AMD CPU speculation... and expert conjecture

[griptwister]

AMD should have invested in having their own fabs when they had the chance...

AMD needs to be like, "We've got a 12 core excavator chip on 14nm process running @4.5 GHz coming to our FM3 Platform which will be supporting DDR4 CUZ YOLO!"

*edit* On another note, I've got a FX 8350 coming my way anytime next week! So excited! It looks like this chip won the silicon lottery too...

 

[de5_Roy]

Exactly. Since first day (since AMD adquisition of ATI) AMD concept of APU has been that of a heterogeneous compute processor, where the GPU is used as a giant FPU

i meant kaveri's higher integer processing power since everyone was arguing GFLOPS. the gpu-being used as a fp co-processor is dependent on various conditions. the gpu-as-a-fp-coprocessor is a big advantage only when those conditions are met.

Kaveri is the first APU that fulfills that long dream and however some people continue believing that an APU is something for cheap gaming or cheap graphics. LOL

kaveri is aimed at entry level market, which tends to be cheap and benefits from integrated solutions. it's not some people's belief, it's amd's way of positioning apus.
from what i see, kaveri has a rather large price range option from $150 to $180. i call it large because $150 is near the top of the low-end while $180 is upper-midrange. if kaveri sells at $180, it'll be considered more as a cpu than an apu. then if it can't outperform a core i5, it will eventually come down to $150 then settle at $140 (provided that amd clears trinity and richland stock and not repeats llano debacle). that's suited for cheap gaming. the gfx is most definitely cheap - 512 shaders sans gddr5, dependent only on the soc buses and imc, amd cannot price it high.
cheap is not a derogatory designation. in kaveri's case, that means customers may have a winner like 6800k or 5800k. people who think cheap is derogatory are deluding themselves with unrealistic, non-existent grandeur.

Spoiler

Well, I do believe the lower GFLOPs are due to Bulk not reaching the same clocks as FD-SOI. The formula uses the speed for each component (CPU and GPU) to calculate the FLOPs, so it's kind of obvious that the process DOES matter in terms of how APU behaves now. Indirect effect, but it still sucks.

And in those same lines, not giving official Turbo speeds is really something to be wary of. They can't back out of the 4GHz mark after Richland broke it so easily. I am doubting the initial batch of Kaveri parts will be as good as Richland when OCing and hence will have lower performance after tweaking. I wonder how the GCN part will stack against VLIW4.

Cheers!

EDIT: Looks like HSA software will save Kaveri. If it doesn't, then Kaveri starts to smell like a flop.

i looked at amd's GFLOPS calculation again. they consider 3.7 ghz for cpu and 700 mhz for igpu. amd's GFLOPS calculation hold true iff these are the nominal base clockrates (not the .8-1 ghz idle clockrate) that kaveri can certainly hit when both cpu and igpu are loaded. in real situations, the apu shares it's thermal budget among all of it's components and will very likely be outside a controlled lab environment. software performance aside, thermals will decide whether the apu can maintain those clockrates under load or throttle. i read that both trinity and richland throttle when loaded and those are s.o.i..
i've been saying this for a while, without hsa, kaveri is just a regular apu like richland but faster with higher integer performance and more powerful igpu. as long as amd doesn't !@#$ up igpu driver support, kaveri should be fine for everyone even without hsa. without hsa, kaveri would face much stiffer competition from broadwell since intel isn't sitting still (short of making a proper gpu).
i mean, how many entry level customers are gonna seek out installing hsa runtimes and other drivers unless amd bundles them inside the box(and that those files just work)? even majority of pc/apu buyers won't bother.

which is why I was talking wether or not it gets adopted. If HSA fails, Kaveri fails.

fails? on what? at entry level, it'd be hard to fail, since the competition only offers $160+ dual core cpu with annoyingly underpowered igpu.

Spoiler

first off, there is no proof of clock speed.

