AMD post-Bulldozer x86 CPU architecture

[AtenRa]

Both of you are actually handpicking. That puma+ @2.2Ghz is obviously throttling a lot in CB MT , maybe even at lower Ghz than the atom.
I will not comment until I see a proper mullins benchmark with real power consumption data. But I'm sure future 14nm cats will be a lot more competitive in performance/watt.

You cannot have everything in life, Mullins will most probably have higher power consumption than BayTrail.
One SoC (Mullins) gives you higher ST performance and 2-3x times faster Graphics and the other(BayTrail) have lower power consumption.

 

[Sweepr]

The reason I piked the Single Thread benchmark is because currently in Tablets and Extremely Low Power Devices, ST performance still matters more. And Puma is also faster in every ST benchmark in AT's review.

While ST is relevant for those devices, you can't ignore MT performance and BT-T is still very competitive here.

AMD could have created a smaller Die with lower iGPU performance and still be faster than BayTrail.

Bay Trail-T SoC (102mm²) is not the fastest Intel chip in terms of CPU perf/mm². Haswell 2C+GT1 is in another league perf/wise with a very similar 10xmm² die size (Kabini SoC is around 110mm², though it has an integrated NB). Also, I don't think 8 EUs would add up a lot of die space (probably <10mm²) and it would help BT-T a lot in this kinds of comparisons. Again, that's an aging Gen 7 architecture, things would look very different if Intel had a more competitive gfx architecture @ 22nm back in 2012/2013.

Truly if you thing of it, AMD is in another league and way ahead of Intel in the Mobile (Tablet) segment.

You can't say that without new benchmarks in an actual shipping product (not a reference platform), proper power consumption figures and at least a few 7-10'' design wins comparable in terms of size/weight/battery life to the current BT-T products available. The fact that 2013 BT-T can deliver the same MT performance in 7-8'' tablets as the fastest Mullins in a bulky 11.6'' reference platform begs to differ from this statement, when it comes to CPU performance.

Intel will most probable be in front in 2015 due to 14nm but the pain for Intel will come in 2016 when AMD will be on the low power high density 14nm FinFet process. If they could outclass Intel with 28nm, just imaging what they will bring with 14nm FinFets.

While Silvermont is already very competitive in CPU perf/watt and Intel is agressively pushing iGPU performance with Cherry Trail-T later this year, they will have a completely new CPU+GPU architecture (Goldmont + Gen 9 graphics) with 14nm Broxton, and this chip is expected to launch sometime in H2/2015, quite a while before the yet to be announced 14nm part you just mentioned.

Old ?? That is current BayTrail

Current Bay Trail-T are C0 parts with higher base clocks and 15% higher iGPU max Turbo clocks.

Let's stick to the thread subject now, shall we?

 

[witeken]

The reason I piked the Single Thread benchmark is because currently in Tablets and Extremely Low Power Devices, ST performance still matters more. And Puma is also faster in every ST benchmark in AT's review.

AMD could have created a smaller Die with lower iGPU performance and still be faster than BayTrail.
Truly if you thing of it, AMD is in another league and way ahead of Intel in the Mobile (Tablet) segment.
Intel will most probable be in front in 2015 due to 14nm but the pain for Intel will come in 2016 when AMD will be on the low power high density 14nm FinFet process. If they could outclass Intel with 28nm, just imaging what they will bring with 14nm FinFets.

Why do you ignore power consumption? You could say that AMD isn't far behind Intel with its FX-9590, but then you look at the TDP and it's indeed in another league entirely... yet still slower than Intel.

'Leagues' aren't created by performance, but by power consumption. By your logic, I could come up with an Intel Y CPU that destroys your Puma core, but since I don't know how much power they both consume, I can't claim anything about one being ahead in terms of efficiency.


There are always 2 things that you need to separate when comparing CPUs:

1. Ultimately, the whole package is what matters for the end user, so if it has higher performance, that's great for them. If your Puma core has higher performance than Bay Trail, it's probably a good core since you can find both in tablets and laptops (note that conclusions about performance can be more nuanced when CPUs throttle).

