AMD mulling break, spinoff

[cbn]

Two other approaches AMD could take to make a 4C/8T Zen:

1. Die harvested 8C/16T Zen.

Problems: It would be a low volume item based on a relatively large die (octocore plus L3 cache). Also, there is no iGPU.

2. Die harvested 4C/8T Zen APU with large iGPU.

Problems: If iGPU size is very large, then a small iGPU version harvested from a large die would also be low volume.

Overall, I think if AMD wants to compete with Intel they need to drive volume. Basing mainstream level 4C/8T Zen processors on large die products doesn't allow them to do this.

 

[VirtualLarry]

If AMD gets to Sandy Bridge level of IPC, I think they have a chance of being competitive.

I agree. I think those saying that it will reach Haswell-level IPC may be a tad optimistic, but either way, I'm willing to give AMD the benefit of the doubt, that they have indeed improved IPC significantly. They can do it, Phenom II wasn't far off from Intel's Core2 IPC, and they were contemporaries.

 

[dark zero]

Actually, going SMT, getting the correct instructions and putting the correct TDP it will make even goes at the best 5% less powerful as the Skylake... on moments of Canonlake. So they have a real chance here.

 

[JDG1980]

Even in that case, remember Phenom and 50% faster than Core 2? Something AMD had pubished and then removed again on their site.

Do you happen to have a link for this? I did some googling around, but wasn't able to find references to it.

AMD got away with lying many times before. Even insider trading. And it doesnt look like Lisa Su will last long in the chair. She is already having her fair share of the burden. Remember the lawsuits from the stockhodlers because she couldnt even follow AMDs own rules on stock options awards?

After looking up the stock option lawsuit, it looks like it wasn't something that Lisa Su did wrong, but rather a problem with the way the Board of Directors assigned the stock options that she was granted as part of her salary. She's apparently making $850,000 a year (which I think includes the options) - that's good money, of course, but not at all unusual for a CEO, especially not one with this many responsibilities and this precarious a company to handle. The stock option thing looks like a tempest in a teapot. As I'm not a stockholder, I have no reason to give a damn about it.

 

[JDG1980]

Fury won't change the picture on the notebook market, which is where Nvidia is mopping the floor with AMD GPUs. Basically three years being spanked on that market wasn't enough for AMD to change its course, and on top of that Fury won't come for the high volumes SKUs. Basically Fury is a lame duck for all that matters, it will be incapable of challenge Nvidia domination on the GPU market on the brackets that matter.

AMD could still make some decent profits on the Fury cards. Sure, volumes will be low, but per-unit margins are going to be higher than with cheap laptop chips or low-end dGPUs. Of course, this has to be offset against the R&D costs, but I expect that a large portion of Fiji R&D was for the HBM, and that expertise should help them get a leg up on more mainstream HBM2 products next year.

I definitely agree, though, that failure to create competitive products lower in the stack is a big problem for the laptop market. Frankly, I'm amazed that Apple decided to go with AMD GPUs for both the Retina iMac and the new MacBook Pro, especially since many of these products are so outdated (Pitcairn, Cape Verde). I wonder if they did so simply because they don't want to see a Nvidia monopoly in the GPU market and decided to throw AMD some pity business to keep them afloat. From a technical perspective, they should have gone with something like the GTX 970M or 980M (GM204) for the Retina iMac, and a GTX 950M (GM107) for the new MacBook Pro. Those would have provided superior performance at equivalent or lower TDPs.

 

[JDG1980]

P.S. Also remember Vishera (AMD's current 8 thread processor) is a big expensive die without iGPU. Whereas a 4C/8T Zen with small iGPU would be much smaller and have the versatility to be used in a wider range of devices.

What do you define as a "small" iGPU? Remember, compute units only make up a part of a GPU's die space. You've also got some room occupied by fixed-function units (UVD, VCE, TrueAudio, etc.) and memory controllers (which, in an APU, are going to be shared between CPU and GPU in any case).

I suspect there would be negligible savings in going below 8 CUs (512 shaders). Note that Carrizo, despite being designed for low-power laptop usage, has that many CUs - and this is on a 28nm process. With 14nm FinFET, the die space occupied will be even smaller (though it may be more expensive on a per-sq.mm. basis, at least at first).

Based on AMD's leaked roadmap, it doesn't seem like we're going to get what you are looking for. They plan to bring out Summit Ridge (server/HEDT Zen with no iGPU) in mid-2016, then Raven Ridge (mainstream Zen APU) in 2017. Summit Ridge will be 8C/16T (though of course there will probably be die-harvested parts), but we don't know yet what Raven Ridge will be - the roadmap doesn't say.

