AMD Raven Ridge 'Zen APU' Thread

Page 3 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Sweepr

Diamond Member
May 12, 2006
5,148
1,142
131
Zero notebook will use it. Practically DDR4-2133 and DDR4-2400 SO-DIMM is available in mass for a reasonable price, in a year DDR4-2666 SO-DIMM might be feasible but that's it. And even this might be too high because Kabylake goes up to DDR4-2400, means demand for higher SO-DIMM DDR4 is limited.

I agree. That's what I meant, wasn't talking specifically about DDR4-3200.

Even dual-channel DDR4-2400 SO-DIMM would be 38.4 GB/s, a little bit more than a 1/3 Polaris 11 bandwidth. And it's sharing this amount (and the TDP) with 4 Zen cores.


If there really is a DDR4-3200 out of the box support, then it is more likely referred to desktop where such modules are cheap and available. Although I'm sceptical, it could be also the maximum with OC.

This could very well be the case. Fudzilla says ''DDR4 up to 3200 speed''.
 

Phynaz

Lifer
Mar 13, 2006
10,140
819
126
Dual Core Intel Core i5 Iris Graphics 6th Gen Skylake (Core i5 6360U 15W TDP and Core i5 6267U 28W TDP) have a recommended price of $305

http://ark.intel.com/products/family/88393/6th-Generation-Intel-Core-i5-Processors#@Mobile

And you actually believe that a Quad Core + HT 8x Threads ZEN with Polaris 10 Graphics paired with 4GB of HBM2 memory will not be able to sell for the same price ???

If AMD create an APU like that it will make those Dual Core Intel CPUs obsolete or force Intel to lower their price bellow $200.

You apparently think AMD has the same or similar brand value as Intel. You are wrong.
 

majord

Senior member
Jul 26, 2015
433
523
136
Agreed. Is DDR4-3200 even a JEDEC standard? Very few notebook OEMs will use it, let alone in dual-channel.

Notebook is another story, but the same theory holds true.

If they can even get DDR4-2400 into devices, combined with Polaris architecture would be a massive leap in effective bandwidth over the DDR3-1600/1866 we see today in Carrizo notebooks. with a corresponding leap in performance

That said, it's sad Apple already have 51Gbps at their disposal via their 128bit LPDDR4 interface - on a tablet of all things. Perhaps it's time to ditch this SO-DIMM rubbish on mobile?

The Cannonlake 'threat' is going to come down to the 10nm process, and what improvements they make to the GPU architecture.. which is anyone's guess at the moment. Iris Pro has proven to be a waste of time in practice, due to its perf/watt being essentially rubbish. Leaving them with the opposite problem to AMD - Plenty of effective BW, but unless you can open up TDP, it still performs poorly.

I also thought they had more headroom here after their emphasis on process + architecture efficiency gains (up to 2.8x) with Polaris. 704 SPs is 'only' 37.5% more than the current 512 SPs config had for years now. Let's not forget it will face Cannonlake in notebooks next year, rumoured to pack up to 40 EUs in GT2 config (up from 24 EUs with Skylake/Kaby Lake)



Well firstly, 2.8x was an extreme case based on worst perf/watt available desktop products. In APU's we're talking about moving from Tonga to Polaris, which as can be seen on desktop has a similar SP increase over its equivilent TDP predecessors. with the rest being made up by a 20+% increase in clockspeed.

I think you'll see the same play out in APU's, but with possibly even higher % clockspeed increases in the 15-35w class, for a combined perf/watt uplift in the expected 50-50% range.
 

24601

Golden Member
Jun 10, 2007
1,683
39
86
Notebook is another story, but the same theory holds true.

If they can even get DDR4-2400 into devices, combined with Polaris architecture would be a massive leap in effective bandwidth over the DDR3-1600/1866 we see today in Carrizo notebooks. with a corresponding leap in performance

That said, it's sad Apple already have 51Gbps at their disposal via their 128bit LPDDR4 interface - on a tablet of all things. Perhaps it's time to ditch this SO-DIMM rubbish on mobile?

The Cannonlake 'threat' is going to come down to the 10nm process, and what improvements they make to the GPU architecture.. which is anyone's guess at the moment. Iris Pro has proven to be a waste of time in practice, due to its perf/watt being essentially rubbish. Leaving them with the opposite problem to AMD - Plenty of effective BW, but unless you can open up TDP, it still performs poorly.





