The Kaveri Pre-Launch Thread (A10-7800 and A10-6800k @3,5 Ghz)

[monstercameron]

 

[Erenhardt]

Thanks! That looks really impressive. New process node and GCN are probably causing this.

 

[dosmastr]

which i3 ?

 

[Homeles]

Thanks! That looks really impressive. New process node and GCN are probably causing this.

28nm bulk should be worse on power. The improvements are from elsewhere.

 

[Headfoot]

I have what seems like a stupid question, but I'm having a hard time figuring it out.

If I'm about to build an entirely new rig and plan to splurge on some significant discrete GPU hardware, do I have any reason to be excited about Kaveri? So far, it seems that the big draw is for its graphical performance, but what if I want my CPU to be a CPU? Is Kaveri a good option I should be considering?

No, not really. Not unless you're on a tight budget but from the sounds of it you arent. Top end is still going to be 4770k/4960x, midrange 4670k. Kaveri might shake up the budget gamer space but the high end is going to be pretty much the same

 

[cbn]

Ok, so back on topic. What secret sauce is behind having a 2x larger IMC than Richland?

Is it that big because of the failed attempt at propelling GDDR5M? Is it actually a DDR3/DDR4 IMC, saving the latter for a future platform update (and thus screwing whoever bought FM2+ thinking it would last)?

The GDDR5m sounded fancy for embedded/mobile solutions, while DDR4 sounds reasonable for desktop, where you can actually populate four slots of RAM. Either way AMD made a decision and somehow that IMC is bigger that it should (for DDR3 only, that is).

My guesses regarding the memory interface section of the die map: either intentionally fudged or quad channel DDR3 support. Perhaps will get some quality analysis along the lines of that mysterious AMD die shot thread which we now know would have to be Excavator or beyond (was some speculation it was Steamroller at the time).

Quad channel makes sence, but will be not getting it on desktop, maybe some OEM will go for a itx with quad ddr3 with a mobile kaveri.

Here are some price comparisons for dual channel and quad channel (8 GB total):

2 x 4GB 2400 1.65 volt kit starts at $74.99 --> http://www.newegg.com/Product/Produc...CE&PageSize=20

4 x 2 GB 1600 kits starting at $84.99 ---> http://www.newegg.com/Product/Produc...PC3%2012800%29

4 x 2 GB 2133 1.65 volt kit for $96.99---> http://www.newegg.com/Product/Produc...PC3%2017000%29

4 x 2 Gb 2400 1.65 volt kit for $128.99---> http://www.newegg.com/Product/Produc...PC3%2019200%29

Being able to run four 1600 dimms in quad channel would give a 50% increase in bandwidth over dual channel 2400 speed kit for ~$10 more. (With the added benefit of the 1600 kit having 1.5 volts vs. the 1.65 volts of the 2400 kit)

Likewise, four 1.65v 2133 dimms in quad channel would give a 77% increase in bandwidth over a dual channel 1.65v 2400 speed kit for $22 more.

Four 1.65v 2400 dimms in quad channel would, of course, be double the bandwidth of a 1.65v 2400 speed dual channel kit for $54 more.

Based on that pricing I would say the quad channel is definitely worth it even if a person needed to buy the more expensive 2GB DIMMs. (EDIT: Let me just add in the caveat that the 2133 4 x 2GB DIMM kit is obviously a better deal than the 2400 4 x 2GB DIMM kit.)

 

[Slomo4shO]

Based on that pricing I would say the quad channel is definitely worth it even if a person needed to buy the more expensive 2GB DIMMs.

Only problem is that all the available FM2+ boards only support dual channel.

 

[inf64]

I seriously begin to think AMD will do some EDRAM action on Carrizo. It will be useless to add more iGPU resources when they will just sit idle being underutilized and wasting power and die area. They just have to add some additional level of fast-on package-memory for the next gen part. Since x86 cores will share whole address space with iGPU this would be good news for traditional workloads too.

 

[Asterox]

28nm bulk should be worse on power. The improvements are from elsewhere.

