Kabini Rumors

[Abwx]

Well, yes. But probably more realistic than the people expecting Steamroller +20% IPC single-threaded.

We all know that it s 20% at the module level , so why distort
people sayings just to make your expectations look credible.?...

 

[Abwx]

http://www.tomshardware.com/reviews/core-i7-4770k-haswell-performance,3461-6.html

We are about 20-25% behind trinity in the tested games on an ES with crappy drivers.

That makes 33-38% behind Richland wich is the real comparison.....

 

[Ventanni]

Damn it lol.... AMD needs to get their CPUs to be more power effecient.
Thought that HDL automated design thingy was gonna be in kavari...

The question is, if a computer could beat hand drawn designs why the hell didnt they use that methode long before now? so Steamrollers ect could have had it too?

Their passing up ~30% power figours due to bad hand made designs? wtf.

Do you know if Jaguar suffers from the same fate?
ei poorly handdrawn design, that a automated design by a computer could beat?

I remember reading a huge thread on this specific subject here once.

Design automation is heavily used in GPU development, so this would be an example of AMD using technology learned from one department (GPU) for the benefit of another (CPU).

And I'm pretty sure (not positive though) Jaguar uses this technology to improve power efficiency and reduce die area.

 

[Vesku]

AMD announces the embedded jaguar SKUs:

http://www.amd.com/us/press-releases/Pages/amd-embedded-gseries-2013apr23.aspx

Anyone know where mini-itx with these on them might pop up for sale?

Models and pricing

Models available at launch include:

GX-420CA SOC with AMD Radeon™ HD 8400E Graphics
Quad-core, 25W TDP, CPU freq. 2.0GHz, GPU freq. 600MHz
GX-415GA SOC with AMD Radeon™ HD 8330E Graphics
Quad-core, 15W TDP, CPU freq. 1.50GHz, GPU freq. 500MHz
GX-217GA SOC with AMD Radeon™ HD 8280E Graphics
Dual-core, 15W TDP, CPU freq. 1.65GHz, GPU freq. 450MHz
GX-210HA SOC with AMD Radeon™ HD 8210E Graphics
Dual-core, 9W TDP, CPU freq. 1.0GHz, GPU freq. 300MHz
GX-416RA SOC
Quad-Core, 15W, CPU Freq. 1.6GHz, No GPU

Pricing ranges from $49 - $72 for the SKUs.

 

[Cerb]

AMD announces the embedded jaguar SKUs:

http://www.amd.com/us/press-releases/Pages/amd-embedded-gseries-2013apr23.aspx

Anyone know where mini-itx with these on them might pop up for sale?

Maybe nowhere, if Zacate ones were any indication. OTOH, x86 embedded, other than Pentium and i3 CPUs, with ECC again, could provide some sales, if (a) priced right, and (b) if they get some SBC makers onboard. Anything that gives them real cash flow would be good, now.

 

[erunion]

. I think you are off on Kaveri though, unless you think Haswell will have a 6-9 month span as main segment.

probably 12 months for haswell. But Probably 10-11 for Richland, putting Kaveri just a couple months before broadwell.

AMD said kaveri will ship in 2013, not be available.

 

[Pilum]

Conroe was as much as 60% faster than Netburst . I mean seriously , there are so many thing wrong and out of place about Bulldozer/Piledriver that it's impossible not to squeeze out 20% ....

With an unlimited budget or unlimited time, yes, that's doable. AMD has neither.

The comparison to Conroe really doesn't apply, Core2 didn't happen out of the blue. It was derived from the Pentium-M, which came out in 2003. The work on Pentium-M probably started around 1999/2000, so we're looking at 6-7 years of engineering to give us Conroe; and probably with a far higher development budget than AMD could afford right now.

And whatever AMD may do to improve decode, scheduling and cache hierarchy, I'm petty certain they won't get close to Intel in integer single-threaded performance unless they widen the integer clusters. This may happen with an additional integer pipeline or realizing the "AGLU" concept (executing certain simple integer instructions in the AGU pipelines). But AMD has not mentioned any such improvement in the integer clusters at all, so I highly doubt we'll see this in SR.