Well firstly, 2.8x was an extreme case based on worst perf/watt available desktop products. In APU's we're talking about moving from Tonga to Polaris, which as can be seen on desktop has a similar SP increase over its equivilent TDP predecessors. with the rest being made up by a 20+% increase in clockspeed.

I think you'll see the same play out in APU's, but with possibly even higher % clockspeed increases in the 15-35w class, for a combined perf/watt uplift in the expected 50-50% range.

AMD's main problem, even more primal than any others, is that their IMC is simply trash tier compared to Intel's so far. With Skylake, all bets are off, as that IMC scales linearly up to at least DDR4-4000. The current AMD APU's IMC scaling is absolutely laughable past DDR3-1333.
 

beginner99

Diamond Member
Jun 2, 2009
5,208
1,580
136
Dual Core Intel Core i5 Iris Graphics 6th Gen Skylake (Core i5 6360U 15W TDP and Core i5 6267U 28W TDP) have a recommended price of $305

http://ark.intel.com/products/family/88393/6th-Generation-Intel-Core-i5-Processors#@Mobile

And you actually believe that a Quad Core + HT 8x Threads ZEN with Polaris 10 Graphics paired with 4GB of HBM2 memory will not be able to sell for the same price ???

If AMD create an APU like that it will make those Dual Core Intel CPUs obsolete or force Intel to lower their price bellow $200.

Well the CPUs you mentioned are expensive because of the low power usage compared to their performance and are laptop CPUs. They target a completely different market than that APU would so comparing them doesn't make much sense.

Zen 8-core has a tdp of 95w, so naive we assume TDP of 45w for 4 core. RX 480 has a tdp of 150w (ignoring power issues). Even if you lower clocks a lot and use a cut die of P10, you will never get anywhere near the lower power usage required for laptop.

The APU you mentioned would be for desktop and there $300 is a hefty price simply because it also makes future upgrades more expensive. With current lack of progress in CPU performance, you can keep a top of the line CPU for 5 years easily and do GPU upgrade(s) in between. If you buy an APU with a slow CPU, you will need to upgrade both sooner. If you add a dGPU to the system it could become CPU bottlenecked.

Such an APU would make sense in an ideal world were most applications can make use of the iGPU and most games can do dx12 multi-gpu. But we are far away from that and not everything can run on the GPU. We still need fast CPUs
 

AtenRa

Lifer
Feb 2, 2009
14,000
3,357
136
Well the CPUs you mentioned are expensive because of the low power usage compared to their performance and are laptop CPUs. They target a completely different market than that APU would so comparing them doesn't make much sense.

Zen 8-core has a tdp of 95w, so naive we assume TDP of 45w for 4 core. RX 480 has a tdp of 150w (ignoring power issues). Even if you lower clocks a lot and use a cut die of P10, you will never get anywhere near the lower power usage required for laptop.

The APU you mentioned would be for desktop and there $300 is a hefty price simply because it also makes future upgrades more expensive. With current lack of progress in CPU performance, you can keep a top of the line CPU for 5 years easily and do GPU upgrade(s) in between. If you buy an APU with a slow CPU, you will need to upgrade both sooner. If you add a dGPU to the system it could become CPU bottlenecked.

Such an APU would make sense in an ideal world were most applications can make use of the iGPU and most games can do dx12 multi-gpu. But we are far away from that and not everything can run on the GPU. We still need fast CPUs

Are you really implying there can be no Quad Core ZEN APU at 15W TDP with 11CU Polaris iGPU at 14nm FF ??
 

redzo

Senior member
Nov 21, 2007
547
5
81
So essentially Raven Ridge will look a lot like Intel's current desktop 4C/8T quad-core offerings. It'll interesting to see if RR performs in the same league. But at least they could put some pressure against 2C/4T i3's.
If it performs as an intel 4c/8t(close to) is going to be priced near an intel 4c/8t. If it's priced as an intel i3 2c/4t I consider zen a total disappointment as it would mean its performance is half vs an intel core.
 

majord

Senior member
Jul 26, 2015
433
523
136
AMD's main problem, even more primal than any others, is that their IMC is simply trash tier compared to Intel's so far. With Skylake, all bets are off, as that IMC scales linearly up to at least DDR4-4000. The current AMD APU's IMC scaling is absolutely laughable past DDR3-1333.

IIRC this is restricted to Write performance. I think one would have to wait to see how the IMC on Zen performs , but it's a fair point it is weak.. performance scales for IGP well past that though.

Well the CPUs you mentioned are expensive because of the low power usage compared to their performance and are laptop CPUs. They target a completely different market than that APU would so comparing them doesn't make much sense.

