Should AMD include an unlocked multiplier on Carrizo BGA SKUs?

[cbn]

For those who don't know about Carrizo it is AMD's upcoming mobile APU SOC with two Excavator modules (ie, quad core), 512 stream processor iGPU, DDR3 memory controller (possibly DDR4 memory controller as well). It will be available up to 35 watt TDP.

The original rumor was that Carrizo would come to FM2+ desktop as a 65 watt SKU with the motherboard's southbridge chip handling I/O duties rather than the SOC. Since that original rumor we have learned that Carrizo would not come to FM2+ and instead only be available in FP4 BGA format.

BGA format, of course, limits the chip to applications where the processor is soldered onto the motherboard. Typically this is mobile, but processors like this are also seen on Mini-ITX desktop boards as well.

Historically BGA processors (at least the ones I know about) have only come with locked multipliers, but I think there may be merit in releasing at least some SKUs with a unlocked multiplier. Certainly for a Mini-ITX application I think it would make sense to allow the end user to overclock the processor to a higher level of performance. This, especially, as the chip was planned at one time to operate at 65 watts.

Another alternative, of course, would be to allow certain SKUs to have a 65 watt configurable TDP. However, I am not sure if the FP4 BGA format is validated/certified to handle 65 watts.

P.S. Regarding the actual SKUs that get the unlocked multiplier, which ones do you think it should be? Only the top 35 watt SKU with 512sp iGPU? An Athlon x 4 SKU? How about Dual core SKUs?

 

[DrMrLordX]

I voted no, and for the following reasons:

In my opinion, desktop Carrizo on a BGA board would be a cost-effective way for AMD to sell their 10-35W lineup of chips to various end users that would want such things. Obviously there would need to be some changes to the form factor of the boards so that they could fit in existing cases, unless they just plan on OEMs selling them integrated into UCFF units and/or AiO units (in which case they can slap in mostly notebook hardware and be done with it).

It will be essential for them to make this hardware available to HSA developers (what with Carrizo having HSA features unavailable on Kaveri), and MiniITX would let the devs configure their machines to their own specifications which is flexibility that they need.

That being said, I do not see unlocked multipliers being terribly useful for UCFF users or serious HSA developers.

Furthermore, there is the issue of what happens when you enable hardware IOMMU support on a Kaveri machine with a Linux kfd kernel (3.19 at least). I'm not sure of the root of that problem, but basically, you'd better be happy running base clockspeeds with the kfd device initialized. Even turbo frequencies provide 0 gain in performance under those circumstances, at least according to my testing. I have been unable to get actual turbo working with the 3.19 driver installed for various reasons, so that is a use case I have yet to test, but manually locking in 3.7 ghz (which is the default P1 state speed) provided no performance gain over 3.4 ghz with Linux 3.19 on Kaveri with hardware IOMMU enabled. So, that's another tick against overclocking when enabling HSA support on AMD APUs (something any HSA developer would want to do with a desktop Carrizo machine).

 

[cbn]

I voted no, and for the following reasons:

In my opinion, desktop Carrizo on a BGA board would be a cost-effective way for AMD to sell their 10-35W lineup of chips to various end users that would want such things. Obviously there would need to be some changes to the form factor of the boards so that they could fit in existing cases, unless they just plan on OEMs selling them integrated into UCFF units and/or AiO units (in which case they can slap in mostly notebook hardware and be done with it).

Can you provide some clarification on why an unlocked multiplier changes what you are mentioning?

P.S. These BGA chips aren't sold by AMD already integrated into a board.

Its up to the motherboard manufacturer to actually make the board around the chip.

For example, there would be nothing to prevent an OEM from making a motherboard for a AIO or UCFF using an unlocked mulitiplier chip. It simply won't be overclocked due to the form factor.

But that same unlocked chip sold to the OEM also allows the option of a Mini-ITX or Micro-ATX board built-up with extra power delivery*. This gives the unlocked chip extra desirability beyond what it would have had as a locked chip.

So having the unlocked multiplier available has basically zero downside from my point of view.

