AMD Carrizo Pre-release thread

[SAAA]

ISSCC Tips Hot Circuit Designs

No surprise, 28nm is the node for Excavator. A bit surprising are other numbers tho, compare with SR:

...

AMD Kaveri GF 28nm SHP 2.41B 245 mm2 9.837M
AMD Richland GF 32nm SOI 1.30B 246 mm2 5.285M
AMD Llano GF 32nm SOI 1.178B 228 mm2 5.166M

Looks like the HDL achieved the density increase (28% more transistors with basically same die size). Dunno why AMD cites the lower die area in the paper when Kaveri is right about the same size as Carrizo. Also I doubt they went for SP count increase due to mem. BW woes they have and will have (DDR3 dual ch.). What they spent those additional transistors on is unknown for now.

I still wonder what did they add in Kaveri to use 1.1B more transistors and here they are with +0.7B, again.
Hopefully this is more SPs or better ones, say with reduced clocks to improve even more performance/W...

The density is really high now but I won't discuss it more than for comparisons against Kaveri itself.
I mean there must be a reason this and other SoCs (say the A8X with 3B too) have so many transistors compared to the node they are built on.
You can read here and there how pitiful the density are for Intel but look at IBM 22nm chips, Power8 at >5B and >650m^2 gives about the same density as Xeon parts, and tell me who's lying or how/what are they really counting because really I dont know.

 

[Shehriazad]

+0.7B sticking to the same die size in 28NM??? Wasn't Kaveri already hella cramped for a 28NM chip?

Am I tripping for thinking that this is hella weird? Especially if they want to fit HBM on there. (Which is still a possibility).

Also as the noob that I am I have to ask...what's HDL? xD

 

[maarten12100]

+0.7B sticking to the same die size in 28NM??? Wasn't Kaveri already hella cramped for a 28NM chip?

Am I tripping for thinking that this is hella weird? Especially if they want to fit HBM on there. (Which is still a possibility).

Also as the noob that I am I have to ask...what's HDL? xD

Yes the density is rather high. I actually have a Kaveri chip and it is bottlenecked by the low ram speeds. The only solution is to decrease resolution however that will just bottleneck you by cpu speed. So you will lose either way.

If Carrizo can fix the bandwidth problems and increase the cpu perf by 20% over steamroller all while cutting the power in half then I will be buying it.

I doubt they will add many more transistors I actually think they will rework the design to be far more efficient. As for HBM I don't think it will be ready for Carrizo especially not the mobile launch supposedly in December. Though the mobile Carrizo is the one that needs it the most.

HDL stands for High Density Libraries meaning that a node is composed more dense than without those libraries. It gives a major reduction in die size but increases the leakage of power. There is also UHDL wich stand for Ultra High Density Libraries. Even denser than the first :thumbsup:

 

[krumme]

What do you guys mean by where did the 0.7B transistors go? - i mean isnt this the beefed up core with eg. separate fpu for each core? (meaning the module approach is practically dead - there is hardly anything left shared)

I look at it oposite - if its only 3.1B there is no significant improvements in numbers of sp. That again is bad news because it means AMD aparently thinks this apu is only meant for low end. But yes its the bd core, we will have to wait a year or two before we get the new interesting stuff - but still damn. We could do with some competition, because Intel is just ripping us big time.

 

[Shehriazad]

*snip*

Ugh...so doesn't HDL in the end mean that the frequency will forcibly take a hit again and thus ruin any possible IPC gains yet again? (Since Kaveri already did the same.... higher IPC...but ruined by the fact that it clocks 400 MHZ lower than the previous generation).

Sheesh....I really hope AMD doesn't mess this up. I really expected Carrizo to be clocked at 4 GHZ base again. Not a fan of the thought of looking forward to a 3.3 GHZ chip even if the IPC is raised to "make up for it".

I mean fine...I won't see a Hexacore FM2+ CPU...but at least don't destroy my dreams of actually seeing a Carrizo at 4 GHZ base core clock


In a perfect world Carrizo would be made in 20NM, Have 2GB of HBM, be clocked at 4.1 GHZ base with a 10% IPC gain over Kaveri (and thus finally destroying Richland hard on the CPU side) and have its' GPU cores with a full feature set of RX 300 with the ability of being crossfires with a R7 350

Sadly I fear that another Kaveri is going to happen instead...and I hope that I'm wrong at least a little bit. (Don't get me wrong...Kaveri wasn't terrible...but they cut corners on all the wrong ends in my eyes)

 

[NTMBK]

Ugh...so doesn't HDL in the end mean that the frequency will forcibly take a hit again and thus ruin any possible IPC gains yet again? (Since Kaveri already did the same.... higher IPC...but ruined by the fact that it clocks 400 MHZ lower than the previous generation).