Zen 8-core has a tdp of 95w, so naive we assume TDP of 45w for 4 core. RX 480 has a tdp of 150w (ignoring power issues). Even if you lower clocks a lot and use a cut die of P10, you will never get anywhere near the lower power usage required for laptop.

This makes no sense at all. Sorry. how else do you think mobile parts based on the same architecture as their desktop counterparts achieve their lower TDP? Magic?

In some cases there's different process variants for mobile, and obviously they're lower leakage bins, but the rest is all lower clock speeds. This is just "101" type stuff
 

beginner99

Diamond Member
Jun 2, 2009
5,208
1,580
136
Are you really implying there can be no Quad Core ZEN APU at 15W TDP with 11CU Polaris iGPU at 14nm FF ??

No I'm implying there can't be such an SKU with Polaris 10 as you mentioned in your original post. 11CU Polaris vs polaris 10 is obviously a huge difference.
 

AtenRa

Lifer
Feb 2, 2009
14,000
3,357
136
No I'm implying there can't be such an SKU with Polaris 10 as you mentioned in your original post. 11CU Polaris vs polaris 10 is obviously a huge difference.

Aaaa, just realized that. Well my bad, I meant to say Polaris architecture not the actual Polaris 10 chip.
 

Insert_Nickname

Diamond Member
May 6, 2012
4,971
1,691
136
If it performs as an intel 4c/8t(close to) is going to be priced near an intel 4c/8t. If it's priced as an intel i3 2c/4t I consider zen a total disappointment as it would mean its performance is half vs an intel core.

Well, its not a total loss if that's the case. It'll put some pressure on Intel to shift 4C/4T down to i3 territory.

Product stack could be;
2C/2T - Celeron
2C/4T - Pentium (note Intel is already doing this on certain mobile CPUs)
mix of 2C/4T for low-end (x1xx) i3's and 4C/4T high-end (x3xx) i3's.
4C/4T - i5 (essentially a high-end i3 with more features/better IGP, higher clocks)
4C/8T - i7 as usual.

Personally, based on limited testing of my Carizzo Athlon x4 845, I'm expecting somewhere around IB-level performance, perhaps a bit better then that in integer workloads.

Which isn't half-bad when you've got eight full cores with SMT.
 

cbn

Lifer
Mar 27, 2009
12,968
221
106
I think AMD is going to have a hard time convincing OEMs to use DC DDR4 3200.

DDR4 3200 should eventually be JEDEC. (In the same way DDR3 1600 and DDR2 800 were JEDEC).

With that mentioned, current 2 x 4GB SO-DIMMs (according to Newegg product listings) only go up to DDR4 2666:

http://www.newegg.com/Product/Produ...01&IsNodeId=1&bop=And&Order=PRICE&PageSize=30

While the 2 x 8GB SO-DIMMs go up to DDR4 3000:

http://www.newegg.com/Product/ProductList.aspx?Submit=ENE&N=100007609 601190333 600213067&IsNodeId=1
 

CriticalOne

Member
Apr 17, 2015
26
0
16
What AMD really needs to work on before concerning itself with L4$, EDRAM, or HBM2 on the CPU package/die is to upgrade the APU's IMC. The current one in Steamroller is terrible, with it getting around 1/2 the bandwidth that a similar Haswell CPU gets at the same frequencies.
 

cbn

Lifer
Mar 27, 2009
12,968
221
106
Some info on HBM2:

http://www.storagereview.com/samsung_begins_mass_production_of_hbm2_4gb_dram_package

Samsung indicates that they plan to produce an 8GB HBM2 DRAM package within this year, which will allow designers benefit from space savings by over 95 percent compared to GDDR5 DRAM. As a result, more optimal solutions for compact devices that require high-level graphics computing capabilities will be possible.

8GB per package is pretty good.

I just wonder how much those could be tuned for even lower power at the expense of some bandwidth.
 
Last edited:

itsmydamnation

Platinum Member
Feb 6, 2011
2,743
3,075
136
What AMD really needs to work on before concerning itself with L4$, EDRAM, or HBM2 on the CPU package/die is to upgrade the APU's IMC. The current one in Steamroller is terrible, with it getting around 1/2 the bandwidth that a similar Haswell CPU gets at the same frequencies.