2. But that a core has higher performance while being in the same product category, doesn't mean anything about the real characteristics of the CPU. If Puma is only a low double digit percentage faster than Silvermont, but Silvermont consumes 2x less power, Silvermont is clearly the superior CPU (or combination of architecture and process).

Since we have no insight in power consumption, it's impossible to say that Puma is ahead of Intel. And since we have even less of an idea (read: totally none) of what AMD's nor Intel's 2016 architectures and transistors will look like, I advice you not to make any predictions about that.

Enough already. At this point the discussion needs to be AMD's next CPU architecture; anything else is thread crapping
-ViRGE

 

[itsmydamnation]

'Leagues' aren't created by performance, but by power consumption. B

bull, if I could produce a chip with twice the single thread performance of haswell with the same clock scaling but it consumed three time the power i would still sell everyone that i make and they would be playing in the exact same market ( datacentre/server).

Just because intel say power consumption matter across segments, doesn't mean it actually does.

 

[Homeles]

bull, if I could produce a chip with twice the single thread performance of haswell with the same clock scaling but it consumed three time the power i would still sell everyone that i make and they would be playing in the exact same market ( datacentre/server).

Just because intel say power consumption matter across segments, doesn't mean it actually does.

Niche products sell well to niches? You don't say!

Power consumption is important to the vast majority of markets that Intel and others cater to. Even your precious AMD is well aware of that.

 

[TreVader]

Intel needs to drop Bay Trail and focus on just making BroadWell faster.

They've cut down core functionality so much it's making Bay Trail look very weak in ST and because intel will never really stack up to AMD in iGPU they NEED ST performance to be on a level where they can at least stand toe to toe agains the best Apple, Samsung, and AMD SoCs.


While apple is widening the pipeline with A7 and adding functionality, intel has been chopping away at Haswell to make Bay Trail. They should just stop and figure out a way to make a Haswell quad that can fit into a phone. A8 is going to provide Haswell performance if they clock it anywhere close to 1.4Ghz and it's a quad.

And another infraction. The subject of the thread is AMD.
-ViRGE

 

[TreVader]

Mullins looks very promising though. I really hope AMD doesn't produce tons of A4 and A6 parts and have only a few high end A10 quad core mobile products.


The thing that bothers me most about AMD is their lack of focus on the very top of the line (because that's what I buy).

 

[Idontcare]

Just because intel say power consumption matter across segments, doesn't mean it actually does.

Actually it was AMD who said this and started the whole ball rolling in this direction back in 2002-2004 with their push to 90nm SOI and resultant marketshare grab in all market segments on the basis that power consumption mattered most.

Now that Intel beats them at their own game, how can the story change without causing a bit of a chortle in the audience?

Regardless, in any event it is not the supplier who is defining what matters to the segment; rather, it is the customer who decides whether or not the segment is viable on the basis of whether or not the products in that segment are flush with the necessary features as to compel the customer to purchase products within that segment.

It is within this perspective that Intel (and AMD previously) have noted that the customer appears to care enough about power consumption, regardless the segment or product, as to drive a requirement that you either capture this care-about in your feature list for the product offering or you risk having the customer not pursue purchasing your product (as AMD has experienced with their big cores, and Intel as well prior to recent years).

 

[itsmydamnation]

that all well and good, but you can't reconcile AMD's current big core position against what i said because what i said didn't occur.

also customers cant buy what doesn't exist. simple end to end life cycle TCO ( including software) big powerful core come out way cheaper. power consumption and cooling costs are a rounding error.

an example for the consumer space, look at the sony EXPRIA tab Z , a completely flawed device ( tiny battery for its specs) yet works brilliantly.

 

[mrmt]

bull, if I could produce a chip with twice the single thread performance of haswell with the same clock scaling but it consumed three time the power i would still sell everyone that i make and they would be playing in the exact same market ( datacentre/server).

That's roughly AMD's business proposition for Servers and Datacenters. The price they charge for Opterons is far below what Intel charges for Xeons of comparable performance, but as power consumption *do* matter for those customers, AMD market share was almost wiped out.

The only market segment that seems to be oblivious to power consumption is the desktop market, and it's not for anything that Rory Read is saying that AMD must grow on this market.

Given these facts, I can't really say how you can consider power consumption a bull of any kind.

