New AMD Roadmap for 2013 (donanimhaber.com)

[pelov]

What AMD brings to the table that would be unique? Because they won the console battle but this does not mean anything, as we know that console wins are a race to the bottom.

Whether or not they would make an interesting option for Apple would hinge upon Temash and Kabini's GPU performance. If AMD can get the Temash chips at sub-5W TDP and great GPU performance then Apple would definitely be interested.

Keep in mind that it was Apple's push toward GPGPU that created openCL too, then AMD has been pushing it as a standard ever since. That too would imply some incentive.

There's also discrete GPUs as well as Apple being interested in AMD's APUs for two generations now, though they missed out with Llano because of GloFo's yield and capacity issues early on with 32nm. Obviously the x86 license would be tossed out, but with AMD dabbling in ARM it certainly opens up some new doors

 

[mrmt]

Whether or not they would make an interesting option for Apple would hinge upon Temash and Kabini's GPU performance. If AMD can get the Temash chips at sub-5W TDP and great GPU performance then Apple would definitely be interested.

Interested? Sub 5W chips aren't really Apple thing, as it is too hot to bring to tablets and too weak to power macbooks. Plus whenever Intel came with some better Atom interation Apple would toss AMD out of the window.

All in all Apple seems pretty committed to custom ARM designs, they definitively want to bring chip design on board.

Keep in mind that it was Apple's push toward GPGPU that created openCL too, then AMD has been pushing it as a standard ever since. That too would imply some incentive.

Not at all. As you can imagine OpenCL is, well, open, so everybody can come and design OpenCL chips. AMD has no advantage here too, or at least not the advantage needed to bring Apple to the other side of the table.

There's also discrete GPUs

Discrete market is going to shrink.

as well as Apple being interested in AMD's APUs for two generations now, though they missed out with Llano because of GloFo's yield and capacity issues early on with 32nm. Obviously the x86 license would be tossed out, but with AMD dabbling in ARM it certainly opens up some new doors

Do you really think that yields were the definitive issue for Apple not going to AMD? They weren't. If, let's say, Apple wanted to launch a Llano Macbook in december, this contract would have to be locked in at least one year before.

Designs would have to be made, orders placed, drivers developed and tested, support trained, notebooks assembled and then you have your macbook ready.

AMD didn't have to know how Llano yields would be one year before launch.

What probably happened is that AMD floated the idea, Apple liked the AMD chip and said that it would suit them good, but then they looked at AMD then current track record, manufacturing capacity and product roadmap and gave up this idea. It is clear by now that partnering with AMD would have been a very bad idea.

 

[lifeblood]

Apple buying AMD is brilliant! Apple could do some of that Reality Distortion voodoo, make everyone believe that Bulldozer is faster than anything Intel has, everyone will buy it, then AMD will be saved!

Of course, they'll be owned by Apple so I will have to hate them...

 

[Vesku]

Now that it has been "upgraded" to use 16-core bulldozer-based operton CPUs, the compute power that comes from the CPUs themselves has actually fallen from ~2.3PFlops to only ~2PFlops...the Nvidia GPU's account for 90% of the spec'ed 20PFlops capability of Titan.



That is quite the upgrade from AMD's side of the equation, 33% more cores and only 10% lower performance :thumbsup:

Is there any wonder why so few opterons are sold in the server segment anymore? Who wants them when that is what they have to offer versus the competition (or just stick with your existing stars-based opterons you bought 2 yrs ago).

That's floating point operations, so it actually went from 12 Magny FPUs to 8 (4 in AVX mode) Bulldozer FPUs. 33% less operating units 10% less performance. Not the best choice for floating point intensive supercomputing projects. Must have been some good salesmanship on Cray's part.

IMO, AMD has a good play with APUs using the CMT concept. It's just they bungled allocation of resources and picked the wrong end of the industry in terms of targeting initial designs. Trinity should have been out before the non-APU version of Bulldozer. Then they could've shaved at least a quarter or two off of GCN APU launch timeframe. Plus they probably would have been less obsessed with clockspeeds and more focused on Perf/W if the APU had priority. Not sure if there is a corporate executive with enough brass to refocus a multi-delayed multi-billion dollar project.