2nd, this "benchmark" has been increasing constantly.

now its 3311 integer and 653 fp.

http://www.cosmologyathome.org/show_host_detail.php?hostid=187215

there is no reported clock speed, and seeing as this started out at 2544 integer score, what are the chances that this isn't always running at 1.8 ghz?

lolwhat? i guess i can't use it for basing my speculations anymore. but most of my reasonings stand.

edit:
AMD’s GloFo deal could lead to inventory issues
http://www.fudzilla.com/home/item/33169-amd%E2%80%99s-glofo-deal-could-lead-to-inventory-issues

 

[juanrga]

Here's the issue Juan, in the one benchmark where Kaveri beats PD, they underclocked to 1.8. Which sounds good, until you factor in what happens when a SINGLE PD core gets over 100% work. Here's a hint: You take a giant performance hit. I'm guessing that's what happened here; basically, PD's design ended up bottlenecking when underclocked. If Kaveri were OC'd, rather the PD UC'd, then I'd expect the opposite (and expected) result of PD thrashing Kaveri.

So as always, you need to qualify your statement, to "Kaveri beats PD when PD is underclocked to half its normal speed", also noting how the PD arch is not designed to run at those low frequencies. Or did we suddenly decide to get away from stock v stock testing when it became convenient? We're talking performance, not architectural efficiency, right?

The three architectures, BD, PD, and SR, are high-frequency designs (~4GHz). Therefore your argument doesn't apply.

In fact, it is even possible that it is just the contrary and that Kaveri underperformed @ 1.8GHz because SR is more optimized for feeding the cores at high-frequencies with larger caches, better front-end, and other architectural improvements over PD.

Guess what? Because AMD is leaving it to developers to change they way they do things, an APU is going to be something for cheap gaming.

Nope. I already explained which has always been the goal of an APU. Moreover, Berlin is an APU and it is not aimed to cheap gaming. AMD workstation APUs are not for cheap gaming...

 

[juanrga]

At the same time AMD showed iGPU running BF4 at least 2x faster than GT630, which surprised more than one, because they were expecting less performance.

I would sure as heck hope so.

http://www.geforce.com/hardware/desktop-gpus/geforce-gt-630-oem/specifications

192 cores @ 875 vs 512 @ 700 ... I would hope that kaveri can be faster ...

Who in their right mind in any shape or form would expect less performance out of ~2.5x as much hardware? Thats a dumb statement in itself.

The only dumb here is your belief that you can compare cores of so different architectures with trivial "192 cores @ 875 vs 512 @ 700"

 

[juanrga]

Well, I do believe the lower GFLOPs are due to Bulk not reaching the same clocks as FD-SOI. The formula uses the speed for each component (CPU and GPU) to calculate the FLOPs, so it's kind of obvious that the process DOES matter in terms of how APU behaves now. Indirect effect, but it still sucks.

And in those same lines, not giving official Turbo speeds is really something to be wary of. They can't back out of the 4GHz mark after Richland broke it so easily. I am doubting the initial batch of Kaveri parts will be as good as Richland when OCing and hence will have lower performance after tweaking. I wonder how the GCN part will stack against VLIW4.

I already explained before that only the GDDR5 version of the HD7750 has cores @ 900MHz, the DDR3 version is 800MHz. Both are bulk, the lower clocks are due to memory bandwidth.

The lower iGPU clocks n Kaveri are consequence of the drop of GDDR5 support. I already mentioned that someone said me that Kaveri die has a disabled GDDR5 IMC. The decision to drop GDDR5 was taken very late.

Add to that that Kaveri has lower TDP and power consumption requirements than Trinity/Richland, because is more focused to mobile.

Kaveri on bulk achieves about the same base clocks than Trinity on SOI. Richland achieves higher clocks thanks to a more mature process. Kaveri is the first product made by Glofo in the new node.

About turbos, I have given above a link to an article that claims that AMD has not still decided the turbo frequencies.

 

[juanrga]

As leaked benchmarks show, the Kaveri CPU will outperform a 8350 in integer workloads with software up to four threads

first off, there is no proof of clock speed.

2nd, this "benchmark" has been increasing constantly.

now its 3311 integer and 653 fp.

http://www.cosmologyathome.org/show_host_detail.php?hostid=187215

there is no reported clock speed, and seeing as this started out at 2544 integer score, what are the chances that this isn't always running at 1.8 ghz?

I can see that correspond to 1.8GHz silicon, others also can. Those are benchmarks of ES, there is variability in the scores with each sample/batch/iteration. An ES is used to test and tune the hardware before the final silicon is produced. Ok?