 

[Idontcare]

The question is, if a computer could beat hand drawn designs why the hell didnt they use that methode long before now? so Steamrollers ect could have had it too?

Think about weather forecasting.

For thousands of years man forecast his own weather, and got increasingly better at it in the more recent centuries with the invention of accurate timepieces, thermometers, barometers, and hygrometers.

And then along came computers, which relied on the same inputs and same datasets (plus more, solar radiance, etc), and ever more complex weather models.

In those early days of computer-aided weather forecasting the results sucked, the computers were great if you wanted a tomorrow's forecast 10 days from now. (the computers were way too slow to be practical or timely, let alone accurate or relevant)

Over time, decades, the computers got much faster. And the models themselves got much more accurate.

To the point now where, depending on where you live, a 5-7 day forecast will be pretty darn accurate.

It is like that with computer-aided anything. Be it weather forecasting, IC design, computational chemistry, etc.

It takes time for the computers themselves to become fast enough that it becomes worthwhile to bother trying to use them for the job. And it takes time to develop the software applications such that the scope of the job can be confidently attempted.

 

[Arkadrel]

@Idontcare

Fair enough. Do you know if the jaguar cores are hand drawn?
Could they too see 15-30% gains from simply being redrawn by a automated program?

Does Intel do the same thing? how long have they had it if so?
if not, why not?

 

[MisterMac]

Think about weather forecasting.

For thousands of years man forecast his own weather, and got increasingly better at it in the more recent centuries with the invention of accurate timepieces, thermometers, barometers, and hygrometers.

And then along came computers, which relied on the same inputs and same datasets (plus more, solar radiance, etc), and ever more complex weather models.

In those early days of computer-aided weather forecasting the results sucked, the computers were great if you wanted a tomorrow's forecast 10 days from now. (the computers were way too slow to be practical or timely, let alone accurate or relevant)

Over time, decades, the computers got much faster. And the models themselves got much more accurate.

To the point now where, depending on where you live, a 5-7 day forecast will be pretty darn accurate.

It is like that with computer-aided anything. Be it weather forecasting, IC design, computational chemistry, etc.

It takes time for the computers themselves to become fast enough that it becomes worthwhile to bother trying to use them for the job. And it takes time to develop the software applications such that the scope of the job can be confidently attempted.

....but one of the big flaws as stated by certain engineers in the press about bulldozer was synthetic design?
Automated processes - because it was cheaper - not more effective.

Using more of that - seems dubious if you want to maximize whatever process your on?

 

[Exophase]

Fair enough. Do you know if the jaguar cores are hand drawn?
Could they too see 15-30% gains from simply being redrawn by a automated program?

At Hot Chips and probably other presentations AMD gave demonstrations of Jaguar being almost entirely machine synthesized with only a few custom macros. They've shown that it is even more synthesized than Bobcat was.

Density is probably the biggest advantage to really fine-grained synthesis, you could have totally different functional blocks have very complex boundaries and wrap around each other in ways that'd be almost impossible to design by hand. Another advantage AMD stressed is that it makes it easier to port the design to different processes.

 

[Vesku]

It takes time to input and bake in all the necessary human knowledge to automate tasks. The more complex it happens to be the more time, iterations, and correcting of mistakes or quirks. Oh yes, and also the need for more brute force computing power, Intel runs a massive installation. I recall VIA showed their water cooled overclocked design cluster to a tech journo tour.

 

[mrmt]

Another advantage AMD stressed is that it makes it easier to port the design to different processes.

And for who can consider this an advantage besides AMD?

Every 1st tier player out there chooses a foundry and sticks with it, only AMD has to think about backporting designs between foundries. One has to wonder what kind of trade offs AMD is having to make to achieve such a portability. More time in validation? Less density? Less complex units?

 

[Exophase]

And for who can consider this an advantage besides AMD?