As the much smaller player AMD should be nimble enough to fill in new niches before Intel gets interested in them. Unfortunately their profits looking like a sine wave has been a real drag on AMD's ability to fund exploratory projects.

Edit: I think the ARM announcement is a good thing in this last regard (nimbleness) especially if they can keep the ARM and x86 designs working under the same HSA umbrella.

 

[veri745]

If AMD backs out of the x86 market, can they cancel their cross-licensing agreement with Intel so neither one of them can use x86-64?

When is the next time the cross-license agreement is up for renewal?

 

[mrmt]

If AMD backs out of the x86 market, can they cancel their cross-licensing agreement with Intel so neither one of them can use x86-64?

Not likely to happen. Next generation console chips are most likely x86 chips that leverage heavily on Intel x86 IP, so if AMD walk away from the agreement it would not be able to manufacture Xbox and PS4 chips, and would be liable to billions in damages.

 

[jpiniero]

Not likely to happen. Next generation console chips are most likely x86 chips that leverage heavily on Intel x86 IP, so if AMD walk away from the agreement it would not be able to manufacture Xbox and PS4 chips, and would be liable to billions in damages.

You would have to believe that MS and Sony wouldn't be using AMD x86 unless they felt confident that regardless of what happens to AMD they will be able to get the processors.

 

[carop]

Now that it has been "upgraded" to use 16-core bulldozer-based operton CPUs, the compute power that comes from the CPUs themselves has actually fallen from ~2.3PFlops to only ~2PFlops...the Nvidia GPU's account for 90% of the spec'ed 20PFlops capability of Titan.

Disruptive technology is a term coined by Harvard Business School professor Clayton M. Christensen to describe a new technology that unexpectedly displaces an established technology. As such, NVIDIA disrupted the HPC industry with the release of CUDA in February 2007. NVIDIA brought supercomputing to the masses with its low cost teraflop/sec GPU hardware and co-processor acceleration of both C and Fortran applications. Intel is following suit with its MIC co-processor announcement.

The "Stampede" supercomputer at The University of Texas at Austin appears to be the first large scale deployment of MIC co-processors:

http://www.tacc.utexas.edu/news/press-releases/2011/stampede

Stampede will have a peak performance of 10 petaflops. Sandy Bridge-EP Xeon E5 processors will provide 2 petaflops of peak performance and the MIC co-processors will provide an additional 8 petaflops of performance.

AMD Opterons account for 10% of the performance of Titan and Intel Xeons account for 20% of the performance of Stampede because NVIDIA Kepler and Intel MIC are more power efficient.

NVIDIA Kepler achieves high flops/watt efficiency through the use of an SIMD execution model inside each SM (streaming multiprocessor) that requires less supporting logic than non-SIMD architectures. GPU architects have been able to devote more power and space to 64-bit addressing, additional ALUs, and floating-point units.

Intel MIC achieves high flops/watt efficiency by leveraging the simplicity of the original in-order short execution pipeline of the Pentium design and the power savings of the 22 nm manufacturing process.

As far as AMD Opterons are concerned we all know that AMD has been under performing since the introduction of the Core 2 Duo micro-architecture in 2006. We also know that Bulldozer was not the answer to AMD's problems.

 

[Ajay]

I think the point is "who" is left for Keller to lead? Their is the morale issue for those who remain, and then the question of what type of engineer is going to typically hang around versus moving on, etc, when so much writing is on the wall.

Why would Keller move to AMD to go from being a lead architect to creating ancillary IP for ARM SOCs? I just don't get this. If AMD was building a custom ARM SOC, then OK, maybe they threw a ton of money at him, but since AMD is using a bog standard ARM SOC, Keller must be bored out of his mind.

 

[AtenRa]

I thought it was rather foreboding the fact that when the supercomputer Titan was still called Jaguar it had exactly the same number of CPUs (they were Magny-Cours though, not bulldozer based) but it pulled in ~2.3PFlops.

Now that it has been "upgraded" to use 16-core bulldozer-based operton CPUs, the compute power that comes from the CPUs themselves has actually fallen from ~2.3PFlops to only ~2PFlops...the Nvidia GPU's account for 90% of the spec'ed 20PFlops capability of Titan.



