Why have AMD APUs failed on the market?

[DrMrLordX]

Aida64 5.00.3300:

CPU:
Queen - 21632
PhotoWorxx - 8419 MPixel/s
Zlib - 183.6 MB/s
AES - 9192 MB/s
Hash - 2186 MB/s

FPU:
VP8 - 3440
Julia - 7071
Mandel - 3615
SinJulia - 1481

Cinebench 11.5:

CPU - 3.52
OpenGL - 36.17FPS

Okay, I wanted to compare my CPU to a 5 ghz 6800k, but obviously that wasn't going to happen here. So I took the approximate ratio in clockspeed between my chip at max clocks (4.7 ghz) and a hypothetical 6800k at max clocks (5 ghz), and then applied that ratio to the stock clockspeeds of the 6800k. That should at least give me some idea of the performance delta between my 7700k and a 6800k at max clocks.

What I wound up with was my 7700k @ 3.9 ghz (1.3125v vcore, more than it needs), no turbo, 2 ghz NB (I stepped it back from 2100 mhz), 1028 mhz iGPU, DDR3-2400 10-12-13-32 2T. The lack of turbo sort of puts me behind, I guess, though whatever. Here's what I got with that config in Win10 9926:

AIDA64 5.00.33.00
CPU Queen 19159
CPU PhotoWorxx 10054
CPU ZLib 178.5
CPU AES 8670
CPU Hash 2737
FPU VP8 3839
FPU Julia 6591
FPU Mandel 3476
FPU SinJulia 1520

Cinebench R11.5
CPU 3.66
OpenGL 47.34

Bear in mind that I have observed some oddball differences in results between Win7 and Win10 on this machine, though Win10's performance has matured a good bit with the latest build.

Regardless, they trade blows in the AIDA64 tests, but I think you'll find that the 7700k definitely shows its power in Cinebench or in a wide variety of other benchmarks (y-cruncher is a big one for Steamroller, for example) at this clockspeed ratio. Plenty of people would cry foul over my choice to OC the 7700k, but, fact is, I didn't buy it to run it at stock settings. I wanted to (among other things) confidently best older AMD APUs pushed to their limits, and I think Kaveri can do just that, given sufficient effort. So, from the enthusiast point-of-view, I think Kaveri is a step up over Richland.

 

[cbn]

So basically, what they are supposed to do is cancel yet another project, take a loss on all their R&D expenses incurred in production of Bristol Ridge, and start all over again with a design that would come out maybe ~4 months later than what would have been Bristol Ridge's launch date?

No. Just, no.

We don't know how far Bristol Ridge has gone, but if the product doesn't fit the market then it would be better to cancel it than to put it in production and suffer more losses.

 

[cbn]

I don't know that I really agree with that. Intel's top end consumer desktop processor is the 4790k which is an absolute monster, and that's just for the LGA1150 socket. Let's not talk about "enthusiast" LGA2011. AMD would have a very difficult time challenging the 4790k. If Intel felt threatened in the least, they could start pushing code initiatives (if they were on the ball) to encourage more widespread adoption of AVX2, putting AMD even further behind.

If AMD cannot challenge Intel on absolute desktop CPU performance then IMO they need to focus stronger on value.

Regarding AVX2, Excavator will have that ---> http://en.wikipedia.org/wiki/Advanced_Vector_Extensions#AVX2 whereas current lower end Intel processors like Pentium don't. So Intel pushing AVX2 on the market should actually help AMD in the lower price ranges. (This assuming Intel keeps AVX2 away from 2016 Pentiums. Current Pentiums lack both AVX and AVX2).

 

[DrMrLordX]

We don't know how far Bristol Ridge has gone, but if the product doesn't fit the market then it would be better to cancel it than to put it in production and suffer more losses.

Depending on whom you ask, none of AMD's products "fit the market". AMD has an APU line, and they are obviously interested in continuing the line, hence Kaveri refresh followed by Bristol Ridge. Bristol Ridge fits the pattern of AMD's focus on APUs.

Furthermore, what do you think AMD's engineers are doing? Sitting on their butts watching p0rn? If they are in fact releasing any Excavator product (Bristol Ridge or otherwise) side-by-side with Summit Ridge, you can bet that everyone on the team is working their arse off to get it ready on time without major flaws. AMD's back is against the wall. They can not afford to spend 4+months of project time doing nothing.

If AMD cannot challenge Intel on absolute desktop CPU performance then IMO they need to focus stronger on value.