I'm guessing the main question the AMD people would likely have is "Does having unlocked multiplier BGA chips conflict or compete with the existing product line-up in some way that is undesirable?"

*Since the stock TDP is so low I'm thinking the HSF could almost be run passively in some cases. The fan would only speed up when the processor is pushed beyond 35 watts during overclocking.

 

[cbn]

Not everybody has seen a BGA board (usually there are in pre-built devices), but for anyone interested here are some DIY examples on Newegg:

http://www.newegg.com/Product/Produ...deId=1&bop=And&ActiveSearchResult=True&page=1

The most common form factor for x86 is Mini-ITX with Micro-ATX coming in second place.

 

[cbn]

Regarding the cost of the "extra cooler" needed for overclocking a low TDP unlocked chip, I don't see that being much of a problem for a DIY set-up.

In fact, it may be the only difference in some cases between running stock and overclocked is the presence of a fan. (ie, Stock= passive heatsink, OC= adds fan). So not much of a cost adder.

Even going from a 65 watt HSF to 95 watt HSF (using FM2/FM2+ as an example), copper and heatpipers are not needed. Both coolers use the same inexpensive design, just one is a little taller.

 

[Valantar]

The original rumor was that Carrizo would come to FM2+ desktop as a 65 watt SKU
[...]
This, especially, as the chip was planned at one time to operate at 65 watts.

An oft-repeated rumour does not an official plan make.

From what AMD has presented about the new core design optimizations, Carrizo shows power/performance gains against Kaveri at <15W per core pair. In other words, one is likely to see lower performance than Kaveri at significantly higher power than this. Now, given that the highest announced TDP of Carrizo is 35W, one would assume this has two 15W core pairs plus a GPU (which probably uses far more than 5W at full load, but never mind that right now). A 65W TDP would in other words double the power draw per core pair. Due to the properties of the high density design of Carrizo, this would probably perform significantly worse than a Kaveri at similar wattages. Sure, the GPU has a newer design, it's fully HSA compliant, and it's an SoC, which is cool. But overclocking a chip that (semi-purposefully) scales badly with increased power does not really make sense to me.

I'm hoping that the Kaveri refresh tacks on a few tech updates in addition to the clock speed bump. That would be far better than an unlocked Carrizo ITX board.

 

[DrMrLordX]

Can you provide some clarification on why an unlocked multiplier changes what you are mentioning?

Unlocked multipliers would be a feature nobody would use. If anything it would just fill up AMD's lineup with more skus which brings with it attendant costs. Then AMD would have to engage in price differentiation between the locked and unlocked parts . . . and really the highest stock Carrizo they've actually announced is a 35W part. I agree that the 65W part was probably determined to be undesirable and so AMD has saved themselves from money and headaches by canning that for any platform (I could be wrong; time will tell).

An unlocked 35W BGA processor would be kinda silly.

Unlocked parts also send funky messages to OEMs making FP4 boards. Do they need an overbuilt board for enthusiasts? How many unlocked units does AMD plan to sell? That sort of thing.

Don't get me wrong, I'd probably get one and overclock it, but still. What I personally want and what is best for AMD may not be identical, or even closely-related.

 

[cbn]

If anything it would just fill up AMD's lineup with more skus which brings with it attendant costs.

The number of SKUs doesn't necessarily change.

For example, instead of AMD having both locked and unlocked quad core/512sp 35 watt SKUs, the company sells only the unlocked part.

Then let the OEMs decide if they want to include overclocking options in the BIOS or not.

This costs AMD nothing.

 

[Zodiark1593]

Everything should be unlocked.

 

[DrMrLordX]

This costs AMD nothing.

No, this passes the costs on to motherboard manufacturers who then have to populate their lineup with multiple boards based on whether or not the end user is going to push the CPU out of spec (unless they think that the number of users who would do that would be nearly 0, in which case they could produce zero boards that support the feature, making its presence useless).

One of AMD's major problems with Kaveri mobile was that they either could not get their chips into notebooks or that they could not get them into GOOD notebooks. AMD has to provide top-down system design recommendations for OEMs to avoid major configuration and quality issues for Carrizo mobile systems (allegedly they are doing just that). They might not need to do the same for hypothetical AiO/UCCF systems, but we are not even sure if or when Carrizo will have any presence in that space.