That is quite the upgrade from AMD's side of the equation, 33% more cores and only 10% lower performance :thumbsup:

Is there any wonder why so few opterons are sold in the server segment anymore? Who wants them when that is what they have to offer versus the competition (or just stick with your existing stars-based opterons you bought 2 yrs ago).

Sorry but your numbers are wrong,

AMD Dual(2x) Opteron 6174
SPECfp2006 = 30,7

AMD Dual(2x) Opteron 6274
SPECfp2006 = 42,8

33% more cores gives you ~39,5% more in SPECfp2006

You almost have 40% more performance in the same space with the same power consumption (Both Opteron 6174 and 6274 are 115W TDP)

And from AT's Titan article, it is 20+ petaflops

The total computational power of Titan is expected to be north of 20 petaflops.

 

[mrmt]

You would have to believe that MS and Sony wouldn't be using AMD x86 unless they felt confident that regardless of what happens to AMD they will be able to get the processors.

I agree with that. My answer was in the context of AMD trying to pull out of the agreement, which won't happen.

 

[acx]

The table is using double precision FLOP as perf comparison. Jaguar uses Opteron 2435, 6 Istanbul cores with 2 128bit FPUs each at 2.6 Ghz. Titan uses Opteron 6274, 8 Bulldozer modules with 2 128 bit FPUs each at 2.2 Ghz. Each FPU can do two double precision operations per cycle and have the same throughput on both architectures. Opteron 2435 max double precision rate is 62.4 GFLOPs. Opteron 6274 max double precision rate is 70.4 GFLOPs. Titan's CPUs will provide about 13% more max double precision FLOPs than Jaguar CPUs.

Note when looking at Top500 Linpack performance numbers the Rmax value is the theoretical maximum throughput of the system while Rpeak is the measured peak performance on Linpack. The table from anandtech uses peak to mean the max theoretical maximum. Don't confuse the table's "peak" to be the same as Rpeak on Top500 Linpack lists.

BTW Jaguar was upgraded in stages. The Top500 Linpack numbers for Jaguar with all the Opteron 2435 swapped out for Opteron 6274 were done in time for the June 2012 Top500 list. Maximum theoretical performance was 2.63 PFLOPs and measured peak performance was 1.94 PFLOPs. Compare to old Jaguar with Opteron 2435 which had theoretical max of 2.33 PFLOPS and a measured peak of 1.76 PFLOPs. Although the June 2012 list notes that some Tesla 2090's were installed in Jaguar I don't think the Linpack run was done with those GPGPU modules enabled as only a small portion of the cabinets had Tesla 2090's and the theoretical maximum does not seem to include the Tesla 2090 FLOPs.

 

[inf64]

Each BD's FMAC within the FP unit can do 4 DP flops with FMA,double of what Istanbul core can do.

 

[Abwx]

Sorry but your numbers are wrong,

AMD Dual(2x) Opteron 6174
SPECfp2006 = 30,7

AMD Dual(2x) Opteron 6274
SPECfp2006 = 42,8

Seems that viral marketing was so widespread that even people
like IDC did fall in the trap and dont even check numbers ,
instead relying on non sensical rumours....

It tells a lot about the continual trashing that unrelentlessly
occur here and there.

 

[ShintaiDK]

If AMD backs out of the x86 market, can they cancel their cross-licensing agreement with Intel so neither one of them can use x86-64?

When is the next time the cross-license agreement is up for renewal?

They cant cancel the license. Then Intel gets all the IP.

When the license is up for renewal, then the x64 is no longer under IP.

 

[ShintaiDK]

Sorry but your numbers are wrong,

AMD Dual(2x) Opteron 6174
SPECfp2006 = 30,7

AMD Dual(2x) Opteron 6274
SPECfp2006 = 42,8

33% more cores gives you ~39,5% more in SPECfp2006

You almost have 40% more performance in the same space with the same power consumption (Both Opteron 6174 and 6274 are 115W TDP)

And from AT's Titan article, it is 20+ petaflops

Not to state the obvious but...

You do know specfp is singlethreaded right? And you need to use specfp_rate for multithreaded and throughput? Also the software versions aint the same.

