AMD Carrizo Pre-release thread

[Abwx]

According to thoses curves the cores efficency at lower power is improved by almost 50%, Atenra was fully right, the first graph didnt include the improvements provided by the AVFS.

 

[Fjodor2001]

Since the improvement is most impressive in the lower TDP range, could they make a Broadwell Y Core M competitor based on Excavator, at around 4.5 W TDP? Perhaps intended for cheap fanless Ultrabooks...

 

[Enigmoid]

IPC improvement is somewhat meager, efficency at low power seems good but AMD stated that it s not designed to work in a 5-6W range, wich seems to be contradicted by the curve, overall that lookss to be exclusively a Core M competitor.

This CPU would suit perfectly an AM1 like plateform, too bad that AMD seems to be less concerned by long term upgradability, my AM1 is begging for this...

5-6W for a single core module not including PCH, igp, memory.

 

[geoxile]

No announcements of design wins?

 

[el etro]

After reading the second wave of slides i wonder: Why not launch(This Year!) a Desktop Carrizo with up to 65W TDP targets?

 

[jpiniero]

After reading the second wave of slides i wonder: Why not launch(This Year!) a Desktop Carrizo with up to 65W TDP targets?

Budget cuts.

 

[Abwx]

5-6W for a single core module not including PCH, igp, memory.

AMD clearly stated that Carrizo is not designed for a 5-6W TDP range, as for the components you re quoting the memory is irrelevant as this is not integrated in the SoC.

The PCH is likely the one of Beema with some improvements, this latter power comsumption is at most 8% of Beema s TDP wich point to about 1W.

Rest of the uncore should take 25% of the TDP if i was to recycle a Kaveri DT number but we can be sure that this part was also improved substancialy, at the end what matters is the 45-50% better perf/Watt than a Steamroller as there s power saving even within the GPU so the CPU will get a bigger part of the power budget.

With all the datas available i think that one can make a good estimation of how this chip will perform.

A last info, Kaveri has quite higher leakage than Beema, so just adopting this latter process and design guidelines will save power by itself even further, you can check at Notebookcheck the difference between a Kaveri and a Beema at idle.

Since the improvement is most impressive in the lower TDP range, could they make a Broadwell Y Core M competitor based on Excavator, at around 4.5 W TDP? Perhaps intended for cheap fanless Ultrabooks...

AMD has publicly stated that it doesnt suit well a 5-6W TDP (their numbers) so we wont see such lowly clocked parts, that said the 4.5W Core M are not real 4.5W otherwise they would perform extremely poorly, it s all power management to squeeze tha last drop of available thermal headroom.

 

[III-V]

Since the improvement is most impressive in the lower TDP range, could they make a Broadwell Y Core M competitor based on Excavator, at around 4.5 W TDP? Perhaps intended for cheap fanless Ultrabooks...

Not on bulk 28nm...

 

[Abwx]

After reading the second wave of slides i wonder: Why not launch(This Year!) a Desktop Carrizo with up to 65W TDP targets?

Because the improvement over a 65W Kaveri would be something like 10%, that s negligible and well within plateforms variations from a manufacturer to another, the only ilmprovements would be features like H265 or eventualy AVX2 for thoses who are content with perhaps 10-15% gains in very limited scenarios like some encoding.

At 45W it would make sense though but AMD has stated that it will be available on BGA only and at 35W at most, as such this chip will compete against the intel BGAs BDW used in NUCs and other SFFs.

 

[DrMrLordX]

AtenRa, after looking at the AVFS slide, it looks more to me like Carrizo (Excavator + AVFS) will hit the percentage of normalized clockspeed at 15W per module of a Kaveri at ~18-19W (just eyeballing the graph there). Given a projected +5% IPC for Carrizo, we can expect a 30W Carrizo quad to perform equally as well as a hypothetical 40W Kaveri.

Of course, if you go further down the voltage scale and start looking at 7.5W per module on Carrizo, yes, things start looking rosier for Carrizo. Seems to me that 7.5W per module on Carrizo is about on par with Kaveri at ~12.5-13W per module.

Regardless, +5% IPC is still +5% IPC. Carrizo is nice and efficient at the lower end of the power spectrum. At higher voltages, however . . .

After reading the second wave of slides i wonder: Why not launch(This Year!) a Desktop Carrizo with up to 65W TDP targets?