It's important to note that AMD, during Financial Analyst Day, specifically said they want to get away from the low end. That's why I think Raven Ridge will be a more ambitious product than just a next-generation version of the current APUs - after all, the current products are largely failures. I wouldn't be surprised if Raven Ridge was 4C/8T with a really beefy iGPU, the most powerful seen so far - something like 2048 shaders (Tonga equivalent) would be well within plausibility on 14nm FinFET. This would all be tied together with shared HBM2 (8GB or 16GB depending on SKU), thus erasing the disadvantage of existing APUs and providing every bit as much performance as one of today's $200 dGPUs. Of course, it would be more expensive than today's APU products, probably about on par with the cost of Intel's Broadwell Iris Pro offerings. With something like this, AMD could explore new markets in set-top gaming (small form factors become much easier), get more console design wins, and maybe score a contract for Retina iMacs as well.

 

[jpiniero]

The K12 project is one of the most mindbogling projects they ever made. And hopefully they cancel is sooner than later. I cant see any customers for it, and even worse. I cant see it compete with A72 etc. It seems yet again that someone had a meeting and thought something would be a good idea without doing the research.

If anything Zen is worse since they are going to lose the x86 license one way or another eventually. The only way Zen makes sense is if they have some sort of plan to go after the corporate market, and I haven't seen it. Mubdubala may have just finally seen the light and is trying to get something for AMD while they still can.

 

[cbn]

What do you define as a "small" iGPU? Remember, compute units only make up a part of a GPU's die space. You've also got some room occupied by fixed-function units (UVD, VCE, TrueAudio, etc.)

A small iGPU is a GPU that serves basic needs of assisting UI acceleration, flash, etc.

In 2017, for a laptop how many GCN sps do we need for that? On what size screen?

P.S. According to this Anandtech Kabini article each cluster of 128sp is ~16.1 mm2 on 28mm2. On 14nm (assuming 2 to 1 scaling) that would amount to around 8mm2 per 128sp.

Given that one core is 3.1 mm2, extrapolating out gives the size of the die at 31.4x the size of a single core, or 97.3 mm2. The GPU area is approximately 5.2x the size of a core, giving ~16.1 mm2 for 128 GCN cores, compared to 12.4 mm2 for CPU cores. The Video Codec Engine and Unified Video Decoder are not part of these totals, located on other parts of the APU. The memory controller clocks in at ~9.4 mm2 and the display/IO portion runs at ~7.3 mm2.

 

[cbn]

It's important to note that AMD, during Financial Analyst Day, specifically said they want to get away from the low end.

Yes, and with a 4C/8T small iGPU APU AMD does just that compared to a Cat core based tablet SoC or a 2C/4T medium iGPU chip.

4C/8T Zen is more high end that 2C/4T Zen or 4C/4T cat core.

 

[JDG1980]

If anything Zen is worse since they are going to lose the x86 license one way or another eventually.

This won't happen.

 

[Tsavo]

If anything Zen is worse since they are going to lose the x86 license one way or another eventually.

They'll never lose their x86 agreement with Intel unless they go under or sell the business.

 

[JDG1980]

A small iGPU is a GPU that serves basic needs of assisting UI acceleration, flash, etc.

In 2017, for a laptop how many GCN sps do we need for that? On what size screen?

With ultra-high-resolution screens becoming more common on laptops, even standard desktop work needs more GPU grunt than it used to. Note how Intel's Iris/Iris Pro was more or less designed with Apple's Retina MacBook Pro needs in mind. The standard Intel iGPU wasn't cutting it.

I don't see any good reason for AMD to go below 8 CUs (512 SPs) on-die for any of their Zen APUs, even the smallest.

P.S. According to this Anandtech Kabini article each cluster of 128sp is ~16.1 mm2 on 28mm2. On 14nm (assuming 2 to 1 scaling) that would amount to around 8mm2 per 128sp.

Exactly. If AMD can quadruple the graphical power of its APU (128->512 shaders) for only 24 sq.mm. of extra die space, why wouldn't they? Even if it goes unused most of the time, the cost is negligible, and having it there looks good on the specs and offers the option of OpenCL performance in apps that need it. Apple seems to like this, and I'm sure AMD wants to win as many Apple contracts as they can.

 

[JDG1980]

Yes, and with a 4C/8T small iGPU APU AMD does just that compared to a Cat core based tablet SoC or a 2C/4T medium iGPU chip.