Sheesh....I really hope AMD doesn't mess this up. I really expected Carrizo to be clocked at 4 GHZ base again. Not a fan of the thought of looking forward to a 3.3 GHZ chip even if the IPC is raised to "make up for it".

I mean fine...I won't see a Hexacore FM2+ CPU...but at least don't destroy my dreams of actually seeing a Carrizo at 4 GHZ base core clock


In a perfect world Carrizo would be made in 20NM, Have 2GB of HBM, be clocked at 4.1 GHZ base with a 10% IPC gain over Kaveri (and thus finally destroying Richland hard on the CPU side) and have its' GPU cores with a full feature set of RX 300 with the ability of being crossfires with a R7 350

Sadly I fear that another Kaveri is going to happen instead...and I hope that I'm wrong at least a little bit. (Don't get me wrong...Kaveri wasn't terrible...but they cut corners on all the wrong ends in my eyes)

Carrizo is laptop oriented, from the sounds of it. Integrated FCH which can't be used on the FM2+ motherboards, and reduced max clock speed in favour of improved power consumption.

 

[Shehriazad]

Carrizo is laptop oriented, from the sounds of it. Integrated FCH which can't be used on the FM2+ motherboards, and reduced max clock speed in favour of improved power consumption.

Well but the laptop sales don't seem all that great to suddenly focus on that...or am I totally wrong here? I know that the mobile version is supposed to come out before the Desktop version...but that's about it.

The APU market just seemingly started to get interesting...is AMD really going to take a massive dump on this by making Carrizo weaker than Kaveri on Desktop?

I can only hope not. If they really end up released the 8850K (or whatever the name will be) in march/april 2015 and it's just Kaveri with a lower tdp but the same or even worse performance....

Then my tiny trust will have been hurt immensely. If they really put out some 3.3 GHZ Carrizo APU then I'm definitely going to have to sell my setup and switch to Intel...which I would kind of hate...but the performance gap is only getting bigger...and at some point I NEED an UPGRADE.

Oh well....I can still only wait and see...and pray to the holy semiconductor in the sky that AMD doesn't mess this up.

 

[Gideon]

is AMD really going to take a massive dump on this by making Carrizo weaker than Kaveri on Desktop?

Better prepare yourself to be disapointed then. I'm pretty sure AMD stated, that the max TDP for Carrizo will be 65W on desktop AND that they use High Density Libraries. While these give a lot of advantages elsewhere they will limit max clock speed. This makes 4.0 Ghz Carrizo extremely unlikely.

 

[frozentundra123456]

Well but the laptop sales don't seem all that great to suddenly focus on that...or am I totally wrong here? I know that the mobile version is supposed to come out before the Desktop version...but that's about it.

The APU market just seemingly started to get interesting...is AMD really going to take a massive dump on this by making Carrizo weaker than Kaveri on Desktop?

I can only hope not. If they really end up released the 8850K (or whatever the name will be) in march/april 2015 and it's just Kaveri with a lower tdp but the same or even worse performance....

Then my tiny trust will have been hurt immensely. If they really put out some 3.3 GHZ Carrizo APU then I'm definitely going to have to sell my setup and switch to Intel...which I would kind of hate...but the performance gap is only getting bigger...and at some point I NEED an UPGRADE.

Oh well....I can still only wait and see...and pray to the holy semiconductor in the sky that AMD doesn't mess this up.

I have to disagree. Laptops/mobile is the only place that APUs make sense to me, especially with the limited performance available today because of bandwidth restrictions. Unless AMD makes huge strides with Carizzo, you are compromising both cpu and gpu performance relative to Intel plus discrete. In a laptop, where TDP, space, and thermal budget are limited, those compromises are much more acceptable than in a desktop where you can get 50% more performance for very little additional cost. I think AMD is definitely correct to focus on laptops with this chip. Whether they can execute effectively is another story. They have been given a bit of breathing room with all the delays of 14nm from Intel, so they might be able to carve out a nice segment of the market.

 

[Shehriazad]

*snip*

As long as the IPC makes up for it so much that it still manages to get ahead of Kaveri.