Every 1st tier player out there chooses a foundry and sticks with it, only AMD has to think about backporting designs between foundries. One has to wonder what kind of trade offs AMD is having to make to achieve such a portability. More time in validation? Less density? Less complex units?

More portable to other processes would also include new processes made by the same foundry so that could be an advantage for anyone..

As for wanting to be able to move to other foundries, one only needs to look at Qualcomm or nVidia's rumblings about TSMC failing to meet their needs or Apple moving from Samsung.

 

[Khato]

This happends with Kavari right?

...for power constrained designs, sounds like mobil to me.

Quite the well-done marketing slide. They're implying improvements to power and density with the HDL, but at the same time they specifically state that the hand-drawn Bulldozer FPU was optimized for maximum speed and density while giving no indication as to the parameters for the HDL synthesis. What in that slide states that the HDL design can reach the same clock speeds as the hand-drawn? 15-30% lower power per operation isn't very impressive if you're running at lower frequencies... or the more likely case of comparing at same low frequency, but with the HDL design optimized for that frequency and unable to go higher.

Edit: Note that I'm by no means claiming hand-drawn is better. It has its uses, but automated synthesis continues to get better and may eventually surpass it in all cases. But the only reason for it to have a marked advantage over hand-drawn designs like they're claiming is if you have a pretty bad hand-drawn design.

 

[Exophase]

That reminds me.. another advantage to automated synthesis is that it's easier to change what the design is optimized for. It could be interesting to have different cores with different optimizations on the same die (AFAIK nVidia already does this with Tegra 3 and 4, others probably do as well). For instance, with AMD's FX offerings you can't reach the same max turbo with more than one module on, so you could possibly optimize all but one module to hit lower clock speeds if it means they're denser or use less power.

I'm sure there are some serious complications to that though.

 

[mikk]

You can add ~10% to the HD4000 just from the latest drivers. And the same, if not more to the HD4600.

Benchmarking GPUs in prerelease is one of the worst things you can do to make a conclusion.

The same site also seems to mix up TXT and TSX...

They already tested with a 15.31 driver because Luxmark made a huge jump due to the new OpenCL 1.2 driver. I don't expect a speedup from a newer 15.31 driver. I expect a small speedup from a final platform (their ES platform had bandwidth issues which isn't nice for the integrated iGPU).

 

[mrmt]

More portable to other processes would also include new processes made by the same foundry so that could be an advantage for anyone..

The life of a bleeding edge SKU is 18-24 months, after that it gets superseded by the next big thing. That's not a life span big enough to deserve porting to the next node.

But in AMD's case, I can see where this is going: Embedded. In embedded you can get life spans big enough to deserve a porting.

As for wanting to be able to move to other foundries, one only needs to look at Qualcomm or nVidia's rumblings about TSMC failing to meet their needs or Apple moving from Samsung.

See above. Apple is moving new SKUs to TSMC, not porting what they had at TSMC.

 

[Vesku]

Well if it was easy thanks to automation then companies could take some of their older high performance parts and re-release them in lower segments. A64 x2 shrunk to a mainstream 2009-2010 node for example or perhaps a Core 2 series on Intel 32nm?

 

[Exophase]

The life of a bleeding edge SKU is 18-24 months, after that it gets superseded by the next big thing. That's not a life span big enough to deserve porting to the next node.

Compare the die shots of Apple's 45nm and 32nm A5 some time. They're not the only ones who do shrinks (nor is Intel, the most blatant).

 

[ALIVE]

AMD announces the embedded jaguar SKUs:

http://www.amd.com/us/press-releases/Pages/amd-embedded-gseries-2013apr23.aspx

Anyone know where mini-itx with these on them might pop up for sale?