How much bandwidth does a core need ( hint arronspink has posted some good detail on this over the years on realworldtech). Second how much bandwidth can the GPU consume across the IMC. Put those together and you will discover that this is a minor point at best. A bigger concern for excavator core in regards to serial code is with its low L2 no L3 is the memory access latency. Even if it stays around the same for Zen, with its much bigger L3 ( 2mb a core) it will be in a much better position.
 

cbn

Lifer
Mar 27, 2009
12,968
221
106
Unsurprisingly, no HBM mentioned.

Maybe if the next Zen APU (Vega iGPU) is able to boost performance per watt we could see HBM2 integrated into a mobile APU?

And actually I would just love if we could eventually have low profile desktop discrete cards with HBM2 (4GB, etc.) as well. These (as well as the mobile APU) integrated with additional hardware like TV tuner, cable tuner, FM radio tuner.

Remember All-in-Wonder?

Bring back the capability to enjoy multi-media without dependence on the internet. Increase integration and encourage Microsoft to add back features they removed from Windows (thinking of Media center). In other words, advance the standard capabilities of the PC.
 
Last edited:

looncraz

Senior member
Sep 12, 2011
722
1,651
136
I dont believe ZEN Raven Ridge with a small die (less than 200mm2) and only with 2-4MB of HBM2 memory (+ Interposer) is more expensive than Polaris 10 232mm2 + PCB + 8GB of 8GBPS memory.

Not only that, but that Raven Ridge (HBM2) could be sold at higher price than Polaris 10 Custom cards ($299)

If Zen 100% reaches Haswell's IPC - and clockspeeds - we can expect a quad core APU with half of RX 460's graphics performance to fetch only about ~$150 without SMT, maybe $200 with it (depending on its performance).

If AMD were to add HBM - a single stack would add ~$40 to the total cost... effectively without regard to capacity.

* Enlarged die to add HBM controller(s)
* Additional interposer (+handling)
* Custom APU package (to interface interposer to socket)
* HBM stack itself

Due to the need to simplify design and manufacturing, AMD would only use one die for all APUs - so every APU they make would be enlarged by the size of an unused HBM memory controller... just so they can have graphics performance only 33% worse than the RX 460 instead of 50% worse...

There just isn't a business case for it... on the desktop. A 'massive' integrated GPU with HBM, and multiple 8-core dies makes perfect sense for the server or HPC markets - and could more easily justify its development costs... once you have gone MCM, using an interposer might just make things easier overall, permitting a higher-performance inter-die communication bus.
 

looncraz

Senior member
Sep 12, 2011
722
1,651
136
AMD's main problem, even more primal than any others, is that their IMC is simply trash tier compared to Intel's so far. With Skylake, all bets are off, as that IMC scales linearly up to at least DDR4-4000. The current AMD APU's IMC scaling is absolutely laughable past DDR3-1333.

A weak CPU with a strong IMC can still show poor memory performance.

The construction cores' Achilles' heel is its cache subsystem, which is closely related to its ability to maximize available memory bandwidth.

Zen could give better results even with the same IMC - which it clearly will not be using.
 

AtenRa

Lifer
Feb 2, 2009
14,000
3,357
136
If Zen 100% reaches Haswell's IPC - and clockspeeds - we can expect a quad core APU with half of RX 460's graphics performance to fetch only about ~$150 without SMT, maybe $200 with it (depending on its performance).

If AMD were to add HBM - a single stack would add ~$40 to the total cost... effectively without regard to capacity.

* Enlarged die to add HBM controller(s)
* Additional interposer (+handling)
* Custom APU package (to interface interposer to socket)
* HBM stack itself

Due to the need to simplify design and manufacturing, AMD would only use one die for all APUs - so every APU they make would be enlarged by the size of an unused HBM memory controller... just so they can have graphics performance only 33% worse than the RX 460 instead of 50% worse...

There just isn't a business case for it... on the desktop. A 'massive' integrated GPU with HBM, and multiple 8-core dies makes perfect sense for the server or HPC markets - and could more easily justify its development costs... once you have gone MCM, using an interposer might just make things easier overall, permitting a higher-performance inter-die communication bus.

The thing is people only focus on the APU but dont see the rest of the Laptop.

If the APU has 50% LESS performance than the one with HBM2, that means the APU with HBM2 memory is 2X faster (1).

If you want to make a Laptop with 2X the graphics performance of ZEN + DDR-4 3200MHz, you will need the APU + a dGPU.
That will increase the BOM price because of the dGPU cost + GDDR-5 cost + Cooler (2), increase the power consumption (3), increase the heat output of the Laptop, increase the height because of the bulkier CPU+dGPU cooler (4), increase the BOM and complexity of the Laptop motherboard to facilitate the extra dGPU (5), decrease the Battery life(6).

