AMD Raven Ridge 'Zen APU' Thread

Page 10 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Ancalagon44

Diamond Member
Feb 17, 2010
3,274
202
106
The only company that I think might be interested in an HBM equipped APU is Apple, and I don't think they will switch suppliers from Intel to AMD anytime soon. At least, not CPU suppliers.

Yes, an HBM equipped APU would be great for a laptop, where upgradeability is not as much of an issue. But the cost means that it would only be feasible for a high end laptop. I just can't see Samsung or Acer bringing out a high end AMD laptop.
 

ShintaiDK

Lifer
Apr 22, 2012
20,378
145
106
,Yet they ship slow ass e-dram part

EDRAM cost less than the interposer needed. And it doesn't have the production issues either;)

Wut? What disaster? ! Fiji could be had for as low as $300 recently.

Production is stopped. So its old stock. And look on card sales. Its low, very low.

But the point is to see what was the lowest production sales point. And that price was high.

Interposer is a static extra cost.
HBM chips are very expensive.
TSV is expensive.
When combining everything above with the GPU/APU, any failure=loss of all. That's expensive.
 
Last edited:
May 11, 2008
19,306
1,131
126
In light of the new lower cost HBM and since Raven Ridge is a H2 2017 product, i will say a 8GB HBM APU for Laptops is very possibly and viable solution.

For High-End, High Performance, Thin Chassis Segment ($800-1200+)

Raven Ridge with 8GB HBM

1. Single slot solution.
2. Simpler Motherboard design and lower BOM
3. No extra DDR-3/4 dimm slots or integrated Chips
4. Single cooler design, smaller, lighter, cheaper.
5. Higher or same Graphics performance vs dGPU (RX 460m GDDR-5) of the same class with lower power consumption.

1+2+3+4+5 =
Even if the HBM APU alone will be more expensive than a DDR-4 APU, the HBM APU will make the final product (Laptop) be cheaper, lighter, with same or faster performance, consume less power, increase Battery life, be thinner.

Intel doesnt have a product like that, it will make AMDs HBM APU look like heaven to OEM/ODMs.

thoughts ??

I think that will be possible because everybody who designs, loves everything in a single chip.
Because for selling a product, form is just as important as function is. And a laptop that is also a tablet and that is powerful with a detachable wireless keyboard is really something as these models already exist and people are fond of them. The wireless key board needs wireless charging, but that is something interesting to design in such a way that it is not extremely visible and preferably wireless. When the keyboard is attached to the screen/tablet laptop is closed, some small circle with a ferrite core beneath it acts to transmit power and data from tablet to keyboard. Like a combination of wireless charging of electrical toothbrushes and the Qi wireless charging. This way, when using the laptop/ tablet in tablet mode, the wireless keyboard can be charged as well with Qi chargers. An aux micro usb connector for charging would also be handy.

But a x86-x64 APU+HBM SOC would really help. And help for optimizing power. And since consumer HBM can be made on customer specs, it would really help if the HBM memory controller could also lower the clock for the dram dies depending on how much memory bandwidth would be needed. This would further reduce power consumption. The only thing needed to be replaced by a customer would be a 2.5 inch ssd or m2 mass storage device. Simpler design.
 

Glo.

Diamond Member
Apr 25, 2015
5,661
4,419
136
Production is stopped. So its old stock. And look on card sales. Its low, very low.
Yeah, stopped. And they are selling brand new S9300X2 GPUs and Pro Duo based on available supply?

Server and Professional Fiji chips are produced, without any change.

Consumer market sales have stopped. That is all. Because AMD readies replacement for it.
 
May 11, 2008
19,306
1,131
126
EDRAM cost less than the interposer needed. And it doesn't have the production issues either;)



Production is stopped. So its old stock. And look on card sales. Its low, very low.

But the point is to see what was the lowest production sales point. And that price was high.

Interposer is a static extra cost.
HBM chips are very expensive.
TSV is expensive.
When combining everything above with the GPU/APU, any failure=loss of all. That's expensive.


Although expensive, HBM (2 and newer variants) is a new technology with an emerging market. As soon as more manufacturers jump in, more solutions are provided for reliability increase and cost reduction.
Also, HBM2 is just on the market, give it some time to develop.

And good luck with making 8 or 16GB edram next to a gpu, a cpu, mass storage controller, and all other interconnect fabric on a single die. Or you want to go the multiple die route which also needs an interposer...
 

deasd

Senior member
Dec 31, 2013
513
724
136
I would look forward to results about BirstolRidge benching game when pairing DDR4 3000Mhz+, after that it would be much clear that RavenRidge is necessary to use HBM or not.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Raven Ridge-E / "Lumpy Ridge"

22FDX
15h NG2 Bulldozer Dual-core/Single Module
64B Gather/64B Scatter AGU/FPU Load/Store
Single AVX512 FlexFMAC FPU.
Ports A-H => 4x32-bit, 2x64-bit, 1x80-bit / Single Port = 128b/80b, Dual Port = 256b/160b, Quad-port = 512b/320b
Ports A-D => Mul, Divide, Square Root, Transcendentals, Dot/Trigonometry, etc
Ports E-H => Add, Packs, Permute, Shuffle, Rotate, Store, etc

3W and 10W SKUs positioned against Core M AVX512 4.5W and 15W SKUs.

