XBitlabs: Advanced Micro Devices Set to Unveil New Strategy Next Week

[sm625]

Either they come up with a good design or they dont. A good design dominates desktop, server, and notebook segments. A bad design is covered in excuses about market segments and whatnot.

 

[RussianSensation]

1. You cant abandon high end desktop without also abandoning server and i dont think AMD will abandon the server market.

Well not necessarily. There is a new market for very power efficient, not as powerful, server CPUs. In 5-10 years, the really powerful servers may become a much smaller part of the marketplace, required for only the top corporations. It's quite feasible that small and medium-sized businesses will choose to move towards more efficient servers at some sacrifice in performance. Even the top 3 social network companies, as well as HP are interested in these extremely efficient servers.

2. They have no where near as much cash as they would need to break into the ARM/mobile market with players like samsung and qualcomm and Nvidia already in the door.

In desperate times, companies can commit to what at the time sounds like an all-or-nothing strategic change.

Look at Sprint - it put its entire business on the line with the iPhone. Who knows; if AMD's board is so determined to move into mobile and tablet space, expect the unexpected. :awe:

 

[AtenRa]

If Sandy Bridge didn't exist then your point would not be so easily refutable.

But Intel has clearly shown there is a path forward, at least up to the IPC performance point that Sandy Bridge currently fields.

AMD either threw in the towel early, rather than trying as hard as Intel's engineers did to improve IPC, or they guessed and guessed wrong about the value to come from prioritizing TLP and DLP over ILP.

In order for Intel to get a 10-20% more IPC they spend more and more resources, more money and more time each year. Increasing IPC is getting more difficult each passing day and the day of diminishing returns will come soon.

SandyBridge only has 10-15% more IPC Clock to Clock vs Westmere in Legacy code and IvyBridge will have even less than that. It seams that Haswell will get more TLP through more Pipes in the Execution Unit for higher scale of the HT. So, perhaps Intel will start to focus more in TLP from there on.

AMD doesnt have the resources, the money or the time to pursue the higher IPC road and they focused on the higher scaled TLP and DLP. Applications are coded for multi-threading more and more each day especially for Server Workloads.

Did they take the correct choice ?? time will tell

 

[Idontcare]

In order for Intel to get a 10-20% more IPC they spend more and more resources, more money and more time each year. Increasing IPC is getting more difficult each passing day and the day of diminishing returns will come soon.

SandyBridge only has 10-15% more IPC Clock to Clock vs Westmere in Legacy code and IvyBridge will have even less than that. It seams that Haswell will get more TLP through more Pipes in the Execution Unit for higher scale of the HT. So, perhaps Intel will start to focus more in TLP from there on.

AMD doesnt have the resources, the money or the time to pursue the higher IPC road and they focused on the higher scaled TLP and DLP. Applications are coded for multi-threading more and more each day especially for Server Workloads.

Did they take the correct choice ?? time will tell

Oh you'll get no argument from me, AMD has done remarkable things in terms of IPC/R&D-$ for their microarchitectures.

Inasmuch as Intel is showing us what is technically feasible for IPC on x86, AMD is showing us what is feasible when handed a much more constrained set of resources.

 

[Vesku]

AMD was forced to make choices that Intel will be making later, potentially a couple of years later depending on what they can squeeze out of Haswell.

 

[cbn]

I don't think ARM is a part of the strategy; in fact, ARM being so slow gives them an opportunity to attempt to get Brazos into tablets and ultrabooks. Whether they can execute is another story.

What do you mean by ARM being so slow?

Are you talking about processing speed? (The Cortex A15 looks to be pretty fast)

In fact, I would be very interested to see how the bobcat core compares to the A15 core: Which one is faster clock for clock?

 

[exar333]

What do you mean by ARM being so slow?

Are you talking about processing speed? (The Cortex A15 looks to be pretty fast)

In fact, I would be very interested to see how the bobcat core compares to the A15 core: Which one is faster clock for clock?

The A15 is targeting the Atom for performance, on a much smaller process node.

 

[cbn]

Seems to me they need to find a way to advance beyond their competition, in whatever market that is possible, and stick with it. It's time to put the "advanced" back in Advanced Micro Devices.

CUDA HPC seems to be a major target.

But how will they do this? (from a hardware standpoint)

Based on this, the company believes integration between CPU and GPU is more important than having the best discrete CPU or best discrete GPU.

The issue is not the CPU or the GPU, but rather memory bandwidth between the two devices and between the main memory they will share. This will involve stacking memory in 3D configurations on a chip package with these CPUs and GPUs.

Apparently Nvidia agrees with the importance of this AMD hardware strategy, clear evidence is their "Denver" plan with its ARMv8 CPU.

If the "Fusion" strategy pans out, then maybe it will be hard to criticize AMD for lack of single thread performance with Bulldozer. They did the best job they could allocating limited resources to the various projects. Maybe if they spent too much money on Bulldozer, they would have come up short on the integration budget?

 

[Vesku]

I'm not involved in programming for HPC applications but it does seem to be a big advantage for the gpu and cpu to be able to access the same pool of memory.

 

[Medu]

I dont know what they have planned but:

1. You cant abandon high end desktop without also abandoning server and i dont think AMD will abandon the server market.

2. They have no where near as much cash as they would need to break into the ARM/mobile market with players like samsung and qualcomm and Nvidia already in the door.

So hopefully they wont do something stupid but with there recent track record im sure they will anyways.

1. Not true. In fact Intel/AMD are been pushed into making Atom/Bobcat server parts.
2. Probably true. However AMD would have experts in CPU/GPU and chipsets which would make would reduce their learning curve.

 

[cbn]

The A15 is targeting the Atom for performance, on a much smaller process node.