Seems that viral marketing was so widespread that even people
like IDC did fall in the trap and dont even check numbers ,
instead relying on non sensical rumours....

It tells a lot about the continual trashing that unrelentlessly
occur here and there.

Just gonna keep this for fun

 

[Ferzerp]

Let's make the AMD guy's head explode by pointing out that only a *tiny* fraction of the 20 pflops can be attributable to AMD and that the vast majority of it (like 90%) is nvidia.

ORNL didn't chose AMD, they were stuck with it because cray is their vendor and cray is stuck with it because they made agreements before reality had set in.

The opterons are little more than feeders for the real HP in that system, the nvidia GPUs.

 

[ShintaiDK]

Why would Keller move to AMD to go from being a lead architect to creating ancillary IP for ARM SOCs? I just don't get this. If AMD was building a custom ARM SOC, then OK, maybe they threw a ton of money at him, but since AMD is using a bog standard ARM SOC, Keller must be bored out of his mind.

Not to be blunt, but who says Keller is anything like his old days today?

He is 53 now, not 42 or 43.

 

[Rvenger]

Not to be blunt, but who says Keller is anything like his old days today?

He is 53 now, not 42 or 43.

So what are you trying to say?

 

[ShintaiDK]

So what are you trying to say?

That people change. And getting older doesnt mean you get more energy.

 

[Phynaz]

Makes you wonder how/if AMD is going to keep their promise of AM3+ compatibility for Steamroller...

There was no promise from AMD. Just a gossip website saying that AMD was going to do this, without any sources identified.

 

[frozentundra123456]

That people change. And getting older doesnt mean you get more energy.

Well, if you lose your energy at 53, I must be dead already and dont know it!!!


Seriously, 53 is kind of the prime age for businessmen, politicians, etc. Who really knows RR's motivations. Personally, I cant imagine anyone with a resume as good a his going to a troubled company like AMD, but I seriuosly doubt that it is because he is washed up at 53.

 

[AtenRa]

Not to state the obvious but...

You do know specfp is singlethreaded right? And you need to use specfp_rate for multithreaded and throughput? Also the software versions aint the same.



Just gonna keep this for fun

No you dont need to use specfp_rate for Multithreaded and Throughput. Both of my links above have Auto Parallel = YES.

http://www.spec.org/cpu2006/Docs/readme1st.html#Q15

Q15. What is the difference between a "rate" and a "speed" metric?

There are several different ways to measure computer performance. One way is to measure how fast the computer completes a single task; this is a speed measure. Another way is to measure how many tasks a computer can accomplish in a certain amount of time; this is called a throughput, capacity or rate measure.

The SPECspeed metrics (e.g., SPECint2006) are used for comparing the ability of a computer to complete single tasks.
The SPECrate metrics (e.g., SPECint_rate2006) measure the throughput or rate of a machine carrying out a number of tasks.

For the SPECrate metrics, multiple copies of the benchmarks are run simultaneously. Typically, the number of copies is the same as the number of CPUs on the machine, but this is not a requirement. For example, it would be perfectly acceptable to run 63 copies of the benchmarks on a 64-CPU machine (thereby leaving one CPU free to handle system overhead).

Note: a speed run which uses a parallelizing compiler to distribute one copy of a benchmark over multiple CPUs is still a speed run, and uses the SPECspeed metrics. You can identify such runs by the field "Auto Parallel".

 

[jones377]

So am I the only one wondering about Richland? We expected a Steamroller APU with GCN and instead they put a Piledriver APU with Radeon Cores 2.0 on the roadmap. Is it Trinity 2.0 (the same way Brazos 2.0 appeared) or does Radeon Cores 2.0 mean GCN? If so, what is Radeon Cores 1.0 then? We have VLIW5 in 1st Gen Llano APU, VLIW4 in 2nd Gen Trinity APU and now Radeon Cores 2.0 in 3rd Gen. Is it GCN or VLIW4?

 

[Ajay]

Not to be blunt, but who says Keller is anything like his old days today?

He is 53 now, not 42 or 43.

There is no way to tell. Unless you happen to know someone from Apple or PA Semi.