Pure Embedded[IoT-Green HPC] and ULP Consumer only. Received information gathered via contract extension not under NDA.
 

Azuma Hazuki

Golden Member
Jun 18, 2012
1,532
866
131
"Lumpy Ridge?" And this looks something like a reiteration of Stoney Ridge on a process designed for high performance. Interesting idea...Intel's all but stopped putting much into its Celeron and Pentium versions of the Core chips.
 

SPBHM

Diamond Member
Sep 12, 2012
5,056
409
126
dual DDR4 3200 is still a decent amount behind stuff like the HD 7750,
I'm sure they can improve efficiency being based on Polaris/Vega but... it's still going to be half the bandwidth of the RX460, so you can't expect much from this thing, not even close to Xbox One level (with its 256bit 2133 DDR3 + fast cache)... but l3 and quad core is good news.
 

Dresdenboy

Golden Member
Jul 28, 2003
1,730
554
136
citavia.blog.de
Raven Ridge-E / "Lumpy Ridge"

22FDX
15h NG2 Bulldozer Dual-core/Single Module
64B Gather/64B Scatter AGU/FPU Load/Store
Single AVX512 FlexFMAC FPU.
Ports A-H => 4x32-bit, 2x64-bit, 1x80-bit / Single Port = 128b/80b, Dual Port = 256b/160b, Quad-port = 512b/320b
Ports A-D => Mul, Divide, Square Root, Transcendentals, Dot/Trigonometry, etc
Ports E-H => Add, Packs, Permute, Shuffle, Rotate, Store, etc

3W and 10W SKUs positioned against Core M AVX512 4.5W and 15W SKUs.

Pure Embedded[IoT-Green HPC] and ULP Consumer only. Received information gathered via contract extension not under NDA.
Why wouldn't they just combine this FPU with Zen and call it Zen+? ;)

This way it reminds me of MIC cores. Weak Int+fat SIMD.
 

ShintaiDK

Lifer
Apr 22, 2012
20,378
145
106
Although expensive, HBM (2 and newer variants) is a new technology with an emerging market. As soon as more manufacturers jump in, more solutions are provided for reliability increase and cost reduction.
Also, HBM2 is just on the market, give it some time to develop.

And good luck with making 8 or 16GB edram next to a gpu, a cpu, mass storage controller, and all other interconnect fabric on a single die. Or you want to go the multiple die route which also needs an interposer...

And when are we getting all those HBM2 products?
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Speculative: Was not included in above thing that they and them had received from future roadmap.
Why wouldn't they just combine this FPU with Zen and call it Zen+? ;)
Different design methodologies is my guess. Zen is aiming for speed and thus higher supported Fmax. This new unit is aiming for density and higher energy performance. Secondly, nothing that implies 3-cycle dependency latency over 5-cycle dependency latency.
This way it reminds me of MIC cores. Weak Int+fat SIMD.
To be determined/speculation of new integer core for 15h NG2 Bulldozer. This was not specified;

CALU = Complex ALU, AALU = Address ALU, LDDU = Load Data Unit, STDU = Store Data Unit

CALU = ALU ops + Fast LEA[/MOV] + MUL[0] or DIV[1] (2 of these)
AALU = ALU ops + LEA[/MOV] + Branch[2] or Count[3] (2 of these)
LDDU = 1 gather*/x loads per x-cycles[4] (1 of these)
STDU = 1 scatter*/x stores per x-cycles[5] (1 of these)
*AMD specific Insert(Gather/Indexed Load)/Extract(Scatter/Indexed Store) for GPR instruction set (like SSE4a[ABM] but with AVX512 Gather/Scatter) [Macro-op fusion of LEA/MOV + Insert + Clear Upper Half[64-bit to 32-bit MOV/LEA] for x86-64 would undercut what was explored with RISC-V: http://arxiv.org/abs/1607.02318 // https://riscv.org/wp-content/uploads/2016/07/Tue1130celio-fusion-finalV2.pdf]

Ton of PRF/Rename/Scheduler optimization. Energy efficiency & density is preferred over Fmax. 400 mV to 800 mV w/ sub-W power consumption to 2W power consumption.