But with lower TPD AND HDL....I am now officially scared that Carrizo is actually a step back.

They HAVE to do better than the Kaveri 7800.

This one is already a 65W APU with a 3.5 GHZ clock.

If they can't beat their own product of the last generation then 2015 is the darkest year for AMDs CPU section yet.

Possible outcomes I see right now:

-Carrizo 65W at 3.5GHZ but as fast as Kaveri 4 GHZ....acceptable
-Carrizo at sub 3.5 GHZ and not even as fast as Kaveri 4 GHZ.
-Carrizo at nearly 4 GHZ and faster than Richland.

The GPU core can only have 2 things happen to it with a positive outcome.

-HBM -> Fix bottleneck of DDR3 Bandwidth and use full power of the chip
-No HBM but Delta Color Compression and other updated tech to lower bandwidth needs to partially fix DDR3 bandwidth issue.


Sheesh... I didn't look up Carrizo for a while because I was waiting for the official announcements...but all the current guesstimations are looking rather bad and I hope that my FM2+ system wasn't a total failpick.

Edit: Also it should be clear by now that I don't actually give a damn about what happens to the laptop market for now. I solely care about FM2+ Desktop and the promised continuous upgrades that AMD promised to deliver until 2016...if Carrizo is a sidegrade then I'll have to remove myself from that platform.

 

[NTMBK]

Sheesh... I didn't look up Carrizo for a while because I was waiting for the official announcements...but all the current guesstimations are looking rather bad and I hope that my FM2+ system wasn't a total failpick.

Hey, just pop the FM2+ motherboard into a HTPC case with a 65W Carrizo and you should have an excellent Steambox.

 

[Shehriazad]

Hey, just pop the FM2+ motherboard into a HTPC case with a 65W Carrizo and you should have an excellent Steambox.

It's already an ITX board in an ITX case. I'm a sucker for small PCs. Either way the point is that I expect Carrizo to be > than Kaveri.

But if it's just equal in performance and the only real perk being 65W instead of 95 (which isn't a big deal for Desktop PCs, anyway) then I'll slowly but certainly have to think about getting rid of FM2+....and I hoped I could make it last until AMD churns out that new architecture in 2016.

It's a shame, really, since I just had a killer deal on the 860K which caused me to exchange my 7850 for the 860K (I was able to get the 860K so cheap that I could sell the 7850K with a profit since I had already gotten a dedicated card as birthday gift).

But that's going off-topic.

I'll make sure to keep my eyes peeled for more official info on Carrizo...there's still a chance I won't be let down.

 

[maarten12100]

Ugh...so doesn't HDL in the end mean that the frequency will forcibly take a hit again and thus ruin any possible IPC gains yet again? (Since Kaveri already did the same.... higher IPC...but ruined by the fact that it clocks 400 MHZ lower than the previous generation).

Sheesh....I really hope AMD doesn't mess this up. I really expected Carrizo to be clocked at 4 GHZ base again. Not a fan of the thought of looking forward to a 3.3 GHZ chip even if the IPC is raised to "make up for it".

I mean fine...I won't see a Hexacore FM2+ CPU...but at least don't destroy my dreams of actually seeing a Carrizo at 4 GHZ base core clock


In a perfect world Carrizo would be made in 20NM, Have 2GB of HBM, be clocked at 4.1 GHZ base with a 10% IPC gain over Kaveri (and thus finally destroying Richland hard on the CPU side) and have its' GPU cores with a full feature set of RX 300 with the ability of being crossfires with a R7 350

Sadly I fear that another Kaveri is going to happen instead...and I hope that I'm wrong at least a little bit. (Don't get me wrong...Kaveri wasn't terrible...but they cut corners on all the wrong ends in my eyes)

Actually we already have somewhat of a confirmation with the clocks of the prototype being 2,6GHz instead of 3GHz AMD targets efficiency but how well it clocks is yet to be seen.

If you speculate about it Carrizo is basicly as good as the early 20nm nodes could be if made on FD SOI. Whether that is the case we don't know.

on a side not NM stands for Newton Mega which isn't what you'd want to say. What you'd want to say is nm as it stands for nano meter which is what you're looking for.

Carrizo is laptop oriented, from the sounds of it. Integrated FCH which can't be used on the FM2+ motherboards, and reduced max clock speed in favour of improved power consumption.