Going by marketing graphs, I estimate Cortex A15 is faster than atom (in integer) by 14% clock for clock.

Here is how I came up with that number:

Using the above graph we can compare Specint_2000 (which measures single core integer performance) between 1.5 Ghz Moorestown vs a 1 Ghz Cortex A9 core. Measuring the heights of the graphs I come up with 1.5 Ghz Moorestown being 2x faster in SpecInt_2000 compared to 1 Ghz A9. After equalizing for Clock for clock, moorestown atom is 33% faster than Cortex A9 in integer)

According the above graph Cortex A15 increases integer performance 52% over cortex A9.

Therefore Specint_2000:

Cortex A9= 100%
Moorestown Atom= 133% (derived from first calculation)
Cortex A15= 152%

This would make Cortex A15 14% (1.52/1.33= 1.14) faster in Specint_2000 clock for clock compared to atom.

Other factors to consider:

1. Clockspeeds: If ARM A15 is able to achieve good efficiency at 2.5 Ghz in a Tablet things might get interesting.

2. Possible bottlenecks beyond Integer. (eg, floating point)

 

[cbn]

One major Wildcard on the consumer front of the x86 vs ARM war:

Google Android

What will happen to the licensing of this currently "free" OS.

If Google ends up charging for the OS, how much could this affect sales of ARM processors?

The way I see things a having free OS (especially one poised to become a new standard) really makes buying and upgrading cheap CPUs more attractive.

 

[Vesku]

I wouldn't extrapolate too much from a scaleless marketing slide, Computer Bottleneck.

 

[soccerballtux]

sounds like AMD's strat is to...not reveal a strat...just like they...didn't reveal dulldozer performance...

 

[soccerballtux]

Going by marketing graphs, I estimate Cortex A15 is faster than atom (in integer) by 14% clock for clock.

Here is how I came up with that number:

Using the above graph we can compare Specint_2000 (which measures single core integer performance) between 1.5 Ghz Moorestown vs a 1 Ghz Cortex A9 core. Measuring the heights of the graphs I come up with 1.5 Ghz Moorestown being 2x faster in SpecInt_2000 compared to 1 Ghz A9. After equalizing for Clock for clock, moorestown atom is 33% faster than Cortex A9 in integer)

According the above graph Cortex A15 increases integer performance 52% over cortex A9.

Therefore Specint_2000:

Cortex A9= 100%
Moorestown Atom= 133% (derived from first calculation)
Cortex A15= 152%

This would make Cortex A15 14% (1.52/1.33= 1.14) faster in Specint_2000 clock for clock compared to atom.

Other factors to consider:

1. Clockspeeds: If ARM A15 is able to achieve good efficiency at 2.5 Ghz in a Tablet things might get interesting.

2. Possible bottlenecks beyond Integer. (eg, floating point)

wait, I thought a15 was slower than a9?

edit: ok, I'm definitely wrong about this...

 

[cbn]

wait, I thought a15 was slower than a9?

edit: ok, I'm definitely wrong about this...

You are probably thinking of Cortex A5.

Yes, that CPU is definitely slower than Cortex A9.

 

[cbn]

multi-platform OpenCL/DirectCompute/Brook/etc. compilers, FI), focusing on making the GPU division somewhat profitable

AMD needs to figure out a way to get the folks at Khronos working harder on OpenCL.

I was very disappointed today when I checked and found out the latest stable release was the same version 1.1 release from June, 2010.

 

[soccerballtux]

You are probably thinking of Cortex A5.

Yes, that CPU is definitely slower than Cortex A9.

not to be confused with Apple's A5 right?

 

[soccerballtux]

AMD needs to figure out a way to get the folks at Khronos working harder on OpenCL.

I was very disappointed today when I checked and found out the latest stable release was the same version 1.1 release from June, 2010.

AMD is like google. Fail in execution at Polish stage every time.

 

[Idontcare]

AMD needs to figure out a way to get the folks at Khronos working harder on OpenCL.

I was very disappointed today when I checked and found out the latest stable release was the same version 1.1 release from June, 2010.

Did Apple wait for Khronos to work harder before they made their iPhone app store a success?

AMD's problem is they want to lead by standing back and letting everyone else run to wherever they are pointing.

Did Nvidia stand back and wait for someone else to develop CUDA for their hardware?

Had Sanders not retired then he might have had enough fire in his belly to lead AMD to forge its own future rather than relying on others to take them where they wanted to go.

AMD is headed towards the same obscurity the befell Via.

 

[piesquared]

As this page shows, AMD is more involved with OpenCL development than anyone else.

http://developer.amd.com/zones/OpenCLZone/Pages/default.aspx

It's up to developers to write the code and there are plenty of avenues for feedback, discussion and help.

 

[soccerballtux]

Did Apple wait for Khronos to work harder before they made their iPhone app store a success?

AMD's problem is they want to lead by standing back and letting everyone else run to wherever they are pointing.

wisdom, this one speaks.

 

[tatertot]

Not much longer until "Project WINNING!" is revealed...

Rumor has it that Charlie Sheen will be hired for all future APU design work.

 

[soccerballtux]

As this page shows, AMD is more involved with OpenCL development than anyone else.

http://developer.amd.com/zones/OpenCLZone/Pages/default.aspx

It's up to developers to write the code and there are plenty of avenues for feedback, discussion and help.

man look at that webpage they are so involved!

AMD isn't good at engaging developers like Nvidia is...

 

[Revolution 11]

Not much longer until "Project WINNING!" is revealed...

Rumor has it that Charlie Sheen will be hired for all future APU design work.

Which reminds me, I need to get rid of that last bit of AMD stock. Project Win doesn't give me much confidence.