Forgot about possible queues...
Decode(AMD64 -> Macro-op) -> Dispatch & Macro-op Queue[Absorbs Loop cache] <== L1 Op Queue
Scheduler(Macro-op -> Micro-op) -> Micro-op Queue <== L0 Op Queue [2 GPR/1 FPU]
Overall redundant but great for completely gating anything before the dispatch. Also, pseudo-branch prediction which is for loop iteration-branches.
 
Last edited:
Aug 11, 2008
10,451
642
126
Wut? What disaster? ! Fiji could be had for as low as $300 recently.
It got overwhelmed by 980Ti in pretty much every metric. Late to market, slower in most games, very poor overclocking, bad supply, and poor performance per dollar for an amd product. Yea, Fury x was a raging success.
 

Erenhardt

Diamond Member
Dec 1, 2012
3,251
105
101
It got overwhelmed by 980Ti in pretty much every metric. Late to market, slower in most games, very poor overclocking, bad supply, and poor performance per dollar for an amd product. Yea, Fury x was a raging success.
Any sources GDDR5 would make it a better product? Anyway, this is APU thread.
We all know he hates HBM from the very beginning, when the rumors popped up about hbm GPUs, just like he hates console's APUs.

His no HBM2 for APUs looks identical to previous claims about no HBM for GPUs and no APUs for consoles.
 
Aug 11, 2008
10,451
642
126
Any sources GDDR5 would make it a better product? Anyway, this is APU thread.
We all know he hates HBM from the very beginning, when the rumors popped up about hbm GPUs, just like he hates console's APUs.

His no HBM2 for APUs looks identical to previous claims about no HBM for GPUs and no APUs for consoles.
Never said it would be better without HBM, although only seems logical that it could have been brought out earlier and at a better price point without HBM. And the fact that it had HBM certainly did not give it better performance that 980Ti. But trying to defend the product because it has sold at 300.00 (BTW do you have a source for that?) seems absurd. The fact that it supposedly has sold at about half its suggested price is an indication of how uncompetitive it really was.
 
  • Like
Reactions: Sweepr

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
And good luck with making 8 or 16GB edram next to a gpu, a cpu, mass storage controller, and all other interconnect fabric on a single die. Or you want to go the multiple die route which also needs an interposer...

You're not going to put 8 or 16 GB of edram, that would be silly. Something similar to what Skylake Iris Pro has would be enough to get the job done for the most part. And it is still by far the only real solution for increasing bandwidth on Raven Ridge that would be affordable enough to get OEMs to bother. Even then it's still kind of iffy.

Raven Ridge really should be like at least 20-30% faster than Skylake Iris Pro in games that Intel bothers to optimize and 2X in games they don't.
 

ShintaiDK

Lifer
Apr 22, 2012
20,378
145
106
Any sources GDDR5 would make it a better product? Anyway, this is APU thread.
We all know he hates HBM from the very beginning, when the rumors popped up about hbm GPUs, just like he hates console's APUs.

His no HBM2 for APUs looks identical to previous claims about no HBM for GPUs and no APUs for consoles.

So easy with the hate word again are we when we dont get what we want? If you dont agree you are a hater, because then you dont have to argument for it. ;)

Its about saying everyone hates Ferraries because you dont accept the everyone wont drive one. Its a superior car and everyone should use it.

Why dont you deliver some substance and tell how and when HBM costs are coming down to a level that it can compete and deliver some numbers. Raja said HBM wasn't for mainstream. HBM makers says HBM isn't for mainstream in its current form. Nvidia shows what it think when they spend something like 100M$ on the GP102 IC design and fits it with GDDR5X. Samsung commits to GDDR6(GDDR5X).
 
Last edited:
  • Like
Reactions: Sweepr

hrga225

Member
Jan 15, 2016
81
6
11
Regarding HBM on APUs,I think you all are missing another view from different perspective:TAM.Cost of production is chump change if there is enough TAM for your product.
There was an article in which was guesstimated/calculated that you need to sell at least 10 mil. chips to break even on 14/16 FinFet design or 1st SKU with average industry margins for ARM SoCs.That number may be different to X86 SoC and in this point in time,since calculation was done early in process life,but for argument sake lets assume it is correct in this case also.
Since by time AMD introduces to market this APU,they will be selling SKUs which will be recuperating costs of designing on 14nm(Polaris,Summit Ridge,Vega),so with this SKU(one with HBM) AMD will need sell just 3 mil. units(completely arbitrary) to break even.

Now we are back to to beginning:TAM.
 

Glo.