So how many alternate design specs will they have to provide to OEMs to account for overclock-friendly configurations? How many OEMs will want to actually provide units based on all the possible configurations?

AMD sends the wrong message to OEMs when they provide unlocked multipliers. What AMD needs most is a good processor at a good price in stable, desirable OEM laptops that average users will want to buy. That improves their bottom line and the health of their company. It also aids in the proliferation of inexpensive HSA-compliant hardware which helps them deploy their new standard. Something Kaveri hasn't managed to do at all, mind you.

The only way unlocked Carrizo works is in a small selection of DIY boards with the chip pre-soldered in there, with different OC-friendly OEMs providing the boards (Asus, ASRock, MSI, Gigabyte, you know, the usual players). The best way to do that is to keep the unlocked parts out of OEM laptop design configurations and limit them to the DIY sector so there is no confusion as to what to do with these chips, no plurality of design recommendations, and so forth and so on. Of course carrying a different SKU costs more money, so it's only worth it if AMD thinks they'll sell enough units at a high enough price to at least cover the costs of adding another SKU to the lineup and providing reference boards for the DIY OEMs.

If it were AMD's intention to replace AM1 with this kind of product, then it might make sense, though I am not sure if the volume is there yet or if it would emerge in response to the availability of such a product. And even then, I'm still not sure that unlocked 35W BGA Carrizo is the right product for that segment which is generally about efficiency and ease-of-operation with low profile HSFs at a low price-point. Encouraging a bunch of OEMs to "overbuild" the socket/power delivery to accomodate the occasional overclock raises the price of the platform. How many AM1 users, Bay Trail users, etc. really want to overclock those chips? Mostly they just want them for SFF/UCFF HTPCs and the like.

Everything should be unlocked.

It used to be, and we saw what counterfeiters did with overclocked chips (mostly outside the US, but still). Neither AMD nor Intel will ever go back to that.

At least AMD used to have bus locks on everything (courtesy the board manufacturers) so you could run up the reference clock without pushing your SATA controller, PCIe controller, and so forth out of spec. With the APU lineup, that's really hard to manage since many/all of those controllers were integrated with the APU making it harder to manage busses running out of spec. AMD hasn't taken the time to implement their own locks either. Damned rude of them if you ask me.

 

[cbn]

No, this passes the costs on to motherboard manufacturers who then have to populate their lineup with multiple boards based on whether or not the end user is going to push the CPU out of spec

Power, even overclocked, would be so low on these processors that I have hard time believing anything beyond standard components would be used. Nothing exotic required.

Same goes with cooling.

Now, if serious/ extra ordinary overclocking components (cramming extra phases on the board with additional heatsink, etc.) were required then maybe I could see the need for two tiers of motherboards.

 

[DrMrLordX]

If we are talking boards that have no volt tweaking options or anything of the sort, then maybe so, though at that point you are looking at pretty low ceilings for those chips.

 

[cbn]

If we are talking boards that have no volt tweaking options or anything of the sort, then maybe so, though at that point you are looking at pretty low ceilings for those chips.

What would you consider a high amount of power draw?

P.S. Even the cheap Intel boards (with 3 phase power and no heatsink) we see bundled with the G3258 Pentium have voltage tweaking options.

 

[cbn]

One of AMD's major problems with Kaveri mobile was that they either could not get their chips into notebooks or that they could not get them into GOOD notebooks. AMD has to provide top-down system design recommendations for OEMs to avoid major configuration and quality issues for Carrizo mobile systems (allegedly they are doing just that). They might not need to do the same for hypothetical AiO/UCCF systems, but we are not even sure if or when Carrizo will have any presence in that space.

If AMD wants to make things easier on themselves they simply disallow use of the unlocked multiplier on laptops.

In fact, I am having a hard time believing notebook manufacturers (beyond Alienware, etc.) are even interested in overclocking. Thinking back to the Core 2 era I don't recall the major Notebook OEMs even allowing options in the BIOS for the FSB to be raised.

