KaotikSigh, the "GPU" portion doesn't really care if they stick a HBCC HBM memory controller or some other memory controller in there.
What you're calling "Vega" has several separate IP blocks which can be switched around with different blocks, you could replace the GPU portion with different GPU IP block (for example Polaris-level block), UVD with different version of UVD or ditch the whole UVD for that matter etc etc - and that includes the memory controller as a separate IP block.

Perfect example of this would be Fiji and Tong. They share most of their parts from the same IP level, they're both GCN3 etc etc, but one has HBM and one has GDDR memory controller.
Another example is every single APU they've made - they all share most of their blocks with discrete GPUs but they all use shared DDR memory controller with the CPU, something none of the discrete GPUs has or does.

Similar to this, they could do "Vega" (GCN5) with GDDR memory controller if they so choose, and they've made APU with Vega gfx portion and shared DDR memory controller with the CPU.

There is no "AMD made Vega HBM only", AMD made Vega 10 HBM only, just like they made Fiji HBM only and Tonga GDDR only because there's no sense (or even space) to put two completely different memory controllers in the same chip. They could still do for example Vega 11 with GDDR memory controller or Vega 12 or whatever they want to call such hypothetical chip.

bugYes, and? If it's modular it doesn't cost extra to replace the memory controller? We all know the memory controller is a separate block, it's not like it's built into the shares or something. That was not the point.

bugYes, and? If it's modular it doesn't cost extra to replace the memory controller? We all know the memory controller is a separate block, it's not like it's built into the shares or something. That was not the point.

bugI can't wait to see what they have to say about this.

AquinusThere is plenty of room in most towers to use off-package DRAM. HBM should be used for mobile platforms where space and power are at a premium. Even for low end hardware, just use less of it. Even one stack of HBM2 could do wonders for performance on a mixed GPU/CPU package. I would expect it to essentially behave like Crystalwell did on eDRAM-enabled Intel CPUs except with more capacity and performance.

ApocalypseeThis is what I thinking is optimal for APU, but the complexity of interposer and HBM requires additional circuitry and power planes going to complicate matters especially on mobile because lets face it, it still going to require regular DDR memory to operate.

AquinusNot if you cut out a memory channel. If HBM is treated like another layer of cache, they could reduce the size of the DDR4 controller to make room and free up power to be used by HBM. Also, just one stack of HBM means less power as well compared to two. This really isn't any different from how L2 and L3 cache work or how Crystalwell works. There are times when you need fast memory access and there are times when you need bulk memory access and adding a layer between external system memory and cache opens the possibility where more memory helps so you don't have to swap to disk but, there isn't necessarily a requirement for external DRAM to be present. To me, that makes a lot more sense because you're retaining the benefits of both without going 100% in either direction. Something like this would intrigue me for a HTPC or a small form-factor workstation.

AquinusNot if you cut out a memory channel. If HBM is treated like another layer of cache, they could reduce the size of the DDR4 controller to make room and free up power to be used by HBM. Also, just one stack of HBM means less power as well compared to two. This really isn't any different from how L2 and L3 cache work or how Crystalwell works. There are times when you need fast memory access and there are times when you need bulk memory access and adding a layer between external system memory and cache opens the possibility where more memory helps so you don't have to swap to disk but, there isn't necessarily a requirement for external DRAM to be present. To me, that makes a lot more sense because you're retaining the benefits of both without going 100% in either direction. Something like this would intrigue me for a HTPC or a small form-factor workstation.

ApocalypseeImplementing HBM is not as easy as eDRAM. Even though its a single stack it still requires interposer and other circuitry involving with it. I dont believe it could be implemented on the same socket unlike Intel did on Broadwell (except that Intel did the 'new motherboard for new CPU' thing they always did). AMD could not afford to do this, since both Fury and Vega desktop doesnt do well and it will jack up the price significantly for what its aimed for.