Diamond Member
Apr 25, 2015
5,661
4,419
136
Why dont you deliver some substance and tell how and when HBM costs are coming down to a level that it can compete and deliver some numbers. Raja said HBM wasn't for mainstream. HBM makers says HBM isn't for mainstream in its current form. Nvidia shows what it think when they spend something like 100M$ on the GP102 IC design and fits it with GDDR5X. Samsung commits to GDDR6(GDDR5X).
Because everything up to GP102 is updated Maxwell architecture. GP100 is true new Pascal architecture, that Nvidia properly spend money on. Thats why they could use GDDR5 on GP104. Because GDDR5X is so similar to GDDR5 you do not even have to change drastically the memory controller, so no changes in the architecture whatsoever. And the fact that GDDR5X can work with GDDR5 in legacy mode is more meaningful here. Examples?

GP104-400 and GP104-200. GTX 1070, and 1080. Both are exactly the same, but one does have GDDR5 and second has GDDR5X.

I would wait with your dick waving about HBM. It is too early to draw any conclusions about the availability and price affect on market. Or in other words, spreading FUD about this. There is a very good reason why AMD designed Greenland/Vega architecture with HBM2 memory controller, and it is the same GPU family with Raven Ridge APUs.

Because there is quite huge chance that we will see 95W, high-end APUs with this.

And to be quite clear from the other way: I do not believe HBM will go into mass market in APUs. It might be only the mid-range to high-end only option, both on mobile and desktop.
 
  • Like
Reactions: kraatus77

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
So easy with the hate word again are we when we dont get what we want? If you dont agree you are a hater, because then you dont have to argument for it. ;)

Its about saying everyone hates Ferraries because you dont accept the everyone wont drive one. Its a superior car and everyone should use it.

Why dont you deliver some substance and tell how and when HBM costs are coming down to a level that it can compete and deliver some numbers. Raja said HBM wasn't for mainstream. HBM makers says HBM isn't for mainstream in its current form. Nvidia shows what it think when they spend something like 100M$ on the GP102 IC design and fits it with GDDR5X. Samsung commits to GDDR6(GDDR5X).

https://www.youtube.com/watch?v=4qJj1ViyyPY

http://arstechnica.com/gadgets/2016/08/hbm3-details-price-bandwidth/

Raja said they will introduce HBM2 when its cost makes sense for a wider range of mainstream products. btw the samller Vega chip which should launch in Q1 2017 or early Q2 2017 is going to end up competing with GP104. Given that Nvidia has already set a USD 399 - USD 449 price range for GTX 1070 I am sure AMD will undercut GTX 1070 with the partially disabled small Vega SKU as they are coming to market much later. At a price of USD 349 that partially disabled Vega SKU is definitely mainstream. I see HBM2/HBM3/low cost HBM and gddr5x/gddr6 coexist and satisfy different market segment needs. low cost HBM is a a perfect fit for AMD APUs which are bandwidth starved.
 

Glo.

Diamond Member
Apr 25, 2015
5,661
4,419
136
https://www.youtube.com/watch?v=4qJj1ViyyPY

http://arstechnica.com/gadgets/2016/08/hbm3-details-price-bandwidth/

Raja said they will introduce HBM2 when its cost makes sense for a wider range of mainstream products. btw the samller Vega chip which should launch in Q1 2017 or early Q2 2017 is going to end up competing with GP104. Given that Nvidia has already set a USD 399 - USD 449 price range for GTX 1070 I am sure AMD will undercut GTX 1070 with the partially disabled small Vega SKU as they are coming to market much later. At a price of USD 349 that partially disabled Vega SKU is definitely mainstream. I see HBM2/HBM3/low cost HBM and gddr5x/gddr6 coexist and satisfy different market segment needs. low cost HBM is a a perfect fit for AMD APUs which are bandwidth starved.
What if Enthusiast Vega is the bigger one GPU? And is aimed at... Volta?

What if High-End Vega is really going out this year?

Two discreet GPUs will use Vega architecture. All of Raven Ridge APUs will use Vega architecture, because it is exactly the same GPU family with Graphics IPv9, according to data from OpenCL drivers.

What if Vega architecture mainstream is only in high-end APUs with pretty huge price tag, like 499$? How much would you have to pay for RX 470 and quad core Intel i7 CPU? AMD has given us a lot of clues lately what might happen in future with Zen, Vega, and Raven Ridge. Its the matter of connecting the dots.
 
Mar 10, 2006
11,715
2,012
126
What if Enthusiast Vega is the bigger one GPU? And is aimed at... Volta?

What if High-End Vega is really going out this year?

Two discreet GPUs will use Vega architecture. All of Raven Ridge APUs will use Vega architecture, because it is exactly the same GPU family with Graphics IPv9, according to data from OpenCL drivers.

What if Vega architecture mainstream is only in high-end APUs with pretty huge price tag, like 499$? How much would you have to pay for RX 470 and quad core Intel i7 CPU? AMD has given us a lot of clues lately what might happen in future with Zen, Vega, and Raven Ridge. Its the matter of connecting the dots.

You really need to reign in your expectations.