There is something called RCM, and now is fully operational on Kaveri APU as opposed to Piledriver FX CPU and Trinity/Richland APU models.

http://www.cyclos-semi.com/technology/

 

[Galatian]

Are you saying that an i3 can beat a similarly prices kaveri sku in most OCL/compute workloads?
http://www.anandtech.com/show/7032/...-gpu-on-the-desktop-radeon-hd-8670d-hd-4600/4
that is hd4600 vs vliw4 richland...

and iris/pro to mobile kaveri? you lost me.

Ahem...the benchmarks you have shown prove my point...they trading blows. And that coming from the chips most AMD APU fans call insufficient when it comes to OpenCL. The whole marketing for Kaveri is now on OpenCL for Christ sake. AMD makes it sound like they are the only ones that can do OpenCL and I pointed out now on several benchmarks that it is simply not true. I'll take better CPU performance + on par OpenCL performance + lower power consumption any day over lower CPU performance + on par OpenCL performance + higher power consumption and the ability to play some games on medium setting.

Also yes I make a comparison to mobile chips, because let's be real iGPU really shine in the mobile sector. I'm not sure how that is even questionable? Who cares about a iGPU on par of a Radeon 7750 on a desktop, where it really is cheaper and simpler to get a dGPU? Mobile is where it matters. That's why Apple adopted their chips, because...surprise, surprise...AMD has nothing to compare to the Iris 5100 or 5200 Pro power. Their APUs just suck way too much power and need much better cooling. Their mobile version are extremely slower then their desktop counterpart, which is not true for HD and Iris iGPU from Intel.

I'm repeating myself and never get an answer: where could I possibly use an APU on a desktop?

Price is a different question...no surprise. But again the market will pay...since AMD doesn't offer anything closely as powerful as Intels mobile chips, why would they charge less? Let's face it: if Intel would actually drop their prices on their Core i3 and some mobile chips there would be no reason at all to go with AMD IMHO.

Also (...and I have stated that before...is anybody actually listening to what I said?) AMD APUs on mobile are vastly underpowered compared to their desktop brethren. Even the HD4600 which is the standard chip on almost all Intel mobile chips compare to their fastest iGPU in games.

Surely that is exactly what it means? Unless Apple are able to magically make the heat disappear...

But, I can't see any mobile chip drawing that much. I measured the maximum draw I cold coax out of old C2D laptop (Thnkpad T400) while stressing the CPU + CD + HDD + RAM charging the battery at 55W. So unless the rMBP was running FurMark while also running Linx AVX2 I can't see it getting anywhere near that. (Having said all that, I didn't see any power figures in that Anand article.)

ATM, Iris Pro is not in competition with Kaveri as it is in totally different price range. Whether that changes with Broadwell is hard to say, but Intel do tend to charge a premium for their large dies. And from a pure transistor/perf point of view I don't really think Iris Pro is that spectacular. Intel's process node advantage allows them to try these brute-force approaches with a huge die but that doesn't mean it's a great design.

I'm lost...this comes straight from wiki (emphasis mine):

"The thermal design power (TDP), sometimes called thermal design point, refers to the maximum amount of heat generated by the CPU, which the cooling system in a computer is required to dissipate. The TDP is typically not the largest amount of heat the CPU could ever generate, such as by running a power virus, but rather the maximum amount of heat that it would generate when running "real applications." This ensures the computer will be able to handle essentially all applications without exceeding its thermal envelope, or requiring a cooling system for the maximum theoretical power (which would cost more but in favor of extra headroom for processing power)."

Intel states their Iris Pro equipped mobile chips have a TDP of 47W. Apple doesn't magically make it disappear. In fact it's a lot less then running both a CPU + dGPU in a notebook chassis. And this my friends is why iGPU matters most in mobile. And - again - this is where AMD fails. The HD4600 equipped mobile chips - which I already stated earlier have an almost comparable GPU power - use even less and are cheaper then Iris /Pro equipped ones.