In fact the chinese site linked before clearly mentions that if AMD continue tuning the hardware they could compete with Haswell. Time will say.

 

[juanrga]

How about the 8% they missed the CPU target and the 24% they missed the GPU target?

They didn't hit the CPU target because of clockspeed, and they didn't hit the GPU target because of clockspeed.

Both of those factors correlate to several things:

Target TDP limits (influenced by bulk vs. FD-SOI)
Target power consumption limits (influenced by bulk vs. FD-SOI)
Clockspeed headroom (influenced by bulk vs. FD-SOI)

The issue is, bulk substrate means higher leakage, which means more power is required to reach the same clockspeed that can be achieved on FD-SOI because of the insulator.

This means, basically, that while Kaveri runs at 3.7 GHz at 95W TDP, Richland runs at 4.1 GHz + Turbo at 100W TDP or 10% faster on a less advanced process with larger/fewer transistors.

Now, your 20% performance gain just took a 10% hit because of loss of clockspeed. See the difference?

High leakage in bulk substrate is why Intel uses so many tricks to make FinFET work on bulk. Like 3D transistors, and all the other super expensive to R&D stuff they use in their fab just to keep clockspeeds competitive on bulk process.

Additionally, bulk is far more temperature sensitive to power consumption at lower voltages than FD-SOI. If you crank the vcore up on a haswell chip, it gets hot fast. You can turn up the vcore on a FX chip and it doesn't require nearly as drastic cooling to run cooler than the Intel option.

These are all factors that you have not addressed and lead directly into the issues AMD had with bulk process.

They are also why I believe my contact mentioned directly that they would not pursue HEDT on bulk substrate, and would not be doing bulk beyond 28nm.

In the end, if AMD keeps this up, we all lose. FinFET is not the way of the future, it's many times more costly to develop a FinFET on bulk with all the technology Intel has, to compete, rather than take a simple planar UTBB FD-SOI that competes without all the dog and pony show tricks to get there.

But what do I know about CPUs? You clearly are the master of everything about CPUs, if nothing else because you said so, and no one else could possibly be right about something.

FX > Kaveri

i)
They didn't miss 8% of CPU target, because AMD never revealed the frequencies. In my BSN* article I included a table with three different combinations of frequencies. Final Kaveri CPU freq. is only a 3% behind one of them.

The same about the GPU frequencies. AMD never gave them.

ii)
I already explained you that only the GDDR5 version of the HD7750 has 900MHz clocks. The version with DDR3 has 800MHz clocks. Kaveri has 720MHz. All them bulk.

The large drop in iGPU performance from original expectations is explained by AMD dropping the GDDR5 support. The original 1050GFLOP Kaveri was considering GDDR5 memory. BSN* has copies of the original docs reporting GDDR5 support. I already mentioned that someone has said me that the Kaveri die has a disabled GDDR5 IMC. Which means that the decision to abandon GDDR5 was taken very late.

iii)
Trinity and Richland are 100W. Kaveri is limited to 95W TDP. Limit both Trinity and Richland to 95W max and say me about how many you drop frequencies.

My bet is that a 95W rated Trinity would be ~3.7GHz and Richland ~3.9--4.0GHz

iv)
Kaveri has larger and powerful iGPU than Trinity/Richland. Those 100W TDP for the APU are distributed between CPU and GPU.

Take Trinity/Richland and substitute their iGPU with a Kaveri-level iGPU. If the total TDP continue being of 100W, you have to cut the CPU frequencies by some extension.

v)
Sure that Kaveri would achieve higher clocks in SOI. Has someone said the contrary here? NO But are you aware that Glofo has not ready 28nm SOI? AMD decision to go bulk has been a success. I don't understand why you guys don't accept it.

vi)
We don't know turbo frequencies nor its aggressivity. What if AMD follows Intel way of lower base clocks and higher turbo clocks?

The last rumor is that AMD is considering >4.0GHz turbo.

vii)
In my BSN* article I didn't considered a 20% IPC gain over PD, but a 20% minus a 5% for unknows. Palying safety, this means that my predictions of i5 level of performance are off by about a 3%.