AquinusYou're right. Something like this doesn't happen overnight but when you scale up production, these costly things become less costly over time. My point is that it's doable and could be a possible middle ground that could scale from the small to the big but, a change like that takes time and money.

ShamalamadingdongWell, it kinda does cost extra.

You need to make changes to the die meaning you need to make a new mask for the Fab and then it needs to tape-out and you need to validate it. It'll take many months and millions to get such a product ready. It's not as simple as switching a component on an assembly line. It becomes a new chip.

If the memory controller was external, then it should be really simple. But it is true that they don't need to start from scratch to implement GDDR on Vega. The blocks are interchangeable with little work.

Integrated graphics are designed with system memory in mind and always have been for AMD. That's why it's a surprise to no one that Raven Ridge does not have HBM. It's been known for months. The package would also be much different to support dedicated memory and would add considerable cost, size, complexity and power. You'd probably see a $100-200 premium (conservative estimate) on every single device with a Raven Ridge chip (depending on memory configuration). Alternatively they would have to spend those millions and months making two chips. There is a budget constraint going on at AMD. It simply isn't feasible to do so. They need to make one-size-fits-all chips to get as much money out of each chip and so far it's working but it isn't the best for every situation. Hopefully, AMD can start making multiple chips per generation like Intel and Nvidia does; perhaps next year.

ApocalypseeWhat AMD should done is like what they planned on Kaveri by implementing quad channel memory (256-bit) and/or GDDR5 memory. You can read Anandtech article about it. IMO for mobile they could just ditch regular DDR and use GDDR5. There's no need for additional slot just soldered it on the motherboard. GDDR5 have been around since 2007, it shoudnt costs that much. A standard dual channel (128-bit) GDDR5 runs at measly 1GHz (4GHz effective) netted 64GB/s bandwidth. Modern GDDR5 runs at twice the speed of that.

john_Don't shoot the messenger :p

Other sites, Guru3D included, also look amazed by the fact that it doesn't use HBM2.

AquinusYou're right. Something like this doesn't happen overnight but when you scale up production, these costly things become less costly over time. My point is that it's doable and could be a possible middle ground that could scale from the small to the big but, a change like that takes time and money.

ValantarHBM2 supply is still very meagre, and both HBM and implementing it would be crazy expensive. Estimates for Vega say its dual 4GB stacks cost ~$150 before the cost of the interposer and all that jazz. In other words, a single 4GB stack would be $75, and the interposer and thus complicated mounting would be at least another $25, giving us a neat price hike of >$100 for a normally ~$200 APU (possibly a bit more, but Intel's competing solutions list at $303, so definitely less than that). Does that seem like a good idea to you? Even if halving the memory amount brought that down to ~$65, that's still way too much. And there's no indication HBM or interposers will become significantly cheaper over the next year.

AquinusSomething like this doesn't happen overnight but when you scale up production, these costly things become less costly over time. My point is that it's doable and could be a possible middle ground that could scale from the small to the big but, a change like that takes time and money.

john_EMIB might be nice for future discrete GPUs, but probably interposer is good enough for that job. EMIB in laptops with Ryzen CPUs or even APUs, don't know if it makes sense, considering that AMD would have to be able to persuade OEMs to create enough models with AMD CPUs/APUs to justify any payment to Intel, to use that tech. I doubt intel gives any kind of access/license to AMD for the EMIB. They did collaborated closely in the creation of the Kaby lake G, but I think, in this case, things are as simple as "intel is a customer of AMD" and nothing more, for now. Intel gone to AMD and said "We want one of your GPUs to put next to a Kaby lake using this tech. You can help us do it, or we can go to Nvidia, even if we hate the idea".

AquinusTooling and scaling production tends to reduce the cost of manufacturing complex hardware like HBM. What I said prior is 100% applicable:

Let me put it another way: It won't cost that much forever.