Look at the slides. Even on the AVFS slide, Excavator (HDL + AVFS) provides inferior clockspeed scaling with power beyond maybe ~24W per module as compared to Kaveri. That's slightly better than the non-AVFS slide for Excavator. What you are looking at is this: Kaveri and Excavator (HDL + AVFS) can achieve about the same clockspeed at 50W given a 2M configuration. Just for comparison, consider Kaveri @ 65W TDP: that's the A8-7600 with base clock of 3.3 ghz. Carrizo should have lower base clock at the same TDP thanks to its inferior scaling with power. 65W Carrizo wouldn't be all that and a bag of chips, and it would probably stink for overclocking. On the plus side, the die is smaller, so it would be cheaper to produce, so they could theoretically sell it for less.

In the 10-35W power envelope, it should be quite nice. And cheap.

Because the improvement over a 65W Kaveri would be something like 10%,

Are you sure it would even be that good? I'm guessing the base clock would have to be lower than the 7600's 3.3 ghz.

 

[III-V]

They really shrunk the size of the module. The CPU area used to be a fair bit taller (y -axis) than the L2 cache, and now it's a fair bit shorter.

 

[Abwx]

Are you sure it would even be that good? I'm guessing the base clock would have to be lower than the 7600's 3.3 ghz.

You can already get 10% better perf at 65W with a current Kaveri, all you have to do is to set the voltage at a normal stability margin rather than the huge voltage that AMD implemented as default value, here Hardware.fr review of a mediocre sample they deliberately selected when it was launched, first line is the stock setting and the power is measured under Prime 95 :

At stock settings CPU comsumption is 74.4 x 0.9 = 67W
User setting and stock frequency yield 55.2 x 0.9 = 50W

AtenRa, after looking at the AVFS slide, it looks more to me like Carrizo (Excavator + AVFS) will hit the percentage of normalized clockspeed at 15W per module of a Kaveri at ~18-19W (just eyeballing the graph there). Given a projected +5% IPC for Carrizo, we can expect a 30W Carrizo quad to perform equally as well as a hypothetical 40W Kaveri.

Carrizo should have lower base clock at the same TDP thanks to its inferior scaling with power.

Are you sure that you are not missing something, and even a huge something...??.

 

[geoxile]

What happened to integrated voltage regulation?

Also, does AMD support S0i1?

 

[podspi]

Whether this will beat what Intel has is ... questionable. But the power improvements + die size improvements suggest to me that if they price this right they might have a winner on their hands.

Perf/watt has always been the thing really killing AMD. FX-8 series aren't that bad (in terms of the performance they are capable of and what many people use every day) - but they are too big to produce cheaply and use way, way too much power.

 

[DrMrLordX]

At stock settings CPU comsumption is 74.4 x 0.9 = 67W
User setting and stock frequency yield 55.2 x 0.9 = 50W

Yes, and you can also reliably tune any number of FX CPUs (especially the newer ones) to run at higher clockspeeds @ stock vcore or lower vcore @ stock clockspeed. Sometimes the silicon lottery just runs that way. That is largely irrelevant to the bigger picture. All you have done here is show that AMD has a habit of overvolting some of their chips @ stock (that, or it's the mobo makers' fault). It has no bearing on the clockspeed/power curve that AMD has published in their slides, nor does it change the relationship revealed therein between Steamroller and Excavator (HDL + AVFS).

Fact is, AMD's own slides clearly show Excavator (HDL + AVFS) suffering poor clockspeed scaling with voltage/power beyond somehere around 24W per module compared to Kaveri.

Are you sure that you are not missing something, and even a huge something...??.

Not based on the provided information, no.

Whether this will beat what Intel has is ... questionable. But the power improvements + die size improvements suggest to me that if they price this right they might have a winner on their hands.

Perf/watt has always been the thing really killing AMD. FX-8 series aren't that bad (in terms of the performance they are capable of and what many people use every day) - but they are too big to produce cheaply and use way, way too much power.

The efficiency of Carrizo in the stated power envelope of 10W-35W is really quite impressive if AMD's materials are to be believed. It is certainly a great leap in efficiency vs. Kaveri/Steamroller. I'm sure they'll find continued use for cat cores (Carrizo-L is Puma, after all, and there's Nolan which is supposed to be Puma+), but Carrizo is definitely encroaching on cat territory, especially with the reduction in die size.

Be honest, which would you want at 10W TDP? Carrizo or Carrizo-L? Methinks L wouldn't be that much cheaper, and it would probably be slower.

 

[AtenRa]

I will tell you this,

If Carrizo at 15W TDP reach 15W TDP Haswell CPU performance but with faster than Broadwell iGPU performance i will call it the best product of 2015.

Also, i would really really like 35-45 or even 55W TDP Desktop FM2+ SKUs.

 

[Erenhardt]

wonder about igp performance with dual channel memory and better bandwidth efficiency.
The sweat spot for CPU is around 7W for module. Whole 2M4T CPU should have <20W TDP with possibility of turbo claiming some of un-used IGP TDP. Looking at the graph, its equivalent in freqency to 11W SR module.

