[Digitimes] Kaveri delayed

[frozentundra123456]

I said they needed to kill Kaveri as an APU and focus on GPU less Kaveri/Big Core CPU instead. For sub $200, they could give customers a GPU less Kaveri, mb and discrete graphics (7850) for the cost of an Intel quad core.

What is AMD's higher ASP mobile share right now, 10%? I have not been able to find a single AMD sku with discrete graphics (I'm sure there is but I tried and gave up).

Dont follow your math here. Cheapest Intel quad is 200.00 or even less.

HD7850 = 150.00
motherboard = 50.00
cpu = free??

dont think so. Personally, I am not a fan of APUs on the desktop, but that is the only place where AMD has an advantage over intel.

A more likely scenario for AMD to claim the value proposition is to claim adequate gaming on the APU for 150.00. Personally I dont buy that argument, because I believe you can still get a better value with a discrete card and a cheaper cpu, but that is an argument made by some AMD supporters.

 

[pelov]

But lets lay the blame on GF for Llano anyway...what about Trinity? Was that GF's fault too? The 5800K compares pretty evenly vs 980 BE with two fairly major exceptions. First is it's a 125W Phenom vs 100W Trinity, and second is Trinity has the large graphic core. It should be obvious by this stage that GF's 32nm was in pretty decent shape, but that didn't save AMD when the market fell apart at the end of last year.

Trinity's not a bad chip at all, and I think that needs addressing despite the obvious digression So let's spin that around some and get it back on track.

The Llano delay was most certainly GloFo's fault. There are no ifs, ands, or buts about it - GloFo screwed up big time on 32nm and that process wasn't ready for silicon for an additional 6-12 months after GloFo claimed it was ready.

Trinity shouldn't be compared to Phenom II but rather Llano. The same 32nm process and AMD achieved some significantly higher clocks (and a longer pipeline) but improved performance drastically. The CPU performance was a bit higher - single-threaded went up by a decent amount whereas multi-threaded didn't due to the shared FPU and -20% tax. The GPU performance otoh went up quite a bit. The product was also released only a year after Llano. Technically, Llano was meant to be released early 2011, but due to GloFo providing a very poor 32nm node, AMD had to delay Llano by 6 months, and even then it was supply constrained until Q3. By then GloFo had most of the issues with their 32nm node ironed out, luckily (or unluckily?) just in time for Bulldozer. (which itself was delayed from the 45nm node. Whether that was a result of GloFo's 32nm node being in an awful state or the delay on the 45nm node, or AMD's production samples showing poor performance is up for conjecture. I'm of the belief that it was likely a mixture of all three)

The point is, yes, GloFo most certainly deserves some of the blame here. It's b/c of GloFo's delay on 28nm that AMD had to go to TSMC for Kabini/Temash. In fact, it's likely that the 28nm delay is a direct result of the difficulties GloFo had with 32nm. And though I do think part of AMD's woes lie directly at he feet of GloFo, but it was ultimately AMD's idiotic decisions that put them in this situation in the first place. Had AMD not signed a longterm WSA with GloFo, AMD could have ported some of their future chips to TSMC or Samsung after the fiasco that was 32nm at GloFo. Could either have provided enough capacity? Well, in a struggling PC market and AMD's currently underwhelming lineup, I reckon that wouldn't have been a problem.

 

[SiliconWars]

http://seekingalpha.com/article/165...m-and-activities-conference-transcript?page=2 (August 22nd)

later this year we will be shipping a higher end desktop product called Kaveri.

and we will be launching Kaveri in terms of shipment and shipping product to customers in the fourth quarter of this year.

 

[raghu78]

But lets lay the blame on GF for Llano anyway...what about Trinity? Was that GF's fault too? The 5800K compares pretty evenly vs 980 BE with two fairly major exceptions. First is it's a 125W Phenom vs 100W Trinity, and second is Trinity has the large graphic core. It should be obvious by this stage that GF's 32nm was in pretty decent shape, but that didn't save AMD when the market fell apart at the end of last year.

everybody knows kaveri is delayed due to GF. by atleast 6 - 8 months. you don't need better proof than these roadmaps. Richland was not present at all in Feb 2012 roadmap.