Now, benchmarks leaked after I wrote my article say that SR is a 31% faster than PD and speculate about higher gains in the final silicon

It is still too soon. There are so many unknows.

viii)
It is evident than FinFET is the future. AMD is already tapping out chips on 14nm bulk FinFET and some foundries are running the first tests of 10nm FinFET.

But hey you are the expert that said us that Kaveri was SOI...

 

[gamerk316]

Juan is going to be VERY disappointed. Its almost going to be fun watching him explain this away. I mean, seriously, I foresee him getting run out of here faster then JF-AMD was post BD...

 

[juanrga]

Kaveri is the first APU that fulfills that long dream and however some people continue believing that an APU is something for cheap gaming or cheap graphics. LOL

kaveri is aimed at entry level market, which tends to be cheap and benefits from integrated solutions. it's not some people's belief, it's amd's way of positioning apus.

Kaveri is not entry level APU. This goal is left for Kabini/Beema.

AMD goals with APUs are known since the acquisition of ATI, and never were "cheap gaming". Berlin APU for servers is not for "cheap gaming". The FirePro APU for workstations are not for "cheap gaming".

AMD has been selling some APUs for cheap gaming whereas cannot develop completely its original strategy for APUs.

The final goal is a ultra-high-performance APU. Nvidia has the same plans (and is preparing an ultra-high-performance SoC) and Intel moving now the Phi from PCIe3 to socket is pursuing something similar to compete.

 

[juanrga]

Juan is going to be VERY disappointed. Its almost going to be fun watching him explain this away. I mean, seriously, I foresee him getting run out of here faster then JF-AMD was post BD...

You are getting it in the wrong way. It is the people who expected SR 10-core CPU @ 4.5GHz, FD-SOI, AM4, FX refresh... who are very disappointed and hungry.

All my predictions have been verified, except the drop from 1050GFLOP due to lower GPU clocks (obligated by the drop of GDDR5 support).

However, MANTLE, unified memory addressing in the 290/290X, and other stuff is compensating for that.

 

[szatkus]

Juan is going to be VERY disappointed. Its almost going to be fun watching him explain this away. I mean, seriously, I foresee him getting run out of here faster then JF-AMD was post BD...

+1

 

[gamerk316]

Juan is going to be VERY disappointed. Its almost going to be fun watching him explain this away. I mean, seriously, I foresee him getting run out of here faster then JF-AMD was post BD...

You are getting it in the wrong way. It is the people who expected SR 10-core CPU @ 4.5GHz, FD-SOI, AM4, FX refresh... who are very disappointed and hungry.

All my predictions have been verified, except the drop from 1050GFLOP due to lower GPU clocks (obligated by the drop of GDDR5 support).

However, MANTLE, unified memory addressing in the 290/290X, and other stuff is compensating for that.

I've made my views clear on MANTLE. And Unified Addressing, while nice, still doesn't remove the physical bottleneck of having to transfer data over the bus to the GPU/APU, so there won't be a lot of performance gain there (just makes the code a lot easier to manage). Not shocked about no GDDR5; I predicted it due to cost, remember?

At the end of the day, my conclusions pre-BD are spot on: The BD arch couldn't clock enough to catch Intel (power/heat still limiting factor), and AMD moved to an all-in-one solution with a weak CPU/Medium GPU setup, which is where I advocated they go about 5 years ago.

 

[de5_Roy]

edit: Sony’s PlayStation 4 Costs $381 to Build — Only $18 Under Retail Price — in Teardown
http://allthingsd.com/20131119/teardown-shows-sonys-playstation-4-costs-381-to-build/

Kaveri is not entry level APU. This goal is left for Kabini/Beema.

if kaveri's not entry level, then what is it? do you think intel won't put price-pressure on kaveri? competitive market positioning decides entry level or mid/high end.
kabini and beema are entry level as well as low power/ulp.
you are giving off the impression of being entirely oblivious about kaveri's predecessor's market positioning, including zacate's.
if you have some kind of credible information/speculation about kaveri's pricing, please share.

AMD goals with APUs are known since the acquisition of ATI, and never were "cheap gaming".
Berlin APU for servers is not for "cheap gaming". The FirePro APU for workstations are not for "cheap gaming".

oh the delusion!:lol: you know that berlin is for servers and firepro apus are for pro tasks, right? simply put, those are sold differently from the mainstream desktop skus. desktop apus don't have validation process, optimized drivers, certifications, reliability check and customer service that server and pro products deliver.

