So, where is AMD Seattle?

[mrmt]

But the Playstation 4 design already manages to solve some of these issues. The Jaguar cores still have lower IPC than a gaming console would like, but at least they're power-efficient. The GPU gets the fast RAM (GDDR5) that it needs to run many AAA titles at 1080p. And this is all done on 28nm. Add a better CPU architecture (doesn't need to catch up to Intel's newest, just narrow the gap), the new GPU technology in Tonga and beyond, HBM in place of GDDR5, all on 16nm/14nm FinFET+ for more CPU/GPU oomph at lower power usage... and suddenly you have a very competitive product indeed, for both the PC market and the next generation of consoles.

That product would give them the *budget* gamer market. The high margin/high profit gaming enthusiast market and the workstation market would stay with dGPU solutions, while the bottom of the barrel would stay with AMD. Not a viable TAM.

Particularly I think this is a "go big or go home" issue, AMD either develops something to outright kill 90-95% of the dGPU market or they should stop this entire fusion crap and go back to the basics before they go bankrupt, meaning low capacity graphics capacity, renewed focus in CPU performance as a cheap alternative to Intel (and only that) and renewed dGPU push sans HSA crap. Given that AMD lack the management capacity to conceive this product, they should go back to basics. Long term, we already know that AMD is no match for Intel, they should either refocus on eating Nvidia's bacon or follow Nvidia's lead on the market.

 

[Guest1]

AMD server solutions:

ARM: For those that require best perf/watt and price, and does not have to run x86 SW.
x86: For those that must run x86 SW.

Seems quite logical to provide both solutions. Especially since designing an ARM based chip based on an ARM core already designed by ARM is not that expensive to do.

Also, it's an entry to what will come, AMD's K12 (ARM) and Zen (x86) for future generations.

Correction:
x86: For those that require best perf/watt and price.
x86: For those that must run x86 SW.

http://www.forbes.com/sites/kurtmarko/2014/12/10/arm-vs-intel-servers/

 

[AtenRa]

and the GPU side doesn't perform all that much better than Intel's best iGPUs because DDR3 creates a hard bottleneck. The result is that Trinity, Richland, and Kaveri can barely play AAA titles at 720p30 low settings; a noticeably degraded experience. You lose power efficiency and CPU performance, and don't get much in return.

Clearly you dont have an idea of what your are talking about.

A 55-65W sub $100 A8-7600 is capable of playing almost every game at 720p and lots of them at higher Image Quality settings than low.

GTA V is perfectly playable at 720p Med settings

Hardline (SP) is perfectly playable at 720p Med

Rogue is playable with the 7850K even at 1080p

 

[Arachnotronic]

They seemed happy enough to ditch it for the latest generation. And "backwards compatible" means it is harder to sell HD remakes. (Or 4k remakes, I guess!)

There was a compelling reason to make the ISA transition the last time -- AMD, an x86 shop, was far more capable of supplying semi-custom chips than IBM was.

 

[BigDaveX]

An 8-core Steamroller with L3 would be able to compete very easily with 4C/8T Haswell in MT loads. Take Kaveri vs Core i3 Haswell and extrapolate.

It would? Because comparing a 3.5Ghz i3 (same base clock as a 4770K) to a 3.7GHz Kaveri and extrapolating the clock rates, it seems like they'd just end up back in the situation Piledriver was in against Ivy Bridge; a little faster in some heavily threaded scenarios, noticeably slower in others, and just obliterated in single-threaded loads. Doesn't really seem like it'd have been worth AMD's while to produce a full-blown FX version, especially when you consider that Devil's Canyon came out just a few months after Kaveri did, and would likely have totally erased the performance advantage of any hypothetical Steamroller-based FX.

 

[mrmt]

AMD server solutions:

ARM: For those that require best perf/watt and price, and does not have to run x86 SW.
x86: For those that must run x86 SW.

Best performance/watt is a must for almost any server scenario you can think of, that's why Intel is mopping the floor with AMD chips any time of the day. If AMD think it can get away with a subpar x86 architecture made from ip blocks of their ARM chips like Zen is, they are screwed. Zen chips will will end up like Bulldozer Opterons: not making sense inside a server even if they were offered for free.

 

[DrMrLordX]

where did you see RR killing the Fusion business ??? They went full throttle with APUs, OpenCL and HSA.

The original Fusion concept - pre-Llano - was killed before Rory took over. Or just sort of abandoned. But I digress.

Can we get back to Seattle here? The delays are horrid, and they don't say much good about its future prospects, but I expect that AMD is going to launch it anyway. Whether or not they have it out by Computex when they release their other new products for 2015 remains to be seen, and right now it seems to be a crapshoot. Yeah, their marketing department isn't doing a great job of framing it as a desirable product pre-release either. It might be that there isn't much to show.

