AMD - Piledriver chitchat

[scaryjim]

Odd seeing intel using Mali graphics ....

Edit to add: Hang on, are those RockChip manufactured parts? ...

I suspect that's got something to do with the choice of Mali. Also, these are smartphone chips - ultra-low TDPs abound, and I don't know how good HD is under those constraints. It may simply be that it's cheaper and more performant to use a genuine mobile graphics IP, and Rockchip have plenty of experience integrating Mali into SoCs.

[shaithis]

Not sure if legit: http://videocardz.com/55358/xfx-rade...les-for-camera

That looks very small, although as usual I really do not like the look of those custom XFX coolers.......

[DanceswithUnix]

More Zen specifics from FudZilla. Usual pinch of salt being applied

http://www.fudzilla.com/news/process...n-apu-detailed

[watercooled]

Along with the implied die size, quad channel ECC DDR4 implies server/workstation platform. Still, it's impressive if true as it would give AMD something to compete with in that space, and the on-board GPU could be a big selling point for the HPC space where it is currently very common to use CPU + Firepro/Tesla combinations.

[scaryjim]

Along with the implied die size, quad channel ECC DDR4 implies server/workstation platform. ...

Yeah, it's nothing new - we've already discussed the fact that there's an HPC focussed APU coming (see my comment here: http://forums.hexus.net/cpus/241925-...ml#post3447280). Only real surprise is having dedicated GPU HBM as well as multi-channel DDR4, although presumably part of the HSA spec would allow it all to be addressed as one big memory space. Just wonder how hard it'll be to program with unified addressing across two memory spaces with massively different bandwidth....

[watercooled]

The new Xeon Phi seems to do something similar, so it seems like it could become a common thing for the HPC space. I wonder if it will be local, addressable memory or some sort of cache?

I think a few places are assuming this is the Carrizo/Vishera replacement and arguing over whether it will be good for games.

Having read the article again it suggests it's a multi-die (maybe MCM like Interlagos/Abu Dhabi or interposer/a combination of the two) which makes it more plausible IMO - it was sounding too big for a single die.

[watercooled]

This has probably been discussed before, but I've been trying to get some details on the Core M TDP. They of course claim 4.5W but I'm seeing nothing of the sort when looking at the very few sites that actually measure such things.

For example:
http://www.notebookcheck.net/Asus-Ze....136543.0.html
http://www.notebookcheck.net/HP-Envy....130223.0.html
We're seeing even up to a 20W(!!!) idle-load delta.

And with just a 'browsing' workload:
http://www.pcper.com/reviews/Process...Consumption-an
Still >8 Watts.

So how exactly is this now TDP and not still SDP which Intel is supposed to have done away with? Because from all of the measurements I can find, these are no more 4.5W TDP than the 4.5W HSW-Y parts were?

According to Intel's Core-M datasheets (yes I did try to do some digging to find out - page 62), they appear to be claiming with ambiguous language that max Turbo can draw 1.25x the TDP in bursts for PL2 state, but the off-by-default PL3 can be even more.

Looking further down the datasheet to p88, the Iccmax is listed at 18A for the core (not including uncore). At the specified 1.57v minimum voltage (so best-case) the power supply must be able to provide just under 28W to the core. That's quite a bit for a "4.5W" processor, and weirdly in the ballpark of those notebookcheck deltas!

So yeah, the 'TDP' of these CPUs seems pretty misleading. Intel do state, quietly, that the TDP is for max load at the base clocks (so around 1GHz for the CPU), but I fail to see how this is much different to sticking a 'normal' processor in a thermally-constrained chassis and having it throttle back? Perhaps the algorithms are more optimised if they're expecting throttling vs having to do it to protect the processor, but it still seems pretty much exactly what SDP was...

The point being, these just aren't '4.5W' processors, not even close. Perhaps with some customised definitions of TDP they are (in that they need a ~4.5W thermal solution to remain functional). But not in the same way that many tablet SoCs are said to be ~4W. E.g. the Nexus 9: http://www.notebookcheck.net/HTC-Goo....130446.0.html

[scaryjim]

... Perhaps with some customised definitions of TDP they are (in that they need a ~4.5W thermal solution to remain functional). ...

Erm ... TDP stands for Thermal Design Power - i.e. the minimum heat a cooling solution must be able to dissipate in order to be ratified for use with the processor.

So assuming you're right, they are exactly 4.5W TDP processors by definition. But it's a TDP based on advanced processor thermal management, rather than a TDP based on peak processor loads with some headroom for unexpected overrun - which is what we've been used to prior to the last couple of generations, where the power management has become a lot more aggressive. Don't forget AMD are talking about power management in their latest tablet chips being based on device skin temperature, allowing the processor to run very hot for brief periods as that heat won't necessarily transfer to the device skin....

[watercooled]

Thermal output = power consumption.

I could understand them exceeding a given power for *very* brief periods followed by underclocking to average out, but that's clearly not what's happening. They're far exceeding 4.5W for extended periods of time, and so obviously exceeding 4.5W of sustained thermal output.