So how many alternate design specs will they have to provide to OEMs to account for overclock-friendly configurations?

For desktop this wouldn't be hard for AMD and the extra mileage they potentially get out of Carrizo should be worth it.

What AMD needs most is a good processor at a good price in stable, desirable OEM laptops that average users will want to buy. That improves their bottom line and the health of their company. It also aids in the proliferation of inexpensive HSA-compliant hardware which helps them deploy their new standard. Something Kaveri hasn't managed to do at all, mind you.

The unlocked multiplier doesn't change this.

The only way unlocked Carrizo works is in a small selection of DIY boards with the chip pre-soldered in there, with different OC-friendly OEMs providing the boards (Asus, ASRock, MSI, Gigabyte, you know, the usual players). The best way to do that is to keep the unlocked parts out of OEM laptop design configurations and limit them to the DIY sector so there is no confusion as to what to do with these chips, no plurality of design recommendations, and so forth and so on. Of course carrying a different SKU costs more money, so it's only worth it if AMD thinks they'll sell enough units at a high enough price to at least cover the costs of adding another SKU to the lineup and providing reference boards for the DIY OEMs.

Lets assume that some small notebook OEM actually wants to make use of the unlocked multiplier and goes against the official recommendation ( or mandate) to not use the multiplier : Is this something AMD should really be concerned with?

If it were AMD's intention to replace AM1 with this kind of product, then it might make sense, though I am not sure if the volume is there yet or if it would emerge in response to the availability of such a product. And even then, I'm still not sure that unlocked 35W BGA Carrizo is the right product for that segment which is generally about efficiency and ease-of-operation with low profile HSFs at a low price-point. Encouraging a bunch of OEMs to "overbuild" the socket/power delivery to accomodate the occasional overclock raises the price of the platform. How many AM1 users, Bay Trail users, etc. really want to overclock those chips? Mostly they just want them for SFF/UCFF HTPCs and the like.

AM1 and Bay Trail are both based on a small core design that doesn't scale well with additional voltage.

Carrizo's CPU design (although it only has 1MB of cache per module and uses HDL) is a big core built on a high drive current process that traditionally scales well with voltage. Even if HDL and the reduced cache restricts performance I have to believe these cores will go farther than atom and Jaguar/Puma.

P.S. Regarding small low profile heatsinks on Mini-ITX boards. A 15 watt unlocked Carrizo works with that. The heatsink can be the typically sized passive one at stock speeds (but with an added fan header on the board and provisions for mounting an optional case fan on top the heatsink for overclocking). Since overclocking a 15 watt processor to 65+ watts requires nothing beyond standard components price should not increase. One board design to handle both 15 watt and 35 watt unlocked SKUs, the only difference being the 15 watt board comes with a passive heatsink and the 35 watt one adds a fan.

 

[DrMrLordX]

What would you consider a high amount of power draw?

P.S. Even the cheap Intel boards (with 3 phase power and no heatsink) we see bundled with the G3258 Pentium have voltage tweaking options.

The G3258 isn't a mobile processor. You could argue that it could be, perhaps in a semi-DTR laptop, but I digress. I would consider anything that pushes Carrizo up to or beyond its optimal point on the voltage/clockspeed scaling curve (50W or so) to be high power draw for that chip. The cooling solutions that are going to be available in laptops carrying the 35W top-end part are probably not going to do a great job of handling any extra volts. Bear in mind that laptops are where we're going to see Carrizo first.

If AMD wants to make things easier on themselves they simply disallow use of the unlocked multiplier on laptops.

They could, but then you'd have to wonder why they ever unlocked any of the Carrizo chips to begin with. They will be going only into laptops at launch, and it may be awhile before we see them anywhere else.

In fact, I am having a hard time believing notebook manufacturers (beyond Alienware, etc.) are even interested in overclocking. Thinking back to the Core 2 era I don't recall the major Notebook OEMs even allowing options in the BIOS for the FSB to be raised.

They aren't. Hence the absurdity of unlocked Carrizo.

For desktop this wouldn't be hard for AMD and the extra mileage they potentially get out of Carrizo should be worth it.

