Friday, November 24th 2017

Vega 8 Mobile GPU Seemingly Ditches HBM2 Memory, Makes Use of System DDR4 Pool

Update: Industry sources are saying that the 256 MB dedicated system memory to graphics card tasks is currently a firmware limitation, and should be expanded in future firmware revisions to 512 MB and even 1 GB of fully-addressable system DDR4 memory.

AMD's mobile APUs are expected to come with a range of graphics solutions, one of which is the titular Vega 8. Vega 8 isn't supposed to be a high-performance, desktop-class GPU implementation. Instead, it's AMD's counter to NVIDIA's recently announced MX 110 and MX 130 discrete GPUs, which usually deliver relatively low gains (as much as manufacturers want to tout them as extraordinary) when compared to Intel integrated solutions, by far the most widespread IGP. It's expected that Vega 8 performance will slot somewhere around MX110-MX130 levels; and being the low-performance, low cost solution that it's aiming to be, Vega 8 has been made almost unrecognizable from the Vegas we know.
Vega 8 might signify that the graphics chip has 8 NGCUs at its disposal (Vega 64 has 64, and Vega 56 should be pretty self explanatory), which amounts to just 512 shader units. Via HP's recently outed Envy x360 laptops and a carefully taken photo of the Radeon Settings Hardware panel, we can see that this solution eschews the HBM2 memory (which is only natural, considering the pricing and performance of this solution), and instead, seems to make do with just 256 MB of dedicated system memory - out of a possible 8 GB of DDR4 RAM available in the machine.

The lack of actual dedicated GDDR or HBM memory means that costs are shaved down as much as they can possibly be, but also equates to a performance murder. In terms of shaders, the Vega 8 likely counts 512 (8x64), which is equivalent in number, if not in performance (due to Vega's architecture improvements) to the RX 540. However, the lack of high-speed dedicated memory cuts this graphics solution's available bandwidth down to 12 GB/s - and the core runs at a much cooler and power-sipping 300 MHz base, with up to 1.1 GHz boost clocks.Sources: TweakTown, Hardware Info, Thanks @ P4-630!
Add your own comment

49 Comments on Vega 8 Mobile GPU Seemingly Ditches HBM2 Memory, Makes Use of System DDR4 Pool

#1
Fiery
FinalWire / AIDA64 Developer
The device ID (15DD) indicates that this GPU is part of a Raven Ridge APU package. Hence it's absolutely normal that it has no dedicated VRAM.
Posted on Reply
#3
megamanxtreme
I'm still waiting for the sub-$400 variants. Hopefully these MX-110 or 130s hit that low or I have to wait for the AMD variants.
Sucks that they lack dedicated VRAM, Single RAM users can't rejoice.
Posted on Reply
#4
bug
AMD fans are used to (justified) raving about how AMD's IGP runs circles around Intel's. I can't wait to see what they have to say about this.
Posted on Reply
#5
silentbogo
That is so-o-o-o wrong on many levels....
1) As already mentioned, it's a Raven Ridge iGPU core, so no one expected HBM2 in it.
2) This is a pure speculation:
Raevenlord said:
Vega 8 performance will slot somewhere around MX110-MX130 levels
Just from the early "leaked" results R7 2700u was already on par with MX150.
Posted on Reply
#6
Valantar
Jesus, this weird non-news reporting seems to proliferate further and further. It's bad enough that barely-serious sites like TweakTown report this, but TPU? You ought to be better than this.

This is not a dGPU. The Ryzen 5 2500U has "Vega 8 graphics", and the screenshots and pictures are of a HP Envy X360 with that specific APU. The announcement of "Ryzen Mobile with Vega Graphics" had not one single mention of HBM. The lack of HBM or dedicated VRAM should thus not be a surprise to anyone.

These are APU graphics with shared memory. APUs do not have dedicated VRAM. Most likely 256MB is a base number, with a variable buffer that adds to this.

This is not news.
Posted on Reply
#7
_Flare
DDR4-1600 has only 12.8GB/s
DDR4-3200 has 25.6 GB/s

If the Memory-Controller is clever with the same fine-grained DDR4-Handling as the HBCC in the big Vega can.
plus better compression, than the Gap to equal core-count dedicated Cards will shrink.
Posted on Reply
#8
bug
Valantar said:

These are APU graphics with shared memory. APUs do not have dedicated VRAM. Most likely 256MB is a base number, with a variable buffer that adds to this.

