AMD Carrizo Pre-release thread

[NTMBK]

Well the x8 vs x16 lanes for a single graphics card (or anything you mash in that PCI-E slot) should not have an impact.

There's hardly a card that can actually utilize the x16 lanes in PCI-E gen 3.....it's just too much.

Even with PCI-E gen 2 the difference would be like 1.5 fps at best in games (granted you use a high end card...which again...is kinda weird if you use that kind of APU/CPU)


The only downside is that you can't stroke your ego with the fact that you have x16 lanes. PCI-E gen 3 in its entirety is still kind of unnecessary...not saying that it will stay that way...just how it is right now xD

Texture sizes just took a massive jump with the move to next gen consoles... PCIe bandwidth does matter. It may not show up a lot in benchmarks that only measure average framerates, but it can make a big difference to minimum framerates and general frametime consistency.

 

[lefty2]

In the Anandtech article they say Carrizo has full HSA features unlike Kavari.
Does anyone know which features specifically Kavari is missing?

 

[NTMBK]

In the Anandtech article they say Carrizo has full HSA features unlike Kavari.
Does anyone know which features specifically Kavari is missing?

Mid-wave preemption.

 

[lefty2]

Mid-wave preemption.

Thanks. Sounds like it's only a minor improvement then.

 

[pw257008]

List 10 Beema notebooks that can be bought in American stores.

Ok, here's a list. Not all of them are Beema, but most of the AMD sold there are, with maybe one Kaveri, one or two Richland, and a smattering of Kabini. You can also type in A8-6410 on Amazon, or A6-6310, and find results.

 

[geoxile]

I'd really like to know what goes through the head of laptop designers as they design laptops for beema/mullins.

 

[mrmt]

I'd really like to know what goes through the head of laptop designers as they design laptops for beema/mullins.

Costs.

 

[NostaSeronx]

I'd really like to know what goes through the head of laptop designers as they design laptops for beema/mullins.

Costs.

Intel Rebates.

 

[geoxile]

Costs.

But not product viability? Well that explains a lot.

 

[Shehriazad]

Texture sizes just took a massive jump with the move to next gen consoles... PCIe bandwidth does matter. It may not show up a lot in benchmarks that only measure average framerates, but it can make a big difference to minimum framerates and general frametime consistency.

Nope. Not really (yet). If you google for tests for pci-e 8 vs 16 gen 3 you will see that they tested 4K as well.

Had no impact on the performance difference whatsoever. And even the measured results are mostly in the margin of error.


You need to keep in mind that PCI-E gen 3 will move 32gbyte/s in 16x and 16gbyte/s in 8x mode. Now show me a game that needs to move 16Gbyte graphics data per second. Most of the data is handled internally by the GPU...and PCI-E bandwidth exceeds CPU bandwidth in a lot of cases...so what good does it do to eat your cereal with a shovel?

As you can see even PCI-E Gen 2 can easily keep up(even in x8 mode)...and sometimes x16 gen 3 is surprisingly slower than gen 2. No matter what you claim...we are still far from games moving 16gb/s of visual data...and faaaaaaaar from 32gb/s. What you are thinking about is the data that is handled by the bus that handles things like textures and is linked to the GPUs own ram (128 bit bus 256 bit bus and so on) which is internal bandwidth.

 

[III-V]

Well, he was saying it won't show up on typical framerate plots, but will on latency plots. I am still doubtful that there'd be a measurable difference.

 

[NTMBK]

Nope. Not really (yet). If you google for tests for pci-e 8 vs 16 gen 3 you will see that they tested 4K as well.

Had no impact on the performance difference whatsoever. And even the measured results are mostly in the margin of error.


You need to keep in mind that PCI-E gen 3 will move 32gbyte/s in 16x and 16gbyte/s in 8x mode. Now show me a game that needs to move 16Gbyte graphics data per second. Most of the data is handled internally by the GPU...and PCI-E bandwidth exceeds CPU bandwidth in a lot of cases...so what good does it do to eat your cereal with a shovel?

As you can see even PCI-E Gen 2 can easily keep up(even in x8 mode)...and sometimes x16 gen 3 is surprisingly slower than gen 2. No matter what you claim...we are still far from games moving 16gb/s of visual data...and faaaaaaaar from 32gb/s. What you are thinking about is the data that is handled by the bus that handles things like textures and is linked to the GPUs own ram (128 bit bus 256 bit bus and so on) which is internal bandwidth.

Did you even read what I wrote? That kind of average framerate benchmark won't show it. It shows up in minimum framerate dips, and taking longer for full res textures to pop in.

 

[NostaSeronx]

Well if the textures have to be fetched from a hard-drive/ssd.