Now tell me that if you were a Laptop manufacturer you wouldn't want to have an APU + HBM2 that will offer you all those above (1 through 6) at the same or lower cost + lower consumption + the ability to have smaller Battery + the ability to make thinner design ???

The problem for AMD is not the higher cost of the HBM2, the problem is to find a customer for that APU. AMD doesnt have the luxury to create this APU and wait for customers to use it, they need to have the deal before they will make the product due to their current financial state.

Also, i dont know how much more extra space the HBM2 controller will add to the die, in the past AMD APUs had dual memory controllers DDR-3 + GDDR-5 and DDR-3 + DDR-4. They could very well add another controller to the ZEN RR die and be ready to use HBM2 dies when the customer wants them or HBM2 price is lower for the mainstream market. Pure speculation on that part but it could be doable as 14nm FF will allow this and save them the cost of extra die design/masks etc etc.


Edit:
Also, SAMSUNG recently announced 8GB of HBM2 modules to be available by the end of the year if im not mistake.

A Dual Core + HT 4x Threads with 8x CU Polaris iGPU + 8GB HBM2 memory for both graphics and System memory at 4.5W TDP would be a killer product for Microsoft tablet based products like Surface 4 Pro. The will have exceptional high graphics performance for games AND professional 3D/ Open CL workloads and acceptable CPU performance with 8GB of HBM2 memory that will save lots of system space, complexity, BOM, decrease heat output, allow thinner cases etc etc.

Or a Quad Core ZEN APU with 8GB HBM2 for Laptops could easily allow the same savings by completely eliminating the usage of DDR-4 memory.
 
Last edited:

cbn

Lifer
Mar 27, 2009
12,968
221
106
The only problem is for AMD to find a buyer (Apple ???)

If I may be so bold....

If AMD wants to sell an APU to Apple, they first need to make it even more attractive to Microsoft and its OEMs.

Then Apple seeing the advantage the APU is giving to Microsoft and its OEMs will want it for themselves as well.

In my humble opinion this means adding features to the APU (in addition to HBM) that raise the creative and utilitarian level of the hardware.

What can a person do with this new laptop we are imagining that they can't do with one of the many surplus or closeout laptops (including gaming laptops) floating around out there?
 

NTMBK

Lifer
Nov 14, 2011
10,208
4,940
136
If I may be so bold....

If AMD wants to sell an APU to Apple, they first need to make it even more attractive to Microsoft and its OEMs.

Then Apple seeing the advantage the APU is giving to Microsoft and its OEMs will want it for themselves as well.

In my humble opinion this means adding features to the APU (in addition to HBM) that raise the creative and utilitarian level of the hardware.

What can a person do with this new laptop we are imagining that they can't do with one of the many surplus or closeout laptops (including gaming laptops) floating around out there?

Apple don't like to be followers, they like to be leaders. They're the ones who pushed Intel towards the whole "Ultrabook" mentality in the first place, and the ones who made Intel take integrated graphics seriously. (They're still some of the only people who ship U parts with GT3 graphics.)

Remember, Apple almost put AMD's Llano APUs into Macbooks, before GloFo production problems scuppered that.
 

cbn

Lifer
Mar 27, 2009
12,968
221
106
Apple don't like to be followers, they like to be leaders.

I think the hardware is the key to making things change.

For example, take FM radio.

Unlike Internet Radio it is legal to record FM radio for personal usage and Microsoft Media Center (in addition to being able to record DRM protected Cable channels and regular TV) has the ability to play FM radio provided the PC is equipped with a FM radio tuner.

So I think a good FM radio tuner should be one the new features added to the APU (which already has true audio).
 
Last edited:

itsmydamnation

Platinum Member
Feb 6, 2011
2,743
3,075
136
Consider the case of Broadwell C outperforming Skylake in games and some other workloads because of its 128 MB L4 victim cache.

That has nothing to do with throughput, that is a product of decreased latency. The L4 is approx 1/2 the latency of main memory. Games dont even use 256bit ops, simple test increase memory clock but also massively increase CAS latency, if your throughput limited you will still get a performance increase if your latency limited you will get a perf decrease.

Have a look here http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/3 you need to pull very large amounts of data (streaming etc) to really get a throughput benefit.

Even something as wide as a GCN "core" has only ~7gbps of throughput to main memory and its a massively wide vector unit.