But fear not...I'm getting out of this thread. I only posted here because I think AMD is doing a lot of falls advertisement and I have shown several benchmarks proving my point. No one has ever answered my question about a truly useful usage scenario for a APU on desktop, nor does anybody actually bring up numbers for the mobile APU chips. I've shown that Intel is more then price competitive and I have also shown that Intel can even beat the iGPU if it wants to and probably will with Broadwell. All I get in return are the same Benchmarks posted over and over again (that I have shown to be actually on par with Intel), without anybody actually taking the time to analyze them. Furthermore I'm getting thrown into a corner that I'm a Core i5 owner and that efficiency doesn't matter on a desktop, although I clearly stated - and logic dictates - when all things are equal (and they are not because Intel Core i3 CPU is actually even faster) I pick the one that is more efficient. Now I have several uniformed posts thrown at me thinking that TDP is the same as power draw.

I'm sorry...I'm completely lost...you know everybody can pick the CPU they want. But this is still a highly frequented forum and people go here to get an educated opinion on hardware. If I see statements like Kaveri owns I just have to look behind all the advertisement. Their claimed OpenCL performance is not so great after all. The Bulldozer Architecture after their 3rd Revision still looks weak. I just wish that some people around here would acknowledge that, but alas it seems that AMD fans are truly more vocal after all.

Again I couldn't care less and as has been stated before 99% of consumers probably can't even tell the difference between APUs and Intel Core i3. But this is still a hardware forum. Let's stop acting like childs and throw the same benchmarks around that we've already posted on page 1...to prove what point exactly? Nobody has even tried disproving my claims. Is that all you got? Just let my points faint into oblivion?

I'm out...

 

[dosmastr]

I seriously begin to think AMD will do some EDRAM action on Carrizo. It will be useless to add more iGPU resources when they will just sit idle being underutilized and wasting power and die area. They just have to add some additional level of fast-on package-memory for the next gen part. Since x86 cores will share whole address space with iGPU this would be good news for traditional workloads too.

I agree with you from an arch and performance POV... but that would be ALOT of die area (unless you're thinking of different amounts of memory) I just wonder if it would be feasible to produce and sell.

Maybe if they went back to slot A days and loaded ram up on the same card lol

 

[dosmastr]

Galatian,

The reading I did suggested that AMD's chip was built for 2020, not 2014 and certainly not 2011 when most apps are still single threaded.

the amount of parallelism built into the apu is impressive but its not able to leverage that today...and might not be able to in the near future without programs being heavily rewritten for multi-threaded...ness....

Good article:
http://www.extremetech.com/computing/100583-analyzing-bulldozers-scaling-single-thread-performance

I couldn't find the article that says what I paraphrased above... pretty sure it was another extremetech link though.

 

[NTMBK]

he just went shark on their @$$!

What.

 

[dosmastr]

yeah I misunderstood hit statement to be a shark tank (tv show) reference rather than "I give up, I'm leaving." which is why I edited it and removed it.

 

[KompuKare]

Don't know what you are complaining about. I took your original sentence

The Iris 5200 Pro might draw 85 power, but that doesn't mean [it] is dissipating that much heat.

as saying that while "Iris 5200 Pro" (meaning the CPU I guess) might draw 85W it doesn't have to dissipate that much heat. That is 85W =! 85W which is not possible. Maybe your sentence wasn't that clear.

And that struck me as the same kind of statements like the ones about the R9-290X being a good space heaters compared to the GTX 780Ti (seen a few statements like that) when both the 290X and 780Ti draw around 270W it's just that the 290X's core gets hotter. Point being the total energy consumed is the same and hence used as heater, both would be almost identical. So when you said that Iris Pro may draw 85W but that doesn't meant it dissipates that much heat, it struck me as counter-intuitive like asking if the element in one kind of electric heater is more efficient than another; since to me power in = power out.

So if a laptop draws 85W it has to dissipate 85W of heat; simple as. Now which fraction of that 85W is CPU, which is the screen or the power brick's efficiency etc. doesn't make any difference as to how much heat is generated because everything that goes in out has to go out as heat (a few photons excepted).