I actually opened up my Kaveri notebook and spotted the FCH I was stunned at first then I felt the die and it wasn't outputting a lot of heat at all actually it didn't even have a heatsink. I'm not sure how such a part of the chip can shave of so much of the power consumption supposedly.

Better prepare yourself to be disapointed then. I'm pretty sure AMD stated, that the max TDP for Carrizo will be 65W on desktop AND that they use High Density Libraries. While these give a lot of advantages elsewhere they will limit max clock speed. This makes 4.0 Ghz Carrizo extremely unlikely.

Kaveri clocked to 4,7GHz and there have been a lot of overclockers that managed to push it higher while keeping stable enough. Now the consumption and the needed voltage did skyrocket going from 4,3 to 4,4GHz if I recall correctly but it was able to clock higher.

I have no doubt that with a decent board and cooling solution Carrizo can do 4,5GHz

I have to disagree. Laptops/mobile is the only place that APUs make sense to me, especially with the limited performance available today because of bandwidth restrictions. Unless AMD makes huge strides with Carizzo, you are compromising both cpu and gpu performance relative to Intel plus discrete. In a laptop, where TDP, space, and thermal budget are limited, those compromises are much more acceptable than in a desktop where you can get 50% more performance for very little additional cost. I think AMD is definitely correct to focus on laptops with this chip. Whether they can execute effectively is another story. They have been given a bit of breathing room with all the delays of 14nm from Intel, so they might be able to carve out a nice segment of the market.

Desktop Kaveri didn't make sense because of how awful the DDR3 was for it. And of course the fact that a dgpu is cheap.

They have been giving breathing room by Intel's delays however Intel's early 14nm products designated core M are just horrible. There were only 2 reviews out last time I checked for the Yoga 3 pro (horrible piece of junk) and the Asus transformerbook T300FA (cheap but horrible). The performance and efficiency of the core M platform seems to be a let down. Actually Intel's Baytrail and AMD's Mullins can match that efficiency already with ease.

 

[geoxile]

Yes. It should be included since it creates heat when used.

That sounds pretty good. So then the actual TDP for the APU minus FCH is probably 13 to 14W

 

[turtile]

Performance doesn't really matter for AMD's success with this chip. AMD is making what OEMs want. Have you looked at the laptop market recently? Everyone is using low power, low performance Intel CPUs.

Why? Using less power allows for the use of 3 cell batteries and a cheaper motherboard. AMD eliminating yet another chip also cuts down on OEMs' cost.

The laptop market is in a race to the bottom and power saving is the best way to make a cheap laptop. They don't care about performance...

 

[schmuckley]

They should probably make a server chip.

 

[maarten12100]

Performance doesn't really matter for AMD's success with this chip. AMD is making what OEMs want. Have you looked at the laptop market recently? Everyone is using low power, low performance Intel CPUs.

Why? Using less power allows for the use of 3 cell batteries and a cheaper motherboard. AMD eliminating yet another chip also cuts down on OEMs' cost.

The laptop market is in a race to the bottom and power saving is the best way to make a cheap laptop. They don't care about performance...

The cells are litterly 3 dollar pcs and that is for high capacity 4000mAh cells the cylindrical cells go for 1,5 dollar pcs for 2600mAh cells.

And that is when you buy regular if you buy bulk you'd be paying half that price. In other words they want to reduce weight since the cost is next to nothing. We really need a elite division over at AMD that sells high quality notebooks under the AMD brand name.

They should show that they are better than Intel's ULV i3 and i5. A SSD, a high end screen, a big battery and efficient hardware are key to winning over the consumer.

Lipo batteries have a density of up to 250Wh/kg this could be a wonder weapon laptop. Swap out all the metal for magnesium or magnesium alloys and use a lot of carbon fiber(epoxy it cause it conducts though) use SSDs without enclosure don't use too much bezel material. By reducing weight implement as big of a battery as you can fit in the weight budget and make it so that most of the mass is to the front. (so you can pick it up without breaking your wrists)

Actually a common problem with laptops is that the high weight parts like the cells are at the back and as we know F_arm = F * s

 

[geoxile]

Has there been any indication whether Carrizo will include FIVR and the other power saving features? Or is that only on the 20nm APUs

 

[Erenhardt]

Did anyone find an arcitle on memory bandwidth utilization improvement with Tonga's compression technique? If it makes its way to an APU, and the difference between GPUs and IGP in bandwidth utilization is the same we could draw an estimated Carrizo IGP performance improvements at the same specs as kaveri.