They are already doing that. Hell you can get an 860k for $64, or an A8-7600 for $89. The 8310 can be had for $100 or less which is an incredible value.

Regarding AVX2, Excavator will have that ---> http://en.wikipedia.org/wiki/Advanced_Vector_Extensions#AVX2 whereas current lower end Intel processors like Pentium don't. So Intel pushing AVX2 on the market should actually help AMD in the lower price ranges. (This assuming Intel keeps AVX2 away from 2016 Pentiums. Current Pentiums lack both AVX and AVX2).

So how useful will be AMD's implementation of AVX2? How many registers will they have per module in Excavator? Carrizo should answer that question soon enough. Steamroller has AVX, but the implementation is arguably not as good as that in Ivy Bridge or Sandy Bridge.

 

[jpiniero]

We don't know how far Bristol Ridge has gone, but if the product doesn't fit the market then it would be better to cancel it than to put it in production and suffer more losses.

OEMs want new products though. AMD almost has to release something new in 2016. Presumably if Zen does come out it will be for laptops only, so spending the money to finish the Excevator for socketed desktops makes sense.

 

[cbn]

They are already doing that. Hell you can get an 860k for $64, or an A8-7600 for $89. The 8310 can be had for $100 or less which is an incredible value.

On that A8-7600, I think the price is too high for a quad core with iGPU. Future APUs should be priced lower and the volume higher. This, if anything to get better deals with OEMs.

Furthermore, the cost to manufacture also matters. If the die is too large relative to the selling price, I'd imagine things won't be so great for the company.

 

[cbn]

OEMs want new products though. AMD almost has to release something new in 2016. Presumably if Zen does come out it will be for laptops only, so spending the money to finish the Excevator for socketed desktops makes sense.

http://forums.anandtech.com/showpost.php?p=37181034&postcount=697

 

[DrMrLordX]

OEMs want new products though. AMD almost has to release something new in 2016. Presumably if Zen does come out it will be for laptops only, so spending the money to finish the Excevator for socketed desktops makes sense.

I would be shocked if Summit Ridge(Zen) winds up in laptops as an 8M/16T 95W chip. It could happen, and it would actually be pretty cool, but it would also be really weird.

But yeah Excavator APUs and Zen will be serving different market segments, and AMD would be goofy to just throw in the towel on their current line of APUs.

On that A8-7600, I think the price is too high for a quad core with iGPU. It should be lower and the volume higher. This, if anything to get better deals with OEMs.

Furthermore, the cost to manufacture also matters. If the die is too large relative to the selling price, I'd imagine things won't be so great for the company. (see Q4 2014 results thread)

$89 for a 7600 retail is an amazing value. You are getting a Kaveri quad with a slightly gimped R7-250 in one neat little package. I don't even know what OEMs have to pay to get them off a tray. I just don't think OEMs really understand where to stick chips like that (or really any other Kaveri) in their pricing hierarchy. It's faster than lots of other junk chips they shlep in AiOs and similar boxes so they sort of have to charge more for a Kaveri. Sort of.

On a (somewhat) related note, it would appear that AMD actually is still working on the transition of their cat cores to 20nm:

http://www.xtremesystems.org/forums/showthread.php?260770-Aida64&p=5245850&viewfull=1#post5245850

Hmm! Question is, will Nolan's Puma core wind up being significantly different than the Puma core in Beema/Mullins, or is it just being node-shrunk and renamed? And will that make Nolan the first volume-produced 20nm chip from GF?

 

[jpiniero]

http://forums.anandtech.com/showpost.php?p=37181034&postcount=697

I've mentioned this before, but I believe that the GPUs are "easier" than CPUs in getting yield out of it. Don't ask me why though. So a 3M part with a small IGP would be worse in yield than a 2M with a bigger IGP.

Presumably Bristol Ridge is really AMD finishing what will be released in 2015 if not for the budget cuts. It's probably not a different die or very minor changes from the upcoming laptop Carrizo.

 

[cbn]

Presumably Bristol Ridge is really AMD finishing what will be released in 2015 if not for the budget cuts. It's probably not a different die or very minor changes from the upcoming laptop Carrizo.

That rumor from Sweclockers had Bristol Ridge released Q3, 2016. EDIT: Did you mean to write "What would have been released in 2015 if if not for the budget cuts"

It is also different than Carrizo in that it has DDR4 and a southbridge chip. (Carrizo has DDR3 and the southbridge is integrated with the APU making it a SOC.)