I don't know what kind of mileage you are expecting them to get from offering up an unlocked part to the desktop crowd. It might be good for some PR, maybe, if people liked the results. For the most part, I expect that the chips will be a disappointment. Even the best, most-hardened board out there probably wouldn't offer top clockspeeds in excess of 4-4.2 ghz with pretty significant voltage increases. In fact, I think we'd be lucky to see even that.

Lets assume that some small notebook OEM actually wants to make use of the unlocked multiplier and goes against the official recommendation ( or mandate) to not use the multiplier : Is this something AMD should really be concerned with?

Maybe. Depends on the failure rate of the notebooks from people overclocking them. It would be an amusing feature for gamerish notebooks but it would be a cooling nightmare for the OEM unless they just didn't give a darn.

AM1 and Bay Trail are both based on a small core design that doesn't scale well with additional voltage.

How they scale with voltage isn't really the point. The point is that 10-35W Carrizo pre-soldered onto some vendor's BGA board sold as DIY kit is a direct competitor to Bay Trail and AM1 Kabini in the SFF/UCFF segment. Of course a 35W Carrizo will hand either Kabini or Bay Trail their respective heads to them on platters, even without any overclocking. The lower-power parts would probably fare pretty well too. Point is you don't need or want an unlocked processor in that segment either, when you're got something that would be top-dog performance-wise in that segment.

P.S. Regarding small low profile heatsinks on Mini-ITX boards. A 15 watt unlocked Carrizo works with that. The heatsink can be the typically sized passive one at stock speeds (but with an added fan header on the board and provisions for mounting an optional case fan on top the heatsink for overclocking). Since overclocking a 15 watt processor to 65+ watts requires nothing beyond standard components price should not increase. One board design to handle both 15 watt and 35 watt unlocked SKUs, the only difference being the 15 watt board comes with a passive heatsink and the 35 watt one adds a fan.

I'm not convinced that overclocking in tight enclosures would be a great idea, especially when you've got a chip that has voltage scaling issues past ~50W or so. Heat will build up quickly unless it's a well-designed enclosure that can deal with that problem. Also, are you certain that FP4 will be able to push that kind of wattage without some reinforcement here or there? We haven't even seen Carrizo in retail, much less a reference board for hypothetical DIY units.

 

[cbn]

I don't know what kind of mileage you are expecting them to get from offering up an unlocked part to the desktop crowd. It might be good for some PR, maybe, if people liked the results. For the most part, I expect that the chips will be a disappointment. Even the best, most-hardened board out there probably wouldn't offer top clockspeeds in excess of 4-4.2 ghz with pretty significant voltage increases. In fact, I think we'd be lucky to see even that.

I would consider 4 to 4.2 Ghz to be awesome.

Regardless, if Carrizo is deployed on a desktop board I think the VRMs should be stout enough to keep the CPU from dropping below base clocks while the iGPU is active for gaming. And if Carrizo is anything like Kaveri this will mean the motherboard needs to support a TDP of over 35 watts anyway.

So to sums things up, I see at least two reasons for BGA boards to have power and cooling beyond 35 watts:

1. To allow unlocked multiplier overclocking

2. To maintain CPU clocks at the advertised levels while iGPU is active during gaming. (As we have seen with Kaveri testing the advertised CPU clocks are only good when the iGPU is not under load)

 

[cbn]

especially when you've got a chip that has voltage scaling issues past ~50W or so.

We don't know how Carrizo responds to voltage.

It may be Carrizo was not packaged as LGA for reasons that have nothing to do with voltage scaling to 65 watts.

Like you mentioned earlier in the thread adding SKUs costs money. And it might have been AMD saw keeping Carrizo strictly BGA as the more cost effective route to take. (ie, BGA does not preclude OEMs from making Carrizo into a desktop, but having too much LGA inventory would prevent notebook manufacture)

 

[DrMrLordX]

I would consider 4 to 4.2 Ghz to be awesome.

I wouldn't. If that's the best Carrizo production silicon could do, then it would be a retreat from the ~4.5 ghz Kaveris you see running on so many FM2+ machines already. Bear in mind, I am talking about Carrizo at 1.5v+, which is where some Kaveri chips have to go just to get to 4.5 ghz.