4C/8T Zen is more high end that 2C/4T Zen or 4C/4T cat core.

Who do you anticipate being the market for this hypothetical product? Remember, DIY enthusiast CPUs are a very small portion of sales. You can't justify the cost of a new CPU die configuration just for enthusiasts (which is why HEDT chips are basically repurposed server chips). Which OEMs are going to want to go in this direction?

 

[cbn]

Who do you anticipate being the market for this hypothetical product?

My experience with AMD CPUs (E1-2100 cat core APU , Sempron 2650 cat core APU, Athlon 5350 cat core APU, and Athlon x4 860K) is that the cat core iGPU is sized right, but the Jaguar cpu (even on Athlon 5350) is too weak.

On the other hand, the Steamroller CPU Athlon x4 860K CPU is not bad, but I think 512sp iGPU (even if AMD had the bandwidth solved) would be too much.

Instead, I would like to have seen something with quad Steamroller, but with small iGPU (from the cat core family) as AMD's high volume chip.

 

[cbn]

You can't justify the cost of a new CPU die configuration just for enthusiasts (which is why HEDT chips are basically repurposed server chips). Which OEMs are going to want to go in this direction?

http://forums.anandtech.com/showpost.php?p=37497209&postcount=6

 

[cbn]

With ultra-high-resolution screens becoming more common on laptops, even standard desktop work needs more GPU grunt than it used to. Note how Intel's Iris/Iris Pro was more or less designed with Apple's Retina MacBook Pro needs in mind. The standard Intel iGPU wasn't cutting it.

According to the following thread, the problem you are referring to was software related:

https://discussions.apple.com/thread/5547471?start=0&tstart=0

 

[cbn]

Exactly. If AMD can quadruple the graphical power of its APU (128->512 shaders) for only 24 sq.mm. of extra die space, why wouldn't they? Even if it goes unused most of the time, the cost is negligible, and having it there looks good on the specs and offers the option of OpenCL performance in apps that need it.

For a chip (4C/8T Zen) that is meant to be a high end replacement for Cat Core SoCs, 24mm2 on 14nm is a lot of space.

 

[Abwx]

https://translate.google.fr/transla...altung-oder-ausgliederung/&edit-text=&act=url

.......................................................

 

[ShintaiDK]

Do you happen to have a link for this? I did some googling around, but wasn't able to find references to it.

It is getting hard to find on google due to bulldozer replacing the failure in searches.

http://www.zdnet.com/article/amd-posts-blatantly-deceptive-benchmarks-on-barcelona/

It is somewhat easier to search on the 40% projection that grew:
http://www.dailytech.com/AMD+Expect...n+by+40/article5863.htm?www.reghardware.co.uk

And then there was the spec.org part:
http://www.dailytech.com/SPEC+Inval...nchmarks+due+to+Availability/article10030.htm

 

[ShintaiDK]

https://translate.google.fr/translate?sl=de&tl=en&js=y&prev=_t&hl=fr&ie=UTF-8&u=http%3A%2F%2Fwww.computerbase.de%2F2015-06%2Funternehmensstruktur-amd-dementiert-aufspaltung-oder-ausgliederung%2F&edit-text=&act=url

.......................................................

AMD also denied it would sell its fabs.

http://www.eweek.com/c/a/Desktops-and-Notebooks/AMD-Denies-Fab-Sell-Off

 

[nemesismk2]

i am sure that amd will be aound for a long time to come.

 

[CHADBOGA]

i am sure that amd will be aound for a long time to come.

Just like VIA, I guess.

 

[Tsavo]

A neat idea would be for them to spin off their utterly pointless board of misdirectors.

 

[mrmt]

AMD could still make some decent profits on the Fury cards. Sure, volumes will be low, but per-unit margins are going to be higher than with cheap laptop chips or low-end dGPUs. Of course, this has to be offset against the R&D costs, but I expect that a large portion of Fiji R&D was for the HBM, and that expertise should help them get a leg up on more mainstream HBM2 products next year.

AMD basically got a shot on the high end, low volume, high priced desktop SKUs and lost the entire high volume desktop *and* notebook SKUs. Unless HBM can make wonders for GPGPU, they will lose even more market share this year and revenue.

HBM is so far proving to be a foolish venture, as the technology is not ready to go mainstream and is confined to the very high end.

 

[mrmt]

A neat idea would be for them to spin off their utterly pointless board of misdirectors.

One of the requisites for a spin off is that you end up with two viable businesses, and I doubt those clowns could make *anything* work. They probably could bankrupt an oil well in Saudi Arabia if they tried hard enough.