 

[mrmt]

also customers cant buy what doesn't exist. simple end to end life cycle TCO ( including software) big powerful core come out way cheaper. power consumption and cooling costs are a rounding error.

Rounding error? In a data center? Would you mind expanding a bit on the subject?

 

[jpiniero]

bull, if I could produce a chip with twice the single thread performance of haswell with the same clock scaling but it consumed three time the power i would still sell everyone that i make and they would be playing in the exact same market ( datacentre/server).

The server markets don't care much about ST performance. It's all about multithread and perf/watt/$, especially with Virtualization being so hot these days. Consumers? They mostly only care about form factor/battery life/$.

Which leads me back to AMD. The K12, if done correctly, should be able to satisfy the majority of the market. I suppose this new x86 core could be a backup plan in case the K12 is a flop.

 

[itsmydamnation]

The server markets don't care much about ST performance. It's all about multithread and perf/watt/$, especially with Virtualization being so hot these days. Consumers? They mostly only care about form factor/battery life/$.

Which leads me back to AMD. The K12, if done correctly, should be able to satisfy the majority of the market. I suppose this new x86 core could be a backup plan in case the K12 is a flop.

yes they do ( PS i design converged datacentre uarch for a system intergrator for a living) , while facebook, google and amazon might not want ST performance, every single enterprise and government department does. here is why:

1. lots of single thread "legacy" code in inhouse developed applications that you just cant rewrite without massive cost, incredibly painful migrations etc.
2. software licencing costs, have you looked at the different licencing models across software stacks? have you looked at licencing costs of things like MS-SQL for example?
3. cost of writing multithreaded is higher then single thread ( for a single complex transaction).
4. cost of writing sea of cores core is higher then lightly multithreaded ( for a single complex transaction)
5. Most importantly is all of the above costs TIME, time that you could of had your expensive resources creating value for your company not shoe horning applications into threaded solutions.

In the very vast majority of datacentre analysis's ( including software stack) I have done, the most performant single core CPU perf wins by a mile. Government and enterprise system rarely services a million people with simple transactions, they service 100's of people with complex transactions.

Rounding error? In a data center? Would you mind expanding a bit on the subject?

when accounting for real end to end costs of a system, the cost of peoples time is the single biggest cost over its lifetime. cooling and power are cheap have linear scaling in costs and simple trade off can be made to reduce that cost ( run your datacentre @ 23,25,27,30 degree C instead of 21.3) .

That's roughly AMD's business proposition for Servers and Datacenters. The price they charge for Opterons is far below what Intel charges for Xeons of comparable performance, but as power consumption *do* matter for those customers, AMD market share was almost wiped out.

no it isn't, not at all, power consumption has nothing to do with, performance is why AMD is wiped out.

 

[Phynaz]

that all well and good, but you can't reconcile AMD's current big core position against what i said because what i said didn't occur.

also customers cant buy what doesn't exist. simple end to end life cycle TCO ( including software) big powerful core come out way cheaper. power consumption and cooling costs are a rounding error.

an example for the consumer space, look at the sony EXPRIA tab Z , a completely flawed device ( tiny battery for its specs) yet works brilliantly.

You've never seen the power bill from a data center, have you?

 

[frozentundra123456]

bull, if I could produce a chip with twice the single thread performance of haswell with the same clock scaling but it consumed three time the power i would still sell everyone that i make and they would be playing in the exact same market ( datacentre/server).

Just because intel say power consumption matter across segments, doesn't mean it actually does.

I am not an expert in the server market, but on the desktop, what you are saying is exactly the problem with AMD. They have greater power consumption *without* better performance, and for gaming and lightly threaded workloads, worse performance. I could accept higher power usage if it gave commensurately higher performance, but it does not. Kind of like a car that uses more gas, but is slower to accelerate as well.

 

[mrmt]

The server markets don't care much about ST performance. It's all about multithread and perf/watt/$, especially with Virtualization being so hot these days. Consumers? They mostly only care about form factor/battery life/$.

Why did the 16C bulldozer fail more spectacularly than the 8C variants on the market? Because ST does matter. Virtualization makes sure that you'll have your processors loaded most of the time, but once the cores are loaded, the performance of the individual cores matters a lot. Try to run a SAP Hana in a microserver and you'll see what I'm talking about.