My point is they're technically not lying with TDP, which I'm not disputing. However with that very loose definition, why not call a 4770k a 10W processor? Because if you give it 10W worth of cooling I don't doubt it will continue to function, it will just throttle heavily (apart from very old processors with no thermal throttling, many CPUs will function with no heatsink - just throttle a lot). Hence my point about the 4.5W being misleading - they're using a term differently than it is used by others, and has been for a long time.

The issue is that lots of places are churning out the sensationalistic drivel about how these '4.5W' processors are so much faster than other ~4.5W processors - they forget to mention the 'but only if they operate at significantly higher power than that -making the 4.5W comparison meaningless' part.

For instance, you could take pretty much any desktop processor, give it a 5W 'TDP' and boast about how it's so much faster than tablet processors. Just negate to mention the 'but not both at the same time' part. I'd genuinely like to see how well Core-M performs if it's actually restricted to it's claimed TDP/base clocks. That would be a fairer comparison for a start.

Edit: As I said earlier, which I think covers my point fairly well:

So how exactly is this now TDP and not still SDP which Intel is supposed to have done away with? Because from all of the measurements I can find, these are no more 4.5W TDP than the 4.5W HSW-Y parts were?

Edit2: You've actually taken my quote quite out of context - the preceding and following parts you replaced with '...' are really quite important, where I already explain what I mean.

[watercooled]

Something else I've spotted, is the smaller Broadwell-U die looks identical to the Core M die at 82mm^2 and 1.3B transistors. I'd heard rumours that Core M was produced separately on a denser SoC process, so I wonder if the U processors share this?

for Core M Intel went so far as to give Broadwell-Y its own die and design a low-power optimized version of their 14nm process just for it.

From http://www.anandtech.com/show/8355/i...ture-preview/4

Because the SoC process should be denser, but BDW-U looks identical.

[scaryjim]

Thermal output = power consumption.

...

However with that very loose definition, why not call a 4770k a 10W processor?

You see, this is the issue. I have no argument with your first statement - thermal output is at the very least proportionate to power consumption (although it's not quite as straight forward as being exactly equal). However, thermal design is not the same as peak thermal output. When we had stupid processors that always ran at full chat all the time TDP, peak thermal output and peak power draw were roughly analogous because the thermal solutions had to deal with the heat from a processor effectively always running at peak power draw. Once you can intelligently manage power and temperature in real time, you can be more intelligent with the thermal design.

It's not a loose definition to say "This processor is designed to work with 4.5W of heat dissipation", it's utterly accurate. The fact that people have got used to incorrectly assuming that TDP is the same thing as peak power draw isn't Intel's fault, or their problem. The 4770k isn't a 10W processor because it's designed to operate with a cooling system that can dissipate 84W of heat. There's no reason to assume that means the processor can't draw more than that in bursts. AMD explicitly do the same thing with Kaveri, to the point where the same processor can be configured to 95W, 65W or 45W and the peak power draw barely changes.

The problem is that, as TDP is a design characteristic, it means what the person designing the processor says it means. Intel have been explicit for more than ten years than TDP does not equal maximum potential power dissipation (reference). It may be a different definition than other people use, but that's kind of irrelevant. The TDP is the TDP. A 4.5W TDP Intel processor may draw and generate more than 4.5W of heat, but that doesn't stop it being a 4.5W TDP processor. And as long as the design characteristics that engineers use to create Intel solutions are all correct, it doesn't actually matter. If the 4.5W cooling system keeps the laptop running as intended, all well and good (and it saves the system builder money). Why specify a 28W cooling system if you don't need it? *shrug*

Each to their own, obviously, and if it's something you particularly care about then I wish you luck in finding the information, but I don't think it's "misleading" for Intel to state TDP figures based on their own published definition. Difficult to compare, perhaps, but not misleading.

EDIT to add:

This link is interesting as a recent article that discovered it's not just Intel who are cagey about exactly how they calculate TDP: http://www.theinquirer.net/inquirer/...not-comparable

I know AMD used to claim it was the maximum possible power dissipated in a processor, but I'm no longer convinced that stands with some of the work they've done with Kaveri (given 65W processors run at almost identical speeds and configurations to 95W processors, and with almost identical peak power draw).

[watercooled]

Power draw is almost exactly = heat. Unless you want to factor in minuscule amounts lost to e.g. sound and electromagnetism. Simple thermodynamics - it really is that straightforward. A device cannot release more power as heat than it 'consumes', and nor can it radiate less energy than it consumes electrically.

I'm not even complaining about peak vs average TDPs; it's fairly well-known that different manufacturers calculate them differently. E.g. AMD's GPU TDPs have traditionally been more of a never-exceed value whereas Nvidia's more recent values are more akin to SDP - the sort of power they'll use during real games. However Intel's take with the Core M is more like taking one of those GPUs, sticking a tiny heatsink on it and calling it a lower TDP because it's being forced to throttle like mad.