 

[Spawne32]

The desperation for people to believe that there will be a meaningful CPU release for FM2+ before zen in this thread is sad. Just goes to show you how bad a position AMD is in if people are becoming this diluted about their future.

 

[NTMBK]

That rumor from Sweclockers had Bristol Ridge released Q3, 2016. EDIT: Did you mean to write "What would have been released in 2015 if if not for the budget cuts"

It is also different than Carrizo in that it has DDR4 and a southbridge chip. (Carrizo has DDR3 and the southbridge is integrated with the APU making it a SOC.)

http://techreport.com/news/26183/report-amd-next-gen-carrizo-apu-supports-ddr4-integrates-chipset

 

[Insert_Nickname]

So how useful will be AMD's implementation of AVX2? How many registers will they have per module in Excavator? Carrizo should answer that question soon enough. Steamroller has AVX, but the implementation is arguably not as good as that in Ivy Bridge or Sandy Bridge.

AMDs implementation of AVX honestly isn't that bad on Trinity/Richland/Kaveri. It just suffers from lack of execution hardware, because of the module design having only one FPU to two ALUs.

The exception is Jaguar, AVX instructions take two cycles to execute. Effectively halving frequency...

 

[MiddleOfTheRoad]

If AMD cannot challenge Intel on absolute desktop CPU performance then IMO they need to focus stronger on value.

AMD could challenge Intel on absolute desktop CPU performance if they had a fab partner that didn't totally suck. Global Foundaries is probably at least 3 generations behind Intel's fab capabilities (Intel had 2 generations at 22 nm and now has 14 nm in production).

Even ancient chips like Vishera or Thuban would be pretty strong if they were manufactured at 14 nm.

 

[NTMBK]

The exception is Jaguar, AVX instructions take two cycles to execute. Effectively halving frequency...

To be honest, I would argue that this is the best approach to take right now. AVX software won't be especially prevalent for quite a long time to come, and I suspect that when it is more widespread it will assume a baseline of AVX2 & FMA3- there's just too many useful instructions that AVX1 is missing. Given how little difference it will make to consumer workloads in the 5 year lifetime of Jaguar, a 256-bit vector unit would have been excessive.

 

[mrmt]

AMD could challenge Intel on absolute desktop CPU performance if they had a fab partner that didn't totally suck. Global Foundaries is probably at least 3 generations behind Intel's fab capabilities (Intel had 2 generations at 22 nm and now has 14 nm in production).

Even ancient chips like Vishera or Thuban would be pretty strong if they were manufactured at 14 nm.

Yeah, AMD market share ballooned against Sandy Bridge, and Bulldozer servers mopped the floor with SNB-E. Those were the days!

 

[Insert_Nickname]

To be honest, I would argue that this is the best approach to take right now. AVX software won't be especially prevalent for quite a long time to come, and I suspect that when it is more widespread it will assume a baseline of AVX2 & FMA3- there's just too many useful instructions that AVX1 is missing. Given how little difference it will make to consumer workloads in the 5 year lifetime of Jaguar, a 256-bit vector unit would have been excessive.

I agree. Its just something to keep in mind. You shouldn't be buying Jaguar for the AVX support. Its really a nice-to-have, not a need-to-have feature. Its a huge upgrade on Brazos though, were even SSE2 took two cycles to execute.

(As for Jaguar/Puma "special" features, you should be buying Jaguar for the cut-price AES support... )

 

[cbn]

If AMD cannot challenge Intel on absolute desktop CPU performance then IMO they need to focus stronger on value.

They are already doing that. Hell you can get an 860k for $64, or an A8-7600 for $89. The 8310 can be had for $100 or less which is an incredible value.

On that A8-7600, I think the price is too high for a quad core with iGPU. Future APUs should be priced lower and the volume higher. This, if anything to get better deals with OEMs.

Furthermore, the cost to manufacture also matters. If the die is too large relative to the selling price, I'd imagine things won't be so great for the company.

On the Athlon x 4 860K the price mentioned ($64 sold briefly Fry's) was the lowest sale price I have ever seen (typical sale price at Newegg or Amazon is usually $79.99, with some other retailers charging less but adding shipping). Ideally we could have some future quad core (possibly die harvested from a hexcore) on 28nm also come with a small iGPU for OEM usage. Unfortunately at the moment, AMD does not have such a value chip.

The FX-8310 price is also based on a sale (and only one retailer carries that chip, Tiger Direct), and it doesn't come with a cooler. (regular price at TD is $114.99).