You have to take the PCI Express 2.0 x4(UMI(AMD)/DMI(Intel)) as the bottleneck. Next to the speed of the hard-drive/SSD.

 

[NTMBK]

Well if the textures have to be fetched from a hard-drive/ssd.

You have to take the PCI Express 2.0 x4(UMI(AMD)/DMI(Intel)) as the bottleneck. Next to the speed of the hard-drive/SSD.

Oh that's definitely a factor too. SSD -> main memory, then main memory -> graphics memory. Which one matters more depends on how much memory your PC has, and what sort of caching strategy the game uses.

 

[Shehriazad]

Did you even read what I wrote? That kind of average framerate benchmark won't show it. It shows up in minimum framerate dips, and taking longer for full res textures to pop in.

Textures are first loaded from your harddrive...if you have an SSD this pretty much tops out at like 600mb/s.

And after this is loaded into the Vram...the Vram manages this internally. A measly 128 bit bus reaches 88GB/s...but 256/384 and more are the standard by now. So some graphics cards have above 200GB/s....which means that they are many times faster than the PCI-E slot.

With your logic the Vram would be bottlenecked by the PCI-E slot regardless of generation...but that's not how graphics cards work, bro.

Show me hard proof that PCI-E gen 3 8 vs 16 will show "big" minimum fps fluctuations or that texture load times in games are influenced a lot.

Because I ain't buying what you're selling me.


PCI-E does not make the information go any faster. (Gen 2 has slightly higher reaction times, but there is no reaction time differences between 8x and 16x...only the max bandwidth) It just says how much information can pass at once...which ends up being a difference that can not be measured since no game will require a constant stream of 16/32GB of data per second.



This does go heavily off topic...but trying to tell me that there is an actual important difference between x8 and x16 in gen 3 when all the evidence seems to show something different is just magical.

Especially in a chip that will never be meant to be used with high end graphics cards....

 

[NTMBK]

Textures are first loaded from your harddrive...if you have an SSD this pretty much tops out at like 600mb/s.

And after this is loaded into the Vram...the Vram manages this internally. A measly 128 bit bus reaches 88GB/s...but 256/384 and more are the standard by now. So some graphics cards have above 200GB/s....which means that they are many times faster than the PCI-E slot.

With your logic the Vram would be bottlenecked by the PCI-E slot regardless of generation...but that's not how graphics cards work, bro.

Show me hard proof that PCI-E gen 3 8 vs 16 will show "big" minimum fps fluctuations or that texture load times in games are influenced a lot.

Because I ain't buying what you're selling me.

All I know is that swapping from a motherboard with PCIe Gen1 and Hypertransport 1 to PCIe Gen2 and Hypertransport 3 gave me a noticeable improvement in game smoothness and reduced stuttering, with all the components plugged into it unchanged. No, I didn't measure it, and no I can't go measure it because that janky old AM2 motherboard went in a bin. But it was noticeably better.

Just want to help people out, no need to get defensive

 

[Shehriazad]

All I know is that swapping from a motherboard with PCIe Gen1 and Hypertransport 1 to PCIe Gen2 and Hypertransport 3 gave me a noticeable improvement in game smoothness and reduced stuttering, with all the components plugged into it unchanged. No, I didn't measure it, and no I can't go measure it because that janky old AM2 motherboard went in a bin. But it was noticeably better.

Just want to help people out, no need to get defensive

Reduced stuttering and smoothness are rather a chipset and generation tech update thing.

Generation 2 and 3 received a lot of tech updates...but the bandwidth of generation 2 itself is still more than plenty.

It's the tech and lowered reaction times of the newer generations that make things neat.

But the bandwidth itself is an absolute non issue.

So I will believe you if you say that going from gen 1 to gen 2 /3 boosted your smoothness...but not on the basis of x8 vs x16 but rather on the basis of the updated tech being able to send/receive the data in a more reliable and stable fashion.


p.s. kind of unbelievable that they're already working on finishing up gen 4 of PCI-E in 2015...just seems so unnecessary. Especially since it retains full backwards compability...so it won't even be any smaller...sigh xD

 

[mrmt]

But not product viability? Well that explains a lot.

With the kind of margins AMD gets on their financials and still bleeding share, I think that costs play a fundamental role for OEMs when acquiring any AMD product.

 

[DrMrLordX]

p.s. kind of unbelievable that they're already working on finishing up gen 4 of PCI-E in 2015...just seems so unnecessary. Especially since it retains full backwards compability...so it won't even be any smaller...sigh xD

It's simple, really. It's about storage. SATA3 is already reaching its limits, so using PCIe is a somewhat logical solution to the problem. Improving the underlying protocol and standard means they can offer more bandwidth without necessarily offering more lanes.