AMD has to prove that their "ambidextrous" strategy is viable. They are putting as many eggs into the K12 basket as they are Zen, and that's not the only situation where we're supposed to see x86 and ARM coexisting on the same socket. Nolan and Amur are supposed to share the same platform as well.

I expect that the A57 in Seattle has no real problems. My guess is the platform design effort is running into trouble.

 

[monstercameron]

The original Fusion concept - pre-Llano - was killed before Rory took over. Or just sort of abandoned. But I digress.



Can we get back to Seattle here? The delays are horrid, and they don't say much good about its future prospects, but I expect that AMD is going to launch it anyway. Whether or not they have it out by Computex when they release their other new products for 2015 remains to be seen, and right now it seems to be a crapshoot. Yeah, their marketing department isn't doing a great job of framing it as a desirable product pre-release either. It might be that there isn't much to show.



AMD has to prove that their "ambidextrous" strategy is viable. They are putting as many eggs into the K12 basket as they are Zen, and that's not the only situation where we're supposed to see x86 and ARM coexisting on the same socket. Nolan and Amur are supposed to share the same platform as well.



I expect that the A57 in Seattle has no real problems. My guess is the platform design effort is running into trouble.

Platform design running into trouble or is that just fud? Afaik there are no public reviews or third party publications about Seattle but of course we hear that it performs lower and uses more power...does this type of fud remind you of any other chip recently?

 

[NTMBK]

By the way, it's interesting to go back to what AMD has said about Seattle. Right from the start, they lowered expectations. Listen to what Jim Keller actually said about their ARM plans:

We have a step-by-step plan. Seattle is a software constructor play to get that working. In '15 we build SoCs that are ambidextrous so that we've got that plumbing right.

https://youtu.be/SOTFE7sJY-Q?t=2m49s

Seattle in of itself doesn't need to sell particularly well, for AMD's plans- its purpose is to provide a software development platform, so that AMD can (for once) get their software ducks lined up in time for the big Zen launch.

 

[Abwx]

Seattle in of itself doesn't need to sell particularly well, for AMD's plans- its purpose is to provide a software development platform, so that AMD can (for once) get their software ducks lined up in time for the big Zen launch.

Eco system seems less advanced than the hardware, so far they look to delay the launch for this reason, surely at Computex.

Advanced Micro Devices this week confirmed that it plans to start volume shipments of its first 64-bit ARMv8-A-based Opteron microprocessor code-named “Seattle” in the second half of this year............

Originally, AMD planned to ship its Opteron A1100 “Seattle” in the second half of 2014. However, in mid-2014 the company only started to sell “Seattle” software development kits to interested parties.

Since software development typically takes a long time, demand for AMD’s “Seattle” processor today may be negligible, which is why the company does not initiate high-volume manufacturing.

http://www.kitguru.net/components/c...d-opteron-seattle-cpu-to-second-half-of-2015/

Edit : There s technical docs that have been released as late as january 2015...

 

[jpiniero]

Huh, that 24 to 48 Custom ARMv8 core server processor sounds intriguing. Talk about MOAR COREZ.

It would be very AMD to have ARM in servers take off and they manage nothing.

 

[mrmt]

Platform design running into trouble or is that just fud? Afaik there are no public reviews or third party publications about Seattle but of course we hear that it performs lower and uses more power...does this type of fud remind you of any other chip recently?

Every single AMD chip since Barcelona, with the honorable exception of the cat cores?

 

[DrMrLordX]

Platform design running into trouble or is that just fud? Afaik there are no public reviews or third party publications about Seattle but of course we hear that it performs lower and uses more power...does this type of fud remind you of any other chip recently?

AMD has released the SPECInt number for a 25W part, so there is that. And there's lots of delays. We can't really know for sure why there are delays . . . unless . . .

Seattle in of itself doesn't need to sell particularly well, for AMD's plans- its purpose is to provide a software development platform, so that AMD can (for once) get their software ducks lined up in time for the big Zen launch.

Okay, that's about what they're doing with the 300-series GPUs as well, which have allegedly been hardware-ready for months but are delayed so that they can iron out driver issues.

Eco system seems less advanced than the hardware, so far they look to delay the launch for this reason, surely at Computex.

Right, all signs point to a Computex launch. And what you quoted falls into line with what NTMBK said.

At this point it's all speculation until we actually see the launch hardware. If AMD's implementation of A57 runs about the same as everyone else's, and if the underlying platform is functional and stable, then we can reasonably confirm the validity of the claim that Seattle is delayed for no reason other than to give the software side more time to mature. Really, there will be no way for us to go much further than that in knowing what happened with Seattle over the course of the last year unless it launches this year with significant hardware problems.