Feb 2, 2012 Analyst roadmap

http://images.anandtech.com/doci/5503/Screen%20Shot%202012-02-01%20at%202.14.03%20PM.png

CES Jan 2013 roadmap

http://images.anandtech.com/doci/6567/AMD-024.jpg

moreover Llano on 32nm SOI launched in June 2011. it had yield problems till Q4 2011 and AMD suffered a lot.

http://www.anandtech.com/show/5831/amd-trinity-review-a10-4600m-a-new-hope

By the time Trinity launched in June 2012 GF had resolved 32nm yield issues. If things had gone to plan Kaveri should have released at Computex 2013 timeframe (mid of this year). We are hearing AMD say shipping to customers in Q4 2013. They still don't mention a product launch in 2013. so it means AMD will just start volume shipments of Kaveri in Dec 2013 or year end to channel and OEM customers. Kaveri might be announced at CES 2014 or later. Kaveri should be available for purchase at retail by early Feb 2014.

http://vr-zone.com/articles/new-confirmed-details-on-amds-2014-apu-lineup-kaveri-delayed/47455.html

In my opinon and most people around Kaveri is a massively delayed product due to GF's screwup on 28nm high performance bulk process. remember the GF SLP process has just started volume production in Q2 2013 with customers like Rockchip shipping ARM SOCs running at 1.8 Ghz.

http://www.globalfoundries.com/newsroom/2013/20130617.aspx

But for Kaveri AMD needs GF 28nm high performance process at good yields for circuits running at 4.0+ Ghz frequencies. thats another ballgame completely. AMD's problems with GF stem from the WSA as they have to buy USD 1.1 - 1.2 billion of wafers every year at GF. Otherwise AMD could have gone to TSMC for their entire APU production. TSMC has been rock solid with Kabini/Temash and now Xbox One SOCs production confirmed. PS4 SOC should also be at TSMC 28nm. For Kaveri AMD is stuck with GF.

 

[inf64]

Latest post from Stilt(XS) about Kaveri . Quad channel is impossible someone said?

KaveriPI (AGESA):

- Support for KV-B0 stepping.
- "Spectre" & "Spooky" HUMA GOP VBIOS
- DCT0, DCT1, DCT2, DCT3 config array

Loving the GPU codenames

4x64bit 2133Mhz DDR3 would be more than enough for Hawaii derived GPU in Kaveri . With IPC improvements in x86 cores and major GPU perf. jump this thing will become practically an FX6300 with a GPU on board . Now imagine if AMD would launch 6C/6T part with same GPU, who would need FX83xx?

 

[Homeles]

Not impossible -- there's no question of its technical feasibility. You simply slap on two extra 64 bit memory controllers and call it a day.

The issue is that if you want to take advantage of the extra channels, you need to have four ram slots. That is a pretty substantial technical challenge for mobile products, which are striving to have less components on board, not more.

From the same page you linked:

I just see few laptops supporting quad channel. Remember APU laptops, even the top ones, are budget laptops. There's limited space to add 2 more memory slots and the necessary signal routing to them on smaller laptops. Even on laptops that do have the room it still adds to costs.

The problem is that the majority of the products will be taking on the burden of having two extra memory controllers on die for zero benefit. It'll be nice for products that do support four channels, but it is my opinion that much better solutions to the bandwidth problem exist.

 

[inf64]

Stilt already replied to that argument.
I think the most benefit will be had on desktop parts of course as AMD intends to use FM2+ for all of their upcoming desktop products. One day maybe we could slide in 8C EX with 768 GCN SPs or sooner than that a 6C Kaveri. Desktop market is not expanding or growing but still is very important revenue stream.
Laptop APU models won't have such a big need for memory BW(cut down SPs, cut down clocks) and if there is need there might be several premium models that can support 4ch IMC configuration (more expensive of course).

 

[SiliconWars]

everybody knows kaveri is delayed due to GF. by atleast 6 - 8 months. you don't need better proof than these roadmaps. Richland was not present at all in Feb 2012 roadmap.

Feb 2, 2012 Analyst roadmap

http://images.anandtech.com/doci/5503/Screen Shot 2012-02-01 at 2.14.03 PM.png

CES Jan 2013 roadmap

http://images.anandtech.com/doci/6567/AMD-024.jpg

moreover Llano on 32nm SOI launched in June 2011. it had yield problems till Q4 2011 and AMD suffered a lot.

http://www.anandtech.com/show/5831/amd-trinity-review-a10-4600m-a-new-hope

By the time Trinity launched in June 2012 GF had resolved 32nm yield issues.

I'm aware of all that, my point was that GF wasn't to blame for Trinity and all the penalties AMD had to take because of it.

In my opinon and most people around Kaveri is a massively delayed product due to GF's screwup on 28nm high performance bulk process. remember the GF SLP process has just started volume production in Q2 2013 with customers like Rockchip shipping ARM SOCs running at 1.8 Ghz.

http://www.globalfoundries.com/newsroom/2013/20130617.aspx

GF has been in volume production of 28nm since the end of last year. They have had those rockchip SoC's in tablets since the beginning of the year, including this one - http://www.youtube.com/watch?v=HRm6otBw8GU

AMD's problems with GF stem from the WSA as they have to buy USD 1.1 - 1.2 billion of wafers every year at GF. Otherwise AMD could have gone to TSMC for their entire APU production. TSMC has been rock solid with Kabini/Temash and now Xbox One SOCs production confirmed. PS4 SOC should also be at TSMC 28nm. For Kaveri AMD is stuck with GF.