Models and pricing

Models available at launch include:

GX-420CA SOC with AMD Radeon™ HD 8400E Graphics
Quad-core, 25W TDP, CPU freq. 2.0GHz, GPU freq. 600MHz
GX-415GA SOC with AMD Radeon™ HD 8330E Graphics
Quad-core, 15W TDP, CPU freq. 1.50GHz, GPU freq. 500MHz
GX-217GA SOC with AMD Radeon™ HD 8280E Graphics
Dual-core, 15W TDP, CPU freq. 1.65GHz, GPU freq. 450MHz
GX-210HA SOC with AMD Radeon™ HD 8210E Graphics
Dual-core, 9W TDP, CPU freq. 1.0GHz, GPU freq. 300MHz
GX-416RA SOC
Quad-Core, 15W, CPU Freq. 1.6GHz, No GPU

Pricing ranges from $49 - $72 for the SKUs.

an htpc with the GX-420CA SOC with AMD Radeon™ HD 8400E Graphics
Quad-core, 25W TDP, CPU freq. 2.0GHz, GPU freq. 600MHz
sounds wonderfull 25watt for a 4 core cpu and a gpu
so when this baby is out if only it supported ecc :-(

 

[Vesku]

These embedded models do support ECC, it's just you'll have to do some real digging to find a board you can buy in single quantity for a reasonable price (due to being targeted at embedded applications). Going to be keeping an eye out, will update if I have any luck.

 

[ALIVE]

These embedded models do support ECC, it's just you'll have to do some real digging to find a board you can buy in single quantity for a reasonable price (due to being targeted at embedded applications). Going to be keeping an eye out, will update if I have any luck.

well when intel released the atom330 i manage to find and buy the 2 boards made for them in flex atx size mainboard.
one with pci-e x1 and 2 pci
the other pci-e x16(x4 electrical) pci-e x1 and pci
i do not like the mitx mainboard too tiny too limited for my taste
was thinking for an e450 mainboard the asus matx one but then it was close for the refresh of the line with the e1 e2 models which never happened for the desktop. and then it was atom s1260 out or close out which could make a good nas and now again waiting for amd to see if it worths my money. so keep running the 330 waiting to replace them.

 

[CTho9305]

If they are going to take an extra 1-2yrs to develop it and get it into production (versus the competition), it sure as heck ought to be better.

Project management 101.

Is it in production yet or are we still hearing marketing claims? If it's in production, do we have any reasonably comparable designs to look at?

Damn it lol.... AMD needs to get their CPUs to be more power effecient.
Thought that HDL automated design thingy was gonna be in kavari...

The question is, if a computer could beat hand drawn designs why the hell didnt they use that methode long before now? so Steamrollers ect could have had it too?

Their passing up ~30% power figours due to bad hand made designs? wtf.

Do you know if Jaguar suffers from the same fate?
ei poorly handdrawn design, that a automated design by a computer could beat?

Never underestimate the power of inertia. In the past, automated design tools weren't very good, so CPU designs were done by hand for performance. Once you have a team of hand-design experts, they don't want to believe their skills aren't useful. On top of that, it takes a lot of work to get the best results from the tools; if you take an optimized hand design and compare it to a one-off automation experiment, you're likely going to conclude you should stick with hand design. If you do a quick and dirty hand design and compare it to an optimized automated design, you're going to conclude you should stick with automation. A fair comparison requires a large effort.

....but one of the big flaws as stated by certain engineers in the press about bulldozer was synthetic design?
Automated processes - because it was cheaper - not more effective.

Using more of that - seems dubious if you want to maximize whatever process your on?

I don't think that guy knew what he was talking about. If I remember correctly, he'd left AMD long before Bulldozer was finished (a link from 2011 says "a few years ago").

 

[Pilum]

That reminds me.. another advantage to automated synthesis is that it's easier to change what the design is optimized for. It could be interesting to have different cores with different optimizations on the same die (AFAIK nVidia already does this with Tegra 3 and 4, others probably do as well). For instance, with AMD's FX offerings you can't reach the same max turbo with more than one module on, so you could possibly optimize all but one module to hit lower clock speeds if it means they're denser or use less power.

I'm sure there are some serious complications to that though.

Yes, OS support. You'd need new CPUID functions to identify the "fast" core and the OS would need to know how to schedule workloads for that. In embedded that would be rather easy, but on the desktop... pretty much no way.