 

[monstercameron]

Every single AMD chip since Barcelona, with the honorable exception of the cat cores?

I was thinking more about the fud qc had to face about overheating snapdragons.

 

[mrmt]

By the way, it's interesting to go back to what AMD has said about Seattle. Right from the start, they lowered expectations. Listen to what Jim Keller actually said about their ARM plan

If the purpose was to just validate the design framework and the platform details we wouldn't be seeing a production ramp up of the chip, and AMD at least until this quarter is committed to this.

 

[NTMBK]

By the way, it's interesting to go back to what AMD has said about Seattle. Right from the start, they lowered expectations. Listen to what Jim Keller actually said about their ARM plan

If the purpose was to just validate the design framework and the platform details we wouldn't be seeing a production ramp up of the chip, and AMD at least until this quarter is committed to this.

Hey, they have to fill that WSA somehow, might as well seed a bunch of discounted ARM devkits instead of paying yet another fee to GloFo!

(I'm only mostly joking...)

 

[DrMrLordX]

Hey, they have to fill that WSA somehow, might as well seed a bunch of discounted ARM devkits instead of paying yet another fee to GloFo!

(I'm only mostly joking...)

You know, I was thinking the same thing!

 

[Phynaz]

Platform design running into trouble or is that just fud? Afaik there are no public reviews or third party publications about Seattle but of course we hear that it performs lower and uses more power...does this type of fud remind you of any other chip recently?

Standard fallback plan. Start yelling FUD.

 

[Phynaz]

I was thinking more about the fud qc had to face about overheating snapdragons.

Is this FUD?
http://arstechnica.com/gadgets/2015/04/23/in-depth-with-the-snapdragon-810s-heat-problems/

 

[monstercameron]

Standard fallback plan. Start yelling FUD.

It is fud just in the fact that we only have a spec score and a tdp yet a few others claim that it is inferior.

 

[CHADBOGA]

Is this FUD?
http://arstechnica.com/gadgets/2015/04/23/in-depth-with-the-snapdragon-810s-heat-problems/

I think he is a Chuckles fanboy so unquestionably believes every claim he makes.D:

 

[mrmt]

It is fud just in the fact that we only have a spec score and a tdp yet a few others claim that it is inferior.

TDP is already public, and it was AMD itself who provided the only performance estimate. Given the cold shoulder the project got in the last call, when Lisa wasn't 100% committed to the production ramp up, I'd say that real world performance should be really bad.

 

[Nothingness]

It doesnt take more than a quick look on spec.org to see rates is very memory/cache focused.

Example.
i7 4790K 4 cores 4Ghz Haswell.
~225 INT_rates
~160 FP_rates

E5-2637v3(1chip) 4 cores, 3.5Ghz Haswell.
~244 INT_rates (Expected ~197)
~225 FP_rates (Expected ~140)

Benefit of 7MB more cache and quadchannel DDR4 is 23.8% for INT_rates and 60.7% for FP_rates.
Remember this is actually not that bandwidth limited like more cores would be.

Thanks! But that still doesn't make it a memory benchmark. It's sensitive to cache and memory like all server-class benchmarks should be. I'm probably too restrictive in what I call a memory benchmark

I'm still waiting for a technical article describing memory sensitivity of SPEC rate, but I won't be holding my breath. Or perhaps I'll do my own measures.

Anyway if that benchmark really was a pure memory benchmark I would expect Intel to rule it, as they have by far the best memory subsystem. (To clarify, I'm talking about Intel Core, not Silvermont.)

 

[ShintaiDK]

Thanks! But that still doesn't make it a memory benchmark. It's sensitive to cache and memory like all server-class benchmarks should be. I'm probably too restrictive in what I call a memory benchmark

I'm still waiting for a technical article describing memory sensitivity of SPEC rate, but I won't be holding my breath. Or perhaps I'll do my own measures.

Anyway if that benchmark really was a pure memory benchmark I would expect Intel to rule it, as they have by far the best memory subsystem. (To clarify, I'm talking about Intel Core, not Silvermont.)

If you followed the history. You would know AMD for exampel used the _rates benchmark to claim that Phenom was much faster than Core 2.

The bigger the difference between cores/memory the wider the rates gap compared to the non rates.

You can try it yourself as well. Run your memory at different speeds, even single channel. Since you may lack 12, 15 and 18 core parts at home

_rates got enough bandwdth focus to make a much slower CPU look faster, if said slower CPU got faster memory.

In the FP the Xeon was over 30% faster. So its roughly a delta of 70-80% between rates and non rates in this example.

 

[USER8000]

Sleepless.......!