Regardless, if Carrizo is deployed on a desktop board I think the VRMs should be stout enough to keep the CPU from dropping below base clocks while the iGPU is active for gaming. And if Carrizo is anything like Kaveri this will mean the motherboard needs to support a TDP of over 35 watts anyway.

Nobody has ever really been able to indict VRMs as the culprits for CPU throttling during GPU usage. That behavior is linked to something in the microcode that interacts with Windows ACPI drivers. You will not see that behavior under Linux. Whether or not AMD chooses to throttle the CPU on Carrizo during iGPU-heavy workloads remains to be seen.

We don't know how Carrizo responds to voltage.

We have seen AMD's official slides showing Excavator vs Kaveri scaling of clockspeed with per-module power consumption. Their own slides show Excavator falling off at around 50W. Surely their own slides show their latest product in the best possible light.

 

[ShintaiDK]

Carrizo shipping samples:
https://www.zauba.com/import-carrizo-hs-code.html

DDR3 and 1.5Ghz(non turbo?) APU so far.

 

[exar333]

No, this passes the costs on to motherboard manufacturers who then have to populate their lineup with multiple boards based on whether or not the end user is going to push the CPU out of spec (unless they think that the number of users who would do that would be nearly 0, in which case they could produce zero boards that support the feature, making its presence useless).

One of AMD's major problems with Kaveri mobile was that they either could not get their chips into notebooks or that they could not get them into GOOD notebooks. AMD has to provide top-down system design recommendations for OEMs to avoid major configuration and quality issues for Carrizo mobile systems (allegedly they are doing just that). They might not need to do the same for hypothetical AiO/UCCF systems, but we are not even sure if or when Carrizo will have any presence in that space.

So how many alternate design specs will they have to provide to OEMs to account for overclock-friendly configurations? How many OEMs will want to actually provide units based on all the possible configurations?

AMD sends the wrong message to OEMs when they provide unlocked multipliers. What AMD needs most is a good processor at a good price in stable, desirable OEM laptops that average users will want to buy. That improves their bottom line and the health of their company. It also aids in the proliferation of inexpensive HSA-compliant hardware which helps them deploy their new standard. Something Kaveri hasn't managed to do at all, mind you.

The only way unlocked Carrizo works is in a small selection of DIY boards with the chip pre-soldered in there, with different OC-friendly OEMs providing the boards (Asus, ASRock, MSI, Gigabyte, you know, the usual players). The best way to do that is to keep the unlocked parts out of OEM laptop design configurations and limit them to the DIY sector so there is no confusion as to what to do with these chips, no plurality of design recommendations, and so forth and so on. Of course carrying a different SKU costs more money, so it's only worth it if AMD thinks they'll sell enough units at a high enough price to at least cover the costs of adding another SKU to the lineup and providing reference boards for the DIY OEMs.

If it were AMD's intention to replace AM1 with this kind of product, then it might make sense, though I am not sure if the volume is there yet or if it would emerge in response to the availability of such a product. And even then, I'm still not sure that unlocked 35W BGA Carrizo is the right product for that segment which is generally about efficiency and ease-of-operation with low profile HSFs at a low price-point. Encouraging a bunch of OEMs to "overbuild" the socket/power delivery to accomodate the occasional overclock raises the price of the platform. How many AM1 users, Bay Trail users, etc. really want to overclock those chips? Mostly they just want them for SFF/UCFF HTPCs and the like.



It used to be, and we saw what counterfeiters did with overclocked chips (mostly outside the US, but still). Neither AMD nor Intel will ever go back to that.

At least AMD used to have bus locks on everything (courtesy the board manufacturers) so you could run up the reference clock without pushing your SATA controller, PCIe controller, and so forth out of spec. With the APU lineup, that's really hard to manage since many/all of those controllers were integrated with the APU making it harder to manage busses running out of spec. AMD hasn't taken the time to implement their own locks either. Damned rude of them if you ask me.

Good post.

AMD should lock-down the multiplier and engineer different CPU options based on thermals and power efficiency and work with OEMs to ensure the entire package is put together in the 'right way' for different applications.

