensive to devellop and produce to give any profit with competition from Kabini and the Intel lineup now and in the future. There is absolutely no future for AMD x86 big cores. We need more than a revolution here for that to change, and they never happen.

-----------------

And no a A57 at eg. aprox. 1.5mm2 on 20nm is not huge, or A15 at 1.5mm2 at 28nm - comparing Atom to that is nonsense, as both marginal and fixed cost is an order of magnitude lower. Its hard to grasp i think and we will see a lot of "confusion" here, but it doesnt change the brutal way the market works. As when Intel wiped IBM out of the server cpu business.

 

[strata8]

OK, I forgot about FCH but it wasn't made at 65nm?

That's true, yeah.

I did some measurements and Kabini with 6 CUs and 4MB L2 would come in at ~150mm2 - probably less in reality since I just doubled the GPU and L2 area. I'm not sure how much a 128-bit bus will increase the area.

 

[krumme]

Yes that'll be why AMD's CPU revenue was up 12% mostly on Jaguar sales last quarter, while Intel's was up only 1.4% (and their tablet segment was down 3.7%). I think it's quite obvious who is being squeezed by ARM.

Who can find an alternative explanation for the CPU revenue was up 12% than Jaguar sales?

Did someone find a hidden gem in the Richland or AMD server lineup, and didnt tell it to the rest of the world, so the nonstop decline for AMD could stop?

There is not much profit in this business for AMD, but i am sure, in the next ½ a year we will see tons of Jaguar on the notebook market, as bobcat also had a slow start. Add AMD have to provide for the consoles too at H2. And before Atom hits market every jaguar will be quad cores. I saw some dirt cheap Packard Bell trash with dual core jaguar 15W for sale - and that sort of old time segmentation will stop before Atom hits. But it looks like AMD can do it in the start.

 

[rainy]

That's true, yeah.

I did some measurements and Kabini with 6 CUs and 4MB L2 would come in at ~150mm2 - probably less in reality since I just doubled the GPU and L2 area. I'm not sure how much a 128-bit bus will increase the area.

Which means that Brazos as SOC would be rather 60mm2 at 28nm.

I want to make it clear: I like Kabini as chip but I don't think it could replace all Richland models.
Personally, I would keep A8-5545M/A10-5745M/A10-5750M - rest of them out.

 

[mrmt]

What matters as you said earlier is ROI on investment.
As a TSMC 28nm wafer is what 2000usd now, a kabini is under 10usd per pcx. for marginal production cost. It doesnt matter if it was 80mm2 or 130mm2 - the cost is not here. What matters is scalability to other products eg. consoles and lowering design cost - and easyness for the OEM. I guess its build very object like to be easy to port to different designs or processes. That cost something - and thats space. The console deal already shows this was a good investment, but it was an initial risk.

What I was saying fits perfectly what you are saying here. Even if a beefed up Kabini ends up costing only 10%/15% less than Trinity/Richland, it would still be a far better investment than the latter. Not only it costs less to develop, 15% in COGS would do wonders for AMD margins.

The problem for Richland is not its size its the performance and the huge design cost. For a notebook user, who is the customers who is not sattisfied with a 25w dirt cheap kabini implementation and go for the cheap Richland?

FWIW I don't think AMD big core line is more than a WSA quota filler. They are bound to a take-or-pay 1.1 billion contract, meaning that even selling chips at 0% gross margins is better than getting negative margins by paying Globalfoundries for not hitting the quota.

No matter how much AMD invests in the Bulldozer lineage, they won't turn the thing into a winner. They won't be able to beat Intel, they won't be able to scale the thing down to mobile market, they won't be a star on dense servers, meaning that it has no place in AMD's future.

No executive calls his own product an "unmitigated failure" for no reason, and it wasn't criticism directed to x86 chips, as he didn't extend the courtesy do Brazos/Jaguar.

 

[Fox5]

At some point, you have to stop arguing and start asking questions. If Kabini had been optimized solely for TSMC, given DC IMC and 384 SP, it would prob. be half the die size of Rich, not to mention the integrated FCH. Packaging incl., its almost a 3 for 1 deal.

Performance per watt, Richland is still ahead of Kabini. Kabini is tiny, and thus dirt cheap. Kabini's GPU does have the best performance per mm^2 of all the current SOCs though, Intel Atom and all the ARM designs included, and probably the best performance per watt.
The cpu portion doesn't seem particularly competitive with current ARM designs in performance per watt or per mm^2.

 

[SiliconWars]

Well it wasn't Feldman's own product so he can basically get away with saying that. Few at AMD are going to argue anyway.

I don't know if SR will truly fix the BD problems or not, but on paper it would appear to be getting rid of the main obvious bottleneck at the front end.

The problem with Trinity/Richland is that the BD core it's a total paradox for the market they are designed for. It simply couldn't have a worse core for what is supposed to be mainstream "gaming" chip.

BD is pure crap, we know this, but SR should go a long way to helping out with single threaded performance. With that in mind and the supposed similar die size, we'll see what the real problem with Trinity/Richland was.

 

[strata8]

Which means that Brazos as SOC would be rather 60mm2 at 28nm.

I want to make it clear: I like Kabini as chip but I don't think it could replace all Richland models.
Personally, I would keep A8-5545M/A10-5745M/A10-5750M - rest of them out.

Richland is fine in all of the standard voltage parts, where it outstrips any Jaguar based part, but it's the low voltage parts that are the concern. It really depends how much it's been improved over Trinity, because the 19W 4-core A8-4555M was absolute trash and would be walked over in every aspect by the A6-5200.

 

[mrmt]

I don't know if SR will truly fix the BD problems or not, but on paper it would appear to be getting rid of the main obvious bottleneck at the front end.

Impressive how after 7 years engineering a product and having the 45nm version canned AMD engineers still wouldn't get rid of this obvious mistake, isn't it?

 

[rainy]

Richland is fine in all of the standard voltage parts, where it outstrips any Jaguar based part, but it's the low voltage parts that are the concern. It really depends how much it's been improved over Trinity, because the 19W 4-core A8-4555M was absolute trash and would be walked over in every aspect by the A6-5200.

I don't know how reliable is PassMark but it show that A8-4555M (1,6/2,4 Ghz) have 744 points in single thread and 2310 for multi and A6-5200 have 794/2616.
Of course on GPU side Trinity is much faster.

A8-5545M CPU clock is a bit higher (1,7/2,7 GHz) but GPU is different story: 450/554 MHz vs 320/424 Mhz on it's predecessor.
Even bigger difference is with A10-4655M (2,0/2,8 GHz) vs A10-5745M (2,1/2,9 GHz): 360/496 MHz vs 533/626 MHz.

http://www.cpubenchmark.net/singleThread.html

 

[SiliconWars]

Impressive how after 7 years engineering a product and having the 45nm version canned AMD engineers still wouldn't get rid of this obvious mistake, isn't it?

The mistake was assuming the industry would be quick to embrace 8 cores, when the reality has been a long slow slog. BD would be fine if most software was properly multi-threaded, the benchmarks show this. In many ways it was an architecture ahead of it's time.

But there's no need to worry because AMD won't be doing anything daring like this in future, that's why SR is a return to a more normal architecture. Of course their high end CPU's will totally stagnate in the same way Intel's have, but that's the price we pay for Intel's determination to put margins ahead of progress.