ValantarIsn't the whole point of EMIB that it's far cheaper, mechanically simpler and easier to implement than an interposer? Saying that an interposer "is probably good enough" misses the point entirely when the point of EMIB is to do the same for less, not more or better. This is also why I think AMD licensing EMIB makes sense, as it (given a reasonable licensing deal, which KBL-G might give them) would significantly cut production costs for their high-end GPUs, as well as open up the possibility of using HBM on lower end products where the added interposer cost would be too steep. While there's no doubt AMD is getting paid for their GPUs in KBL-G, it seems odd to me that they'd agree to something like that without some exchange of technology or licensing baked into it.

While that is true, this is still not the time to implement HBM with an interposer in a relatively cheap mobile APU. AMD has no control over HBM production, and supply is short and will continue to be so for a while yet (unless HBM production somehow is entirely immune to the current DRAM/NAND production shortages - which doesn't make sense). I'd be shocked to see HBM hit a mid-range APU before it hits midrange GPUs - margins are far higher there, after all. And for now, HBM is for the high end. Sure, Intel is putting it in KBL-G, but Intel has cash to throw around, and entirely owns the market for high-end mobile devices. They're in a far better place to do so. A generation or two from now, the situation will probably be different.

ShamalamadingdongDon't expect EMIB to be nothing but Intel exclusive unless Intel finds issues capitalizing on it and it suddenly makes sense to open it up for licensing. EMIB is a definite ace for Intel. It opens a lot of new avenues (including KBL-G) but allowing the competition to even the playing field and eventually making better products would be a bad move and Intel doesn't need the money that badly - it's the other way around. AMD probably can't afford what Intel would require to relinquish the exclusivity. We already see a lot of what AMD's doing falling flat because they lack resources. Of course the counter-argument would be that they need to spend money to earn money but there are simply limits to what they can do because every move they make could end up tipping the scales towards bankruptcy.

ValantarYou're not wrong, but what does AMD have that Intel needs? GPU tech, or more specifically GPU patents. There's no way Intel will be able to develop their own dGPU - even with the help of Raja Koduri - without stepping on the toes of AMD or Nvidia, so they rather desperately need a patent licencing agreement. The one they had in place with Nvidia expired (even if it still gives them access to patents from before its expiration), so AMD is a natural business partner there. And as the one holding all the rights to EMIB, Intel could essentially tailor a licence that would make sure AMD didn't use the tech in any products that compete with Intel's bread and butter, such as server hardware and the like. An agreement like this could make licencing AMD's GPU patents a heck of a lot cheaper for Intel, that's for sure. And even if Intel has the cash, corporations generally don't like giving away cash when they can avoid it.

ValantarIsn't the whole point of EMIB that it's far cheaper, mechanically simpler and easier to implement than an interposer? Saying that an interposer "is probably good enough" misses the point entirely when the point of EMIB is to do the same for less, not more or better. This is also why I think AMD licensing EMIB makes sense, as it (given a reasonable licensing deal, which KBL-G might give them) would significantly cut production costs for their high-end GPUs, as well as open up the possibility of using HBM on lower end products where the added interposer cost would be too steep. While there's no doubt AMD is getting paid for their GPUs in KBL-G, it seems odd to me that they'd agree to something like that without some exchange of technology or licensing baked into it.

john_If I am not mistaken, interposer is AMD's patent. So, maybe it makes more sense AMD to improve the interposer, than trying to come to an agreement with intel. Especially if the difference in costs and complexity are not significant(you say far cheaper, could be, any links?). Also, while it makes perfect sense to prefer EMIB in a very slim laptop, if AMD is not going to create something like an APU with a very strong GPU and HBM or HBM2 sitting next to that APU, for ultra slim devices, there isn't really any important reason to go and beg intel. On discrete GPUs, having the main chip one or two millimeters higher than the PCB, isn't really a problem.