If the whole chip is rated at 45W TDP then we have about 25W for IGP, which is way more than any FM1 SOC, and probably on par with bandwidth limited 65W kaveri.

With the IGP improvements in mind (better bandwidth utilization and general efficiency), carrizo igp should be better than that of a8-7600, or even any FM2+ APU.

It is clear this chip is targeted at mobile 35W sku. Small, cheap but potent chip to battle against intel CPU + nvidia low end GPU.

 

[Shehriazad]

I really hope that chip comes to AM1 at some point. They would really be a great option there to build some cheapskate PCs. I mean sure...they won't ever be able to stand on equal footing with Kaveri (performance wise) except on the 45W parts maybe, and I accept that..but AM1? That platform could really use some love. Perfect for mini-mini PCs.

 

[DrMrLordX]

AM1 would almost be a better target for Carrizo. Both Carrizo and Kabini have the FCH on-die. I fear that existing boards might not do well with dual-channel memory, but I could be wrong. Maybe with a BIOS update they'd be okay.

It's more likely we'd see Carrizo-L on AM1 and Carrizo in FP4 BGA only, but hey, you never know.

 

[Abwx]

Not based on the provided information, no.

.

According to the slide in post 801 frequencies at about 5W per module are 0.55 for Steamroller and 0.7 for Excavator, we dont know the absolute values but the ratio between the two cores is 0.7/0.55 = 1.27.

This say that at 5W/module Excavator work at 25-30% higher frequency than Steamroller, or that at 5W/module it works at the same frequency than a Steamroller at 8W.

You have indeed made the same analysis, but my question was about this curious statement of yours :

Carrizo should have lower base clock at the same TDP thanks to its inferior scaling with power.

I see that you have edited your post to remove this sentence....

 

[krumme]

Better efficiency and a cheaper platform, seems like a sensible design given AMD's limited options.

In amd entire 30 years life or so this is the only sensible choice from a market perspective and segmentation point of view - except the first bobcat and the consoles. Normally amd would take the braindead tractor method and only if their tractor was strongest by far and by chance they could win. That method didnt earn them any money.

Add this product even comes from a failed arch designed for freq and bandwith. It shows using your brain to think business into things and having a sharp focus brings you a long way. They might actually sell this product.

One can then have a slight hope they bring this business sense - and interesting power and hdl tech, to zen and 14nm next year. But it goes to show the business sense is far more important.

 

[NTMBK]

AM1 would almost be a better target for Carrizo. Both Carrizo and Kabini have the FCH on-die. I fear that existing boards might not do well with dual-channel memory, but I could be wrong. Maybe with a BIOS update they'd be okay.

It's more likely we'd see Carrizo-L on AM1 and Carrizo in FP4 BGA only, but hey, you never know.

Can't happen. AM1 doesn't have the pins for dual channel memory (part of the reason it's such a small socket).

 

[Abwx]

There s 104 pins out of 240 in a memory module that are not connected to the SoC, this let 136 pins per module that convey the signals, so a dual channel could have been possible in an AM1 should AMD have decided so, moreover given the low TDP that require less pins for supply the voltage.

 

[Enigmoid]

AMD clearly stated that Carrizo is not designed for a 5-6W TDP range, as for the components you re quoting the memory is irrelevant as this is not integrated in the SoC.

The PCH is likely the one of Beema with some improvements, this latter power comsumption is at most 8% of Beema s TDP wich point to about 1W.

Rest of the uncore should take 25% of the TDP if i was to recycle a Kaveri DT number but we can be sure that this part was also improved substancialy, at the end what matters is the 45-50% better perf/Watt than a Steamroller as there s power saving even within the GPU so the CPU will get a bigger part of the power budget.

I meant the memory controller.

Beema's PCH is incredibly neutered compared to the desktop stack.

I highly doubt we will see a 45-50% perf/W improvement. Even AMD's own slides do not support this and AMD's slides have exaggerated things in the past.

 

[Abwx]

I meant the memory controller.

When i write the rest of the uncore this mean that the memory controler is included in this uncore.

Beema's PCH is incredibly neutered compared to the desktop stack.

I wonder where you are digging to find such arguments while it s obvious
that it has nothing to do with an hypothetic neutralisation, simply DT parts use at best 55nm processes while Beema s PCH is obviously 28nm, even the cut down version used with mobile Kaveris is using 55nm.

I highly doubt we will see a 45-50% perf/W improvement. Even AMD's own slides do not support this and AMD's slides have exaggerated things in the past.

That s a blank statement, could you please point us where the slides say so.?..