Back to Carrizo. As reported by Anandtech, it's looking murky on the desktop side. The 2015 roadmap for desktop is "unpublished". Really? Okay. Hope they got something good up their sleeves.

 

[Shehriazad]

For now I'll hold on to my panties about the Desktop release...as long as they only say "Mobile first" and not "Nope, no Desktop this year(2015)". Ill think that we will still get our Carrizo Desktop model.

 

[monstercameron]

now we know AMD is just messing with us.
http://www.sisoftware.eu/rank2011d/...efdde4d0e7d2e0c6b489b99ffa9fa292b4c7fac2&l=en
AMD Radeon(TM) R7 Graphics; AMD Radeon(TM) R7 Graphics (896SP 14C 720MHz, 6GB DDR3 1.6GHz 64-bit, Integrated Graphics) (OpenCL)

 

[NostaSeronx]

now we know AMD is just messing with us.
http://www.sisoftware.eu/rank2011d/...efdde4d0e7d2e0c6b489b99ffa9fa292b4c7fac2&l=en
AMD Radeon(TM) R7 Graphics; AMD Radeon(TM) R7 Graphics (896SP 14C 720MHz, 6GB DDR3 1.6GHz 64-bit, Integrated Graphics) (OpenCL)

8CU(Kaveri) + 6CU(Oland)

------
Carrizo is most likely 28nm Advanced FDSOI. GlobalFoundries is very keen on FDSOI over FinFETs in their confidential presentations. AMD most likely saw those slides, lets not forget AMD is GloFo's #1 Customer. Especially, at an AMD IP standpoint as getting the China/India market is clearly 100% FDSOI. Add Samsung's support of 28/20nm FDSOI; it is pretty much shooting oneself in the foot, not to go FDSOI.

 

[Shehriazad]

*snip*

Well...FD-SOI is probably the most balanced option between cost efficiency and upkeep of moores law.

I don't see FinFET 16/14NM being a real thing for the next few years...the cost just seems retarded. (Normal consumers are not gonna pay for it...and the normal server market also would have no real use for it at the price...supercomputers? Well...sure...maybe)


But that makes me think one thing...and I know it's not the most likely scenario...but what if AMD is keeping quiet about the "Big Carrizo" for now because they're "waiting" (yes yes, technically it's ready) for 20NM while producing their Little Carrizos (mobile) with 28NM FD-SOI and HDL?


As far as I'm concerned the date of Desktop Carrizo has been pushed back to match with the release of the RX 300 series.

Which should be in 20NM and feature (SK Hynixs'?) HBM.


A chip like that would be an all kill...especially if it can dual up with the new GPU series.

That would be a huge comeback in the Desktop sector if I have ever seen one. (Especially since at least "momentarily" the CPU side of games is going to be less hungry in 2015/2016 thanks to DX12, Mantle, OpenGL)


But it's probably just wishful thinking. Except for the fact that SoCs are still THE option of the "far" future either way for the consumer PC. The way I see it dedicated CPUs /GPUs are not going to be a thing for forever in the consumer market.

 

[mrmt]

Well...FD-SOI is probably the most balanced option between cost efficiency and upkeep of moores law.

I don't see FinFET 16/14NM being a real thing for the next few years...the cost just seems retarded. (Normal consumers are not gonna pay for it...and the normal server market also would have no real use for it at the price...supercomputers? Well...sure...maybe)

I find amusing that people come with statements about the superiority of SOI over FINFETs when the most profitable foundries of the world are going FINFETs for 14/16nm and, the 2nd tier foundries are also developing finfet, despite licensing SOI nodes from STM. Sure, we can expect this from Seronx, but from other people? Really?

 

[Shehriazad]

I find amusing that people come with statements about the superiority of SOI over FINFETs when the most profitable foundries of the world are going FINFETs for 14/16nm and, the 2nd tier foundries are also developing finfet, despite licensing SOI nodes from STM. Sure, we can expect this from Seronx, but from other people? Really?

Well Im just putting together information I read on the internet myself...so feel free to show me (reliable) sources that FinFET RIGHT NOW would be more profitable per Wafer than SOI.

Because the reports I find always state that SOI comes out ahead(for now) and some even go as far as claiming that FInFET might actually break Moores law in its current state because of its cost.


If you can deliver well documented counter proof....I shall take back what I said.... so no need to go all angrypants here. (because Im definitely not a fanboy for some kind of chip production method...that's about as stupid as it can get)

I'm not some dude who "secretly works at Globalfoundries and has internal info" or anything.

I just pick up what I read on the bigger hardware and tech related sites.