TSMC doesn't have anything like the spare capacity to hand over to AMD, it was never an option and is unlikely to ever be an option. If they started doing big cores at TSMC something more important would have to give.

 

[Homeles]

Stilt already replied to that argument.

Actually, the quote I cited was a response to what you've got linked there.

I think the most benefit will be had on desktop parts of course as AMD intends to use FM2+ for all of their upcoming desktop products. One day maybe we could slide in 8C EX with 768 GCN SPs or sooner than that a 6C Kaveri. Desktop market is not expanding or growing but still is very important revenue stream.

Well, yes. Higher clocked APUs (e.g., desktop APUs) require higher bandwidth.

Kaveri should be pretty interesting on the desktop. 4 channels of 2133 MHz DDR3 would provide ~68.5GB/s bandwidth, if I'm doing the math correctly. That's just shy of the 7750's 72GB/s. If Kaveri does indeed have 16 ROPs, it will certainly need that kind of bandwidth. If clocks are high enough, it should shape out to be pretty fast.

Laptop APU models won't have such a big need for memory BW

You seem to be misunderstanding the argument being made. Quad channel support requires two extra 64 bit memory controllers to be bolted onto every single die, unless AMD starts offering two different die configurations. This means that even the low end Kaveri laptops will bear the cost of that extra die space, even if they don't offer quad channel support. I can grab some images to show you what I'm talking about, if necessary.

 

[Soulkeeper]

The idea of quad channel won't happen with ddr3 imo
ddr4 would be the solution for these problems
"DDR4 also anticipates a change in topology.[disputed – discuss] It discards the multi-drop bus approach used in previous generations in favor of point-to-point where each channel in the memory controller is connected to a single module." --wiki
"It discards the multiple DIMMs per channel approach in favor of a point-to-point topology where each channel in the memory controller is connected to a single DIMM." --wiki

 

[Homeles]

DDR4 won't be released in time for Kaveri.

 

[raghu78]

GF has been in volume production of 28nm since the end of last year. They have had those rockchip SoC's in tablets since the beginning of the year, including this one - http://www.youtube.com/watch?v=HRm6otBw8GU

http://www.globalfoundries.com/newsroom/2013/20130617.aspx

"GLOBALFOUNDRIES and Fuzhou Rockchip Electronics Co., Ltd. today announced that Rockchip’s next-generation mobile processors are ramping to production on GLOBALFOUNDRIES’ 28nm High-K Metal Gate (HKMG) process technology. Based on a multi-core ARM Cortex-A9 design, the RK3188 and RK3168 chips are optimized for tomorrow’s high-performance , low-cost tablets that require long-lasting battery life (see product specifications in annex). The combination of Rockchip’s design and GLOBALFOUNDRIES’ 28nm HKMG process technology resulted in a mainstream tablet System-on-Chip (SoC) capable of operating at up to 1.8 GHz performance, while still maintaining the power efficiency expected by mobile device users. The chips began sampling to OEMs in early 2013 and are now ramping to support a wide range of manufacturers."

So the way I understand it the RK3188 chip was sent to OEM's for product prototyping and testing. This too was in very low volumes in early 2013. only in late Q2 2013 RK3188 has ramped in production.

TSMC doesn't have anything like the spare capacity to hand over to AMD, it was never an option and is unlikely to ever be an option. If they started doing big cores at TSMC something more important would have to give.

looks like TSMC will have capacity to spare with GF taking few customers from TSMC. GF seems to be undercutting TSMC with lower wafer prices. TSMC being the market leader won't get into a price war.

http://www.digitimes.com/news/a20130805PD200.html

 

[SiliconWars]

You can see the RK3188 on show at CES this year in January.

http://www.youtube.com/watch?v=bMQfzsf8ACk

So if we say that was it sampling then, Rockchip was pretty fast and it is clearly in products in March. GF has always had a much faster ramp than TSMC as well so that will help.

On TSMC having spare capacity now, sure, but it's obviously way too late for AMD to do anything about that regarding Kaveri. The fact that Qualcomm and others are moving to GF is proof enough that the low power process works at least.

 

[SiliconWars]

Cointerra Launches New Bitcoin Mining Custom ASIC and Machines


The interesting part (unless you mine bitcoins I guess) -

I also asked exactly which 28nm process at Globalfoundries they were planning to use and he said that they would be using GloFo's 28nm HPP (High Performance Process) which means that this ASIC will likely have a high clock speed and high performance

 

[SiliconWars]

AMD once again deny any Kaveri delay. Seems they had a system running it with the stock HSF. Desktop at the end of this year, mobile early next year.

http://translate.google.com/transla...3-09/amd-bestaetigt-desktop-kaveri-fuer-2013/