AMD has been selling some APUs for cheap gaming whereas cannot develop completely its original strategy for APUs.

amd is selling All of it's desktop apus at entry level segment. desktop is around half of amd's shipments, other half is laptops and others. even mobile apus undercut intel's afaik.

edit:
kaveri can be high end, upper-mid is an easier reach. conditions are:
*intel cannot compete with kaveri's perf/price.
*intel doesn't [strike]bribe[/strike] subsidize pc oems' advertising funds.
*intel doesn't slash core i3 and pentium prices while bringing up silvermont dt skus. i am quite certain those won't be able to compete with kaveri in terms of performance but bulk price is different matter.
*amd chokes market supply just enough (or glofo fails (as usual) with early yields) so that retailers charge high price. since kaveri's new, it'll fly off the shelf for a few months until stabilization.
*hsa is widely and easily available to every pc customer and works out of the box with zero intervention.
*tablets, smartphone sales take a nosedive and majority of people start buying pcs again.

 

[gamerk316]

edit: Sony’s PlayStation 4 Costs $381 to Build — Only $18 Under Retail Price — in Teardown
http://allthingsd.com/20131119/teardown-shows-sonys-pla...

Which means AMD is making next to nothing per part sold; they're basically selling to Sony just above cost. So a significant portion of AMD's limited capacity is being used to make a chip they aren't going to make any money off of, like I suspected months ago.

 

[logainofhades]

Console makers often sell the consoles at a loss. They make up for this in games and accessory sales. I don't see how what Sony is doing makes any difference on how much AMD is making per chip.

 

[gamerk316]

Console makers often sell the consoles at a loss. They make up for this in games and accessory sales. I don't see how what Sony is doing makes any difference on how much AMD is making per chip.

Because this go around, Sony is allegedly making a profit per console sold. And if the BOM is only ~$20 cheaper then the retail price, then it follows that each component is being brought at basically break-even value. AMD probably makes the most profit per part, but I can't imagine its in excess of $10 or so.

In other words: Great for market share, not so much for profit. Especially if this locks out production of consumer chips with higher margins.

 

[de5_Roy]

Which means AMD is making next to nothing per part sold; they're basically selling to Sony just above cost. So a significant portion of AMD's limited capacity is being used to make a chip they aren't going to make any money off of, like I suspected months ago.

afaik, it is designed that way. i see that sony is selling it near-loss, but amd still should make some dough selling the apu. besides, the whole thing is high volume, sales are supposed offset the low profit. console life cycle is like 7-8 years, sony stated that they expect 10 yrs out of ps4. i am more hopeful about some of those defective apus end up being sold by amd. fingers crossed!! 😀

 

[gamerk316]

Meanwhile, over at Intel:

Nothing too surprising, though its sounding like the focus is on the IGP again. Lets see if Intel closes the gap a little bit more...

 

[de5_Roy]

^^ saw that earlier, seems like kabini won't be threatened. mullins should be able to outperform it by clockrate alone. and it's not coming out till later next year.
the socs specs seem like silvermont cores are shrunk to make room for broadwell-class execution units. the "key improvements" stating "new process node" made me chuckle. i was expecting something like "20% more application performance" or something similar.

 

[gamerk316]

Which means AMD is making next to nothing per part sold; they're basically selling to Sony just above cost. So a significant portion of AMD's limited capacity is being used to make a chip they aren't going to make any money off of, like I suspected months ago.

afaik, it is designed that way. i see that sony is selling it near-loss, but amd still should make some dough selling the apu. besides, the whole thing is high volume, sales are supposed offset the low profit. console life cycle is like 7-8 years, sony stated that they expect 10 yrs out of ps4. i am more hopeful about some of those defective apus end up being sold by amd. fingers crossed!! 😀

To date, the PS3 has sold about 81 Million units worldwide, the 360 about 80. Lets say the PS4/XB1 sell about the same over their lifetimes, and sell 200 Million units. And lets assume AMD makes $10 a pop. That equates to about $2 Billion over 10 years, or about $200 Million of profit per year, if we disregard any change in profit margins (which probably will happen as yields improve).