While the sale and non sale prices are not bad on FX-8310, it is unfortunately available at just one retailer. Also I've had a hard time finding games that do a good job scaling beyond six vishera cores at the moment.* (the only exception was BF4 MP Beta at Game GPU and I think one other title I can't remember off the top of my head). Six AMD cores, on the other hand, appears to do quite well in more than one game. (see post #676 for links with some examples). Unfortunately FX-6300/FX-6350 won't drop into existing FM2+ boards and I question if Vishera will be around in Q3 2016 (probably won't).

So if releasing a construction core in Q3 2016, I would think 3M/hexcore has a good amount of CPU cores for a value chip. This with small iGPU so it could be made affordable for OEM. 3M/six core would also provide a tangible improvement for FM2+ owners who can only own up to a maximum of 2M/four cores.

Also regarding die size and the effect on the 28nm WSA, realize there will be other chips using that node as well. (According to rumor) AMD also brought 28nm dGPU silicon ("Fiji") over to GF this year. (Lets hope we also see a smaller 28nm dGPU die designed for GF. This especially as Nvidia is still using 40nm for some entry level dGPUs (GT610, GT620, GT 630, GT 730))

*FX-8350 usually does better in gaming benchmarks than FX-6300, but then it is also clocked higher.

 

[cbn]

http://techreport.com/news/26183/report-amd-next-gen-carrizo-apu-supports-ddr4-integrates-chipset

If Carrizo has already has DDR4, I see less reason for making another quad core (like Bristol Ridge) with 512sp/DDR4.

I guess switching from HDL to non HDL (for the excavator cpu cores) could be one reason. But then it also seems that folks would want Mini-ITX for this type of niche large iGPU APU chip and having the low power and integration of Carrizo would probably help with cost. Remember that FM2+ Mini-ITX boards are rather expensive (for whatever reason).

If releasing a new construction core chip for desktop in Q3 2016, might as well make it a bigger upgrade on the cpu front (and since we are talking conventional desktop it doesn't need the large iGPU).

 

[DrMrLordX]

AMDs implementation of AVX honestly isn't that bad on Trinity/Richland/Kaveri. It just suffers from lack of execution hardware, because of the module design having only one FPU to two ALUs.

The shortage of available registers is the main problem vs. Haswell (for example). Haswell's 32-bit fp performance in AVX2 code is off the chain.

So if releasing a construction core in Q3 2016, I would think 3M/hexcore has a good amount of CPU cores for a value chip.

Since AMD isn't making it, it could not be released by Q3 2016 if Bristol Ridge is any indicator of how long AMD expects to be working on a new desktop Excavator chip. Try Q4 2016 or Q1 2017 if starting today.

Also regarding die size and the effect on the 28nm WSA, realize there will be other chips using that node as well. (According to rumor) AMD also brought 28nm dGPU silicon ("Fiji") over to GF this year.

There is a reason why AMD is moving dGPU manufacture to Globalfoundries: GF 28nm is actually better for GCN cores and low-power, low-clockspeed CPUs. I would not expect much different from the 20nm process allegedly being put to use for the production of Nolan.

 

[cbn]

Since AMD isn't making it, it could not be released by Q3 2016 if Bristol Ridge is any indicator of how long AMD expects to be working on a new desktop Excavator chip. Try Q4 2016 or Q1 2017 if starting today.

We just heard about Bristol Ridge three weeks ago, and we don't even when the project was started. Does a die layout really take 2 years?

Also another thing to consider is that it is not the die layout that is the time bottleneck, but rather the redesign of Excavator without HDL that determines the release point. (This if your theory about Excavator being redesigned for desktop is correct).

 

[DrMrLordX]

We just heard about Bristol Ridge three weeks ago, and we don't even when the project was started. Does a die layout really take 2 years?

I'll let the process engineers (or those people who can hold themselves up as reasonable fascimiles thereof) answer that. But do consider that if AMD cooked up the idea of Bristol Ridge in January, you're still looking at ~19 months to design, test, fab, validate/bin, manufacture, and ship Bristol Ridge parts.

 

[Idontcare]

We just heard about Bristol Ridge three weeks ago, and we don't even when the project was started. Does a die layout really take 2 years?

It can take less, if you really want a stupidly done layout. But yes, most competent companies will take 2+ years.

 

[Abwx]

Automated design doesnt compress only the areas, it also accelerate the layouts devellopement phases, 10 years ago it would had been impossible to release so much iterations in so few time.

 

[svenge]

Most competent companies will take 2+ years.

That still doesn't answer how long it would take for AMD, though.