Ensuring the right CPU in the right product is important for both market perception as well as performance. Intel does a very good at this and AMD should make this a focus, especially as we move away from older, thicker 'standard laptops' into more 'creative' designs and increasingly-thinner and lighter designs.

 

[frozentundra123456]

I did not vote, but I think it is not a very good candidate for an unlocked multiplier. Seems the best application is low power mobile, and apparently doesnt scale very well to high frequencies in any case.

 

[Valantar]

We don't know how Carrizo responds to voltage.

Isn't this slide - directly from AMDs own Carrizo/Excavator presentation - quite clear on this?

In other words: AMD has said that Excavator would, at high power, perform worse than Kaveri at the same power draw. Now why on earth are we still discussing this?

 

[cbn]

Isn't this slide - directly from AMDs own Carrizo/Excavator presentation - quite clear on this?

In other words: AMD has said that Excavator would, at high power, perform worse than Kaveri at the same power draw. Now why on earth are we still discussing this?

That scaling is still definitely worth it for an unlocked multiplier.

And we have to remember current FM2+ Mini-ITX motherboards are relatively expensive.

So an unlocked Carrizo BGA Mini-ITX definitely has a niche for lower cost SFF.

 

[cbn]

Nobody has ever really been able to indict VRMs as the culprits for CPU throttling during GPU usage. That behavior is linked to something in the microcode that interacts with Windows ACPI drivers. You will not see that behavior under Linux. Whether or not AMD chooses to throttle the CPU on Carrizo during iGPU-heavy workloads remains to be seen.

The explanation I have seen is that the CPU downclocking under iGPU load is to keep the chip within TDP. This makes sense to me considering Athlon x 4 860K has the same TDP as the A10-7850K even though the iGPU is gone.

Ideally for these BGA Mini-ITX boards we would see power circuitry capable of supporting at least 50 watts for the CPU plus whatever power the iGPU uses (ideally the iGPU could also be overclocked as well....particularly if DDR4 is included).

So for FP4 Mini-ITX boards we would need the socket to supply at least 65 watts in overclocked form to get the most out of the silicon. Unless there is something about the size of the BGA connection that limits this I don't see supplying a simple 3 phase VRM without heatsink as something that would increase cost over having a 35 watt set-up.

 

[Valantar]

That scaling is still definitely worth it for an unlocked multiplier.

And we have to remember current FM2+ Mini-ITX motherboards are relatively expensive.

So an unlocked Carrizo BGA Mini-ITX definitely has a niche for lower cost SFF.

So what you're really asking for is cheaper ITX FM2+ motherboards - which is something that might actually be made if people asked manufacturers for them (due to broader compatibility and significantly lower production costs than making a new CPU SKU (not to mention they would be made by companies not currently desperately trying to save themselves)). I agree on that - even if they automatically locked CPUs to their 45W/65W states, that would be worth it in many cases if the motherboard was significantly cheaper.

But you're still saying the scaling is worth it? Seriously?

If my eyes aren't completely off, Excavator sits at ~0.95 at 15W in the chart, and ~1.05 at 25W. The curve even seems to droop at the end. In other words, frequencies increase by about 10%, while power draw increases by 66%. How is that anything other than horrendous scaling?

Meanwhile, we know Kaveri can do 3.1GHz base/3.3 Turbo @ 45W (A8-7600), 3.3 base/3.8 Turbo @ 65W (A8-7600) and 3.7 base/4 turbo @ 95W (a10-7850K (although the A10-7700 would be a fairer comparison to the previous two, having the same GPU, it also clocks significantly lower)). Discounting the iGPU here, and counting TDP as actual power consumption (which is of course not nearly true), the GHz/W calculation for base clocks goes from 0.07 to 0.05 to 0.039. Kaveri's efficiency in other words drops off significantly when the core pairs exceed ~20-30 watts. When Carrizo is supposed to scale worse than this, I shudder to think of the amounts of power one would need to push AMDs "normalized frequency" in the chart past 1.1 - especially as AMDs own curated press slide graph seems to level off at ~25W.