 

[mrmt]

In the very vast majority of datacentre analysis's ( including software stack) I have done, the most performant single core CPU perf wins by a mile. Government and enterprise system rarely services a million people with simple transactions, they service 100's of people with complex transactions.

(...)

when accounting for real end to end costs of a system, the cost of peoples time is the single biggest cost over its lifetime. cooling and power are cheap have linear scaling in costs and simple trade off can be made to reduce that cost ( run your datacentre @ 23,25,27,30 degree C instead of 21.3) .

I think you are mistaken here. Cooling and power do *not* have linear scaling, neither in terms of OPEX nor in terms of CAPEX, and the trade off of raising the temperature is anything but simple (it has extreme impacts in your failure rates, with corresponding impacts in CAPEX/OPEX). I think you never saw the business plan or the balance sheet of a data center, or you completely misunderstood what you saw.

no it isn't, not at all, power consumption has nothing to do with, performance is why AMD is wiped out.

There are Opteron processors that have more performance than Xeon processors, but with disproportionate additional power consumption. If what you are saying was true AMD would have been selling tons of these processors, because they indeed offer better raw performance than some Xeon processors, but the fact is, they aren't.

If power consumption was really irrelevant, Intel itself wouldn't be selling low power processors for servers. They would sell only the power hogs and price them accordingly, but instead Intel charges (and consumers pay) a price premium for efficient Xeon processors.

The evidence that power consumption is a consumer trend, not an Intel trend, is in front of your eyes to see. I rest my case on the matter.

 

[itsmydamnation]

You've never seen the power bill from a data center, have you?

yes, yes i have, i have BUILT* datacentres, never seen the real end to end cost for an application have you?

by built i mean detail the technical specifications/requirements of basically everything.

 

[Phynaz]

So your answer is "No".

This is a technical forum. If you wish to disagree with someone find technical means to do so. This kind of sniping is unacceptable
-ViRGE

 

[itsmydamnation]

I think you are mistaken here. Cooling and power do *not* have linear scaling, neither in terms of OPEX nor in terms of CAPEX, and the trade off of raising the temperature is anything but simple (it has extreme impacts in your failure rates, with corresponding impacts in CAPEX/OPEX). I think you never saw the business plan or the balance sheet of a data center, or you completely misunderstood what you saw.

Temperature and humidity variance is what kills hardware not the absolute temperature. But again you keep going back to the "datacentre" the datacentre is but one cost.

There are Opteron processors that have more performance than Xeon processors, but with disproportionate additional power consumption. If what you are saying was true AMD would have been selling tons of these processors, because they indeed offer better raw performance than some Xeon processors, but the fact is, they aren't.

no i didn't, reread what i said, find me a processor from AMD that does that against a Xeon.

If power consumption was really irrelevant, Intel itself wouldn't be selling low power processors for servers. They would sell only the power hogs and price them accordingly, but instead Intel charges (and consumers pay) a price premium for efficient Xeon processors.

and i never said there wasn't a market for them. But that is driven by workload.

The evidence that power consumption is a consumer trend, not an Intel trend, is in front of your eyes to see. I rest my case on the matter.

thats nice.

 

[itsmydamnation]

So your answer is "No".

no it isn't. whats yours?
i ( former job) had to budget and pay the bills for our organisations datacentres.


love how no one here has address the application or the people cost of my posts which is the basis for my position.

 

[Phynaz]

My answer? Up to a couple of months ago I ran one. The power bill came out of my budget.

 

[itsmydamnation]

My answer? Up to a couple of months ago I ran one. The power bill came out of my budget.

that wasn't my question

 

[Phynaz]

I was Director of Systems Operations for a $20B healthcare company. I had 300 people reporting to me, and control of a $75M budget.

Don't tell me I don't know what it costs to run an application.

 

[itsmydamnation]

I was Director of Systems Operations for a $20B healthcare company. I had 300 people reporting to me, and control of a $75M budget.

Don't tell me I don't know what it costs to run an application.

i never did, but i notice how you are yet to answer the question.

As for you, you're done thread crapping
-ViRGE