Now, that sounds like a good thing, until you consider the fact that Desktop APU's likely yield significantly MORE profit per unit sold. In a down year, PC's sell about 400 Million PC's sold per year [though we're currently at about 350 Million]. Figure AMD APU's make it in about 5%, or 20 Million PCs, and they make $20 per unit sold. Do the math, and you see how much money is being left at the table, in this case, $400 Million, or DOUBLE what they would make on both consoles combined.

Obviously, we can't know for sure how much AMD is making selling these things for different platforms, but assuming margins on consoles are razor thin, at 5% market share and $20 margin, AMD would make almost double selling in PC's then they would selling to consoles.

Going farther, if the production currently tasked to the PS4/XB1 were re-tasked to making PC compatible APU's, the extra supply would allow AMD to adjust pricing (more supply leads to more price control) to sell the most units at the highest margins possible.

My issue isn't that AMD is making a nice steady revenue stream selling APU's to consoles, its that if they focus on the PC instead, they could make a heck of a lot more.

 

[iceclock]

sweet griptwister enjoy that good mamba . i cant wait till amd comes out with an intel killer chip if they dont ive lost faith in them. there was a time when amd was better than intel in the gaming area of things, i want that day to come back.

 

[de5_Roy]

Spoiler

To date, the PS3 has sold about 81 Million units worldwide, the 360 about 80. Lets say the PS4/XB1 sell about the same over their lifetimes, and sell 200 Million units. And lets assume AMD makes $10 a pop. That equates to about $2 Billion over 10 years, or about $200 Million of profit per year, if we disregard any change in profit margins (which probably will happen as yields improve).

Now, that sounds like a good thing, until you consider the fact that Desktop APU's likely yield significantly MORE profit per unit sold. In a down year, PC's sell about 400 Million PC's sold per year [though we're currently at about 350 Million]. Figure AMD APU's make it in about 5%, or 20 Million PCs, and they make $20 per unit sold. Do the math, and you see how much money is being left at the table, in this case, $400 Million, or DOUBLE what they would make on both consoles combined.

Obviously, we can't know for sure how much AMD is making selling these things for different platforms, but assuming margins on consoles are razor thin, at 5% market share and $20 margin, AMD would make almost double selling in PC's then they would selling to consoles.

Going farther, if the production currently tasked to the PS4/XB1 were re-tasked to making PC compatible APU's, the extra supply would allow AMD to adjust pricing (more supply leads to more price control) to sell the most units at the highest margins possible.

My issue isn't that AMD is making a nice steady revenue stream selling APU's to consoles, its that if they focus on the PC instead, they could make a heck of a lot more.

imho, consoles seem like amd's backup plan, should they drop into red during transition to ARM.
tsmc makes the console apus, on a mobile node, with bulk silicon. amd's high margin apus(trinity, llano, richland both dt and mobile) are made by glofo, so amd should be relatively safe. glofo's nodes have been using s.o.i. until kaveri (if bulk silicon is confirmed, i don't know yet.). tsmc makes amd's other high volume apus (low power/ulp mobile) for pc markets - think those are the ones being delayed e.g. kabini, temash.
edit:
ironically, high margin apus were being offset by high volume apus since llano. part of the problem was amd's management, market situation and rest was glofo's yields.

 

[gamerk316]

^^ saw that earlier, seems like kabini won't be threatened. mullins should be able to outperform it by clockrate alone. and it's not coming out till later next year.
the socs specs seem like silvermont cores are shrunk to make room for broadwell-class execution units. the "key improvements" stating "new process node" made me chuckle. i was expecting something like "20% more application performance" or something similar.

I expect more iGPU updates. That's about it though.

 

[de5_Roy]

I expect more iGPU updates. That's about it though.

the rather scary part about cherry trail is that with 16 "gen 8" eus and 4 cores, this is a 4 core soc with ivb-class igpu. it can potentially rout amd, nvidia's low end gpu lineup including low-binned apus from both mobile and desktop, and shrink the low end gfx market to only to cater to legacy machines. yeah, amd apus offer discreet class performance at entry level price and stuff, but intel has higher marketshare and far longer reach.

 

[gamerk316]

Which is their plan. Like AMD, they are moving to a single unit type design, but with stronger CPU's rather then GPU's. In a few refreshes, the iGPU could really make some headway though...