ShamalamadingdongIntel could do like Apple and attempt to make their GPU tech a giant black box and hope their competitors don't find any evidence to support patent litigation or hope that competitors are afraid of opening up Pandora's box which in this case consists of large scale patent warfare.

But let's assume Intel trades EMIB for either general GPU patents or Radeon tech. Is it of equal value? I don't think so personally.
And who's to say AMD isn't working on something similar to EMIB (of course without infringing)? It would seem wise to have had something in the works to replace interposers unless AMD has had no foresight as to the money drain HBM and interposers have turned out to be.

I have no doubt AMD wants EMIB or an EMIB-like solution if nothing else but to save money. Intel doesn't like to play ball though; they only do when they absolutely have to.
Examples being opening up Thunderbolt, KBL-G, relationship with Microsoft and Apple (although funnily enough only on Mac; they told Apple to fuck off when they asked Intel to develop a processor for phones and look how much they regret that today as it wasn't the money sink they thought it'd be).
Other than that, Intel have a habit of making sure the competition has a hard time to put it mildly.

Don't get me wrong, I want AMD to be able to execute on their strategy which ultimately relies on tech enabling the connection of chips together and EMIB is a revolutionary way to do that.

ValantarI haven't seen concrete "evidence" of this, but I have read multiple news reports claiming that EMIB's chief advantages are lower cost and easier implementation, due to far smaller silicon pieces needing to be fabbed. Remember, the big issue with Fiji was that the interposer was around 1000mm2, necessary to fit the die, HBM stacks and interconnects. While Vega's interposer is clearly smaller (slightly smaller die, and only two stacks of (slightly larger) HBM2), it's still a significant piece of silicon. Now, I don't doubt that embedding silicon pieces into a substrate is complex, but it would have to be a seriously costly process to approach the cost of fabbing a 7-800mm2 interposer. Remember, the EMIB silicon only needs to be the size of the connections and pathways between them, so for HBM that would likely not be more than 50% larger than an HBM stack. The cost of silicon wafers alone, let alone manufacturing, makes the savings there clear.

john_It could be something like $25 vs $15. And that's when you don't count for how much Intel will be asking. And having to come with some kind of agreement with Intel and considering that whatever assembly is to be done, it will be done at Intel's factories, maybe the easier implementation argument also loses it's value. But I could easily be wrong and this to be just a very bad case scenario to base my point of view.
The 1000mm2 size of the interposer maybe isn't a problem. We are not talking about a complicated chip, and neither for last gens manufacturing technologies. It is in fact a really simple piece of silicon and I think made at 45nm or something.

ValantarOf course, this is all speculation, and my main reason for thinking about this is that for me, it seems unlikely that AMD would enter a close collaboration with Intel without getting something more substantial than cash back. Of course, I might be entirely wrong.

But more to the point: I doubt the price differential is that small. And sure, interposers are "simple" chips with few layers and little work required, but the size is a serious limitation. Heck, Fiji's interposers exceeded GloFo's reticle limit, requiring some special finagling to be able to make them at all. According to Google, standard 300mm silicon wafers cost more than $3 per square inch (that's from 2014, from what I've read prices are higher now). In other words, a 1000mm2 wafer costs a minimum of ~$5 in silicon alone (a square inch is ~650mm2), but given that large square chips have very poor area utilization on circular wafers, the cost is likely to be closer to twice that - and that's before actually etching anything into the chip. The price drop for, say, a 100mm2 EMIB chiplet would scale far better, as yields would be far superior, and there'd be no increase in fab costs. AFAIK the cost of processing a wafer on the same process node is roughly the same no matter the compelxity of the mask, so then you'd be splitting the cost over ... a few dozen interposers per wafer?, compared to hundreds if not thousands of EMIB chiplets per wafer. In other words: fab savings would be significant. Of course, embedding the chiplets into the substrate is probably more complex than soldering(?) an interposer in between a chip and a substrate, which might even the playing field somewhat, but I still have the feeling that EMIB would be dramatically cheaper.