WRT the hypothetical 4770k - I don't care that it may exceed 84W in bursts, and I didn't even mention that TDP, I was just picking some random processor to illustrate my point. Take a 9590 instead if you wish - if you give it a 20W limit, you could likely still get it to operate, but it would probably have to throttle and gate cores to stay below that. If I stick a tiny heatsink on my 1055T forcing it to throttle below 20W, does that magically make it a 20W TDP processor? Because aside from possible throttling algorithm differences, that's essentially what the Core M is doing.

The values speak for themselves - we have a '4.5W' processor exceeding 20W of sustained power draw (and therefore ~20W heat) - IMO that really implies there's something quite wrong with that testing methodology. It's way beyond the average vs peak TDP and in to silly levels of ambiguity. The numbers aren't even in the same ball park - going from 4.5W to say 6W I'd get, but we're talking about over 4x.

A Core M can presumably operate at 4.5W if it's restricted to base clocks. A Core M can do x or y on benchmarks. Those two statements are mutually exclusive, lots of people are assuming they're not.

One of the main concerns, although it's mainly an academic one, is that because of this TDP a lot of places are chanting about how utterly fantastic the Core M is vs other processors, but they're comparing a Core M drawing 20W to processors which really are drawing ~4 Watts. There's no ambiguity or marketing wizardry there - it's a downright silly comparison.

It's not so much the claimed TDP that's irritating me. I still do stand by the fact that it's really bending the meaning of the term but assuming manufacturers know what they're getting then :shrug:. It's more the amount of places including 'professional' reviewers who are taking this TDP and jumping to conclusions without actually testing anything.

Edit: It's really those last two/three paragraphs that sum it up. As I said originally I was just looking for some details out of interest, because very few places test power draw of mobile parts (kudos to notebookcheck for that). I don't deny that it's academic though - aside from enthusiasts, TDP or its interpretation isn't generally a direct concern for consumers.

What's potentially more concerning as far as consumers are concerned, is that variability. As Anandtech showed, it's possible for the lower-end parts to outperform the highest-end ones depending on the thermal solution/processor configuration. So you can't just know what CPU to look for, you'd need to check performance of the exact model you're shopping for.

[scaryjim]

I think the point is that the TDP dictates the cooling system required, not the power draw, and that's always been the case for Intel's definition. Both Intel and AMD are running their tablet chips with peak draws well over TDP in this generation where the temperature headroom exists; presumably if a manufacturer sees fit to kit a Core M laptop with a 35W cooling system then it will perform way above expected levels most of the time.

It's been true for many generations now that performance will depend on the cooling available - Intel's early gen ultrabook chips had huge issues with GPU clocks due to thermal constraints, and both AMD and NVidia have had instances of GPUs that thermally throttled after several minutes of gaming. What we haven't seen yet for Core M are the genuinely in-depth reviews with power draw over time for a lengthy test run, but assuming Intel are aiming for the same thing AMD are (and there are plenty of slides out there from AMD stating explicitly that they're doing this), they will use thermal headroom to run the processor fast and hot to churn through a workload so they can drop to idle and recover the thermal overhead. I suspect that actually a lot of mobile processors do something similar, but when was the last time you saw an at-wall power draw figure in a mobile phone review?

Ultimately things haven't changed - its caveat emptor, and the only way to buy the right device for you is to find reviews of devices running the workload you're going to run. And as for review sites spouting hyperbole without considering every aspect of a new device - that's more an indication of the quality of the review than the quality of the technology, don't you think?

[watercooled]

Yeah I'm not disputing a lot of it is down to the review rather than the spec, like I say a lot of people are holding the TDP up as an all-important figure, when it's pretty meaningless considering all of the different interpretations of it.

I'm not saying it from an AMD vs Intel perspective really (AMD don't make equally-low TDP claims for any of their current big-core processors AFAIK, and their power draw seems quite in-line with the TDP), in fact a Silvermont comparison would be quite appropriate. It's being hailed as more performant than Silvermont in the same power envelope, but we haven't seen any proof of that yet. All we know is it's faster when it draws more power, but we kinda expected that anyway.

With aggressive turboing, I'd expect very short bursts above TDP power as I said, but with Core M we're seeing massive sustained increases. It's not a case of just using thermal capacitance to maintain an average thermal output, they're just running it at a much higher power level. Or looking at it another way, they're letting the thermal solution and thermal sensors throttle back what seems to actually be a fairly standard-power notebook processor.

Interestingly, Notebookcheck do actually measure power consumption of most of the devices they test, including mobile devices; I linked a Nexus 9 above which draws about 4W sustained - more of what you expect for a tablet processor.

[DanceswithUnix]

Interesting article here:

http://www.vrworld.com/2015/04/10/id...owth-strategy/

So, the contra revenue program has stopped. Now Intel is trying to convert the ARM system vendors into PC system vendors. Once part of the PC eco system, then Intel can perform their usual tricks on these companies.

So, with the free Atom chips gone hopefully AMD and ARM will give them the reference designs that Intel are promising and we might see some competition again. Or am I being over hopeful?

[shaithis]

So Intel missed the boat, threw away cash to catch back up and are now trying to divert the boat back onto the course it left years ago?

Hmmm