 

[geoxile]

Did anyone find an arcitle on memory bandwidth utilization improvement with Thonga's compression technique? If it makes its way to an APU, and the difference between GPUs and IGP in bandwidth utilization is the same we could draw an estimated Carrizo IGP performance improvements at the same specs as kaveri.

Isn't it up to 40% more efficient? So 140%.

 

[Erenhardt]

Isn't it up to 40% more efficient? So 140%.

Right, they claimed it on several occasions.

Given how Kaveri is bandwidth starved:

7750 GDDR5 is 50% faster, while 7750 DDR3 is about 7850k with 2400Mhz RAM level.

Even small increase in bandwidth bring lots of performance increase in games:
Less then 5% bandwidth increase (only 5%?):

Brings 10-20% performance increase:

With 40% better bandwidth utilization new APU should blow kaveri out of the water with the same memory speed (and if not TDP limited this time )

 

[JDG1980]

Also I doubt they went for SP count increase due to mem. BW woes they have and will have (DDR3 dual ch.).

Tonga's delta compression seems to allow considerably better memory bandwidth usage than the older GCN designs, so if that technology is incorporated into Carrizo, it should help considerably. That said, Kaveri was bottlenecked below its maximum capacity; it's available with up to 512 SPs, but that doesn't provide much advantage over 384 SPs in most non-GPGPU tasks. So I doubt that Carrizo will go over 512 SPs either. Using Tonga as a benchmark, AMD seems to have achieved about the same performance as Tahiti with the same number of SPs despite Tahiti having 1.36x the amount of memory bandwidth. So if Carrizo uses this GCN technology, we should expect it to have about enough bandwidth to handle 384 x 1.36 shaders, which comes to 522. But they obviously won't use that exact odd number. So they'll probably stick with 512 SPs for the biggest part, except this time it will actually provide better performance than 384 SPs.

 

[JDG1980]

Well but the laptop sales don't seem all that great to suddenly focus on that...or am I totally wrong here? I know that the mobile version is supposed to come out before the Desktop version...but that's about it.

The APU market just seemingly started to get interesting...is AMD really going to take a massive dump on this by making Carrizo weaker than Kaveri on Desktop?

AMD doesn't have competitive CPUs on the desktop, and won't until Zen comes out in 2016. The construction equipment architecture just isn't workable, and will never provide performance to match Sandy Bridge. As former CEO Rory Read put it: "Everyone knows that Bulldozer was not the game changing part [expected] when it was introduced three years ago. We have to live with that for four years." So basically at this time they're trying to salvage what they can, hoping that Carrizo with HDLs and lower clocks will be able to compete on laptops. AMD has scored some surprising design wins with Apple; maybe they're trying to see if they can get Carrizo into a future Retina MacBook Air? That would explain why the slide is talking about a 15W TDP figure...

I don't expect Zen to leapfrog Intel's CPUs (that would be a miracle on the limited R&D budget that AMD has), but I think they'll at least bring IPC up to somewhere between Sandy Bridge and Haswell levels (they have to know that low IPC is what's killing them with the construction equipment cores, and that high-clock, low-IPC designs are a dead end, as Intel found out with Netburst).

 

[geoxile]

I don't think Carrizo will clock lower than Kaveri if they're using SOI

 

[monstercameron]

AMD doesn't have competitive CPUs on the desktop, and won't until Zen comes out in 2016. The construction equipment architecture just isn't workable, and will never provide performance to match Sandy Bridge. As former CEO Rory Read put it: "Everyone knows that Bulldozer was not the game changing part [expected] when it was introduced three years ago. We have to live with that for four years." So basically at this time they're trying to salvage what they can, hoping that Carrizo with HDLs and lower clocks will be able to compete on laptops. AMD has scored some surprising design wins with Apple; maybe they're trying to see if they can get Carrizo into a future Retina MacBook Air? That would explain why the slide is talking about a 15W TDP figure...

I don't expect Zen to leapfrog Intel's CPUs (that would be a miracle on the limited R&D budget that AMD has), but I think they'll at least bring IPC up to somewhere between Sandy Bridge and Haswell levels (they have to know that low IPC is what's killing them with the construction equipment cores, and that high-clock, low-IPC designs are a dead end, as Intel found out with Netburst).

I don't think it has to match sandybridge or any intel processor to be competitive, especially when there are only two players in a growing market. AMD just needs to strengthen it's ties to some strategic partners and start getting more design wins.