As to your other points: sure Intel are financially healthy enough to be able to sell Iris Pro for less but they are very attached to their margins. Iris Pro is a neat engineering solution (and AMD will have to something similar soon adding more shaders is getting to be a waste of die space) but I don't think its very efficient in terms of transistors/perf. For now, Intel have no intention of selling it at a cheap price. What 16nm and Broadwell bring is still unknown.

 

[ams23]

That is not exactly a great result for Kaveri because a GDDR5 HD 7750 variant will have much higher gaming performance than the more bandwidth-constrained Kaveri GPU. So the perf. per watt of the i3 + discrete card would actually be quite a bit better for gaming purposes.

And FWIW, the Blu Ray and HD video consumption is much worse on GCN cards than Kepler cards (even if the former is an improvement over prior gen. AMD cards).

 

[dosmastr]

never stayed up for a launch like this.... will everything go live at 1201 EST? (by everything I mean reviews with NDA's and the actual chips going on sale)

 

[Ventanni]

That is not exactly a great result for Kaveri because a GDDR5 HD 7750 variant will have much higher gaming performance than the more bandwidth-constrained Kaveri GPU. So the perf. per watt of the i3 + discrete card would actually be quite a bit better for gaming purposes.
.

I agree. It's still a nice improvement over the 5800k though.

 

[dosmastr]

Don't know what you are complaining about. I took your original sentence



as saying that while "Iris 5200 Pro" (meaning the CPU I guess) might draw 85W it doesn't have to dissipate that much heat. That is 85W =! 85W which is not possible. Maybe your sentence wasn't that clear.

So when you said that Iris Pro may draw 85W but that doesn't meant it dissipates that much heat, it struck me as counter-intuitive like asking if the element in one kind of electric heater is more efficient than another; since to me power in = power out.

So if a laptop draws 85W it has to dissipate 85W of heat; simple as. Now which fraction of that 85W is CPU, which is the screen or the power brick's efficiency etc. doesn't make any difference as to how much heat is generated because everything that goes in out has to go out as heat (a few photons excepted).

If I understand this....misunderstanding correctly, he is saying not 100% of power in is turned into heat.

saying 85 watts in = 85 watts out is like saying my car turns gas into heat... well yes it does but turns some of it into kinetic motion too, to go down the road.

likewise if it turned all 85watts to heat... none would be used to actually do the work of the processor.

Why would everything that goes in have to come out as heat?
That looks like a false premise to me, though I'm no electronic engineer.

 

[ElFenix]

never stayed up for a launch like this.... will everything go live at 1201 EST? (by everything I mean reviews with NDA's and the actual chips going on sale)

per the slide on the previous page the embargo goes through 8 AM EST tomorrow.

http://forums.anandtech.com/showpost.php?p=35946627&postcount=712

 

[dosmastr]

gratis fenix, so many slides.... the all blur together at this point

 

[ElFenix]

If I understand this....misunderstanding correctly, he is saying not 100% of power in is turned into heat.

saying 85 watts in = 85 watts out is like saying my car turns gas into heat... well yes it does but turns some of it into kinetic motion too, to go down the road.

well, conservation of energy says that turns into heat too due to friction/drag/braking/whatnot.

so, yes, 100% gets turned into heat. whether it is turned into heat doing something useful or turned into heat through leakage is a different question. but that doesn't have any bearing on whether that power has to be dissipated or not. it all has to be dissipated.

 

[JDG1980]

per the slide on the previous page the embargo goes through 8 AM EST tomorrow.

http://forums.anandtech.com/showpost.php?p=35946627&postcount=712

Does that 8 AM time apply just to those slides, or to any information/reviews about the chip?

 

[dosmastr]

I sit corrected, I thought if it was consumed doing the work the processor does it ...got consumed doing that and not heat... again, I now know better.

think the pricing will be boring (same price as an intel part with +/- same performance) or will they price couple parts to spice things up?

wondering if I'll regret buying a lga1150 board earlier today.

 

[ElFenix]

Does that 8 AM time apply just to those slides, or to any information/reviews about the chip?

well, i don't know why you wouldn't write everything to have the same embargo time. much of the information that'll be presented in reviews is going to have come from that conference, i would expect.