This is not news.
Tell that to Intel's CPU that packs AMD graphics.
Posted on Reply
#9
Valantar
silentbogo said:
That is so-o-o-o wrong on many levels....
1) As already mentioned, it's a Raven Ridge iGPU core, so no one expected HBM2 in it.
2) This is a pure speculation:

Just from the early "leaked" results R7 2700u was already on par with MX150.
Benchmarks of Vega 8 place it noticeably behind the MX150. Beats the 940MX fair and square, though.

Edit: of course, the MX150 laptop in question has a lot more thermal/power headroom to boot. I'm looking forward to seeing 25W Ryzen laptops, for sure, both 2500U and 2700U.

https://hothardware.com/reviews/ryzen-mobile-benchmarks-and-performance-analysis?page=2

bug said:
Tell that to Intel's CPU that packs AMD graphics.
Who here is talking about that? This is an article concerning pictures of a HP Envy X360 with a Ryzen 5 2500U APU. The Intel+AMD collaboration has nothing to do with this, and is rumored to operate in power envelopes ~5-8 times higher than Ryzen Mobile (65-100W). Also, Intel specifically names that a "discrete graphics chip". I.e. not an iGPU. And it definitely has no relation to Ryzen.
Posted on Reply
#10
bug
Valantar said:

Who here is talking about that? This is an article concerning pictures of a HP Envy X360 with a Ryzen 5 2500U APU. The Intel+AMD collaboration has nothing to do with this, and is rumored to operate in power envelopes ~5-8 times higher than Ryzen Mobile (65-100W). Also, Intel specifically names that a "discrete graphics chip". I.e. not an iGPU. And it definitely has no relation to Ryzen.
Well, if Polaris can be paired with HBM2, I think it wasn't so unreasonable to expect Vega to do the same. That was my whole point.
But I've said it from the beginning, Vega is a power hog and using it in laptops is going to be challenging.
Posted on Reply
#11
Valantar
bug said:
Well, if Polaris can be paired with HBM2, I think it wasn't so unreasonable to expect Vega to do the same. That was my whole point.
But I've said it from the beginning, Vega is a power hog and using it in laptops is going to be challenging.
Considering they're running 8-10 Vega CUs in a 15W power envelope, it can't be much of a power hog at those clock speeds.

Also, we don't know that the Intel+AMD GPU chip is Polaris - they haven't said more than it being a "semi custom design". Might be Vega 11, for all we know. I suppose Polaris could connect to HBM if they designed a chip with an HBM controller - shouldn't be that big of a challenge, if Intel orders enough chips for the silicon redesign to make economic sense.

As for "unreasonable to expect Vega to do the same" - no, not given a suitable SKU. HBM makes very, very little sense in a relatively cheap 15W chip. Remember, the HP laptop in question starts just above $500. If they make a premium 35W-45W version (possibly with a bigger iGPU), that's another story. But HBM is really, really expensive too - there have been estimates that the 8GB dual-stack setup on Vega 56/64 alone (without the interposer) costs ~$150. That's probably more than half of what a 2500U costs for someone like HP.
Posted on Reply
#12
T1beriu
Dear "news" writer, what have you been smoking?!

The APU was physically shown more than 6 months ago. There was no HBM on it then. You could have broken this "news" then. Oh lordy.
Posted on Reply
#13
medi01
For mere 400 Euros more, you could get awesome i7 with MX940 that are on par with AMD's offering in both CPU and GPU performance (while consuming more power), just think about that.
Posted on Reply
#14
Sempron Guy
since when did AMD announce Raven Ridge will use HBM2?
Posted on Reply
#15
bug
Sempron Guy said:
since when did AMD announce Raven Ridge will use HBM2?
That's the thing, they never announced anything about Raven Ridge, but all Vega we know is using HBM2. This news is significant, because is means AMD has actually produced two memory controllers for Vega (just like Nvidia did for Pascal and Maxwell). It also sets an upper limit on performance expectations.
Posted on Reply
#16
Valantar
bug said:
That's the thing, they never announced anything about Raven Ridge, but all Vega we know is using HBM2. This news is significant, because is means AMD has actually produced two memory controllers for Vega (just like Nvidia did for Pascal and Maxwell). It also sets an upper limit on performance expectations.
The thing is, them not explicitly announcing that RR would use HBM2 could even at launch safely be interpreted as confirmation that it didn't use HBM2. Why? Cost (HBM2 is really expensive), complexity (adding dual memory controllers to a die makes it bigger, more complex, and more expensive, not to mention the interposer you'd need for HBM2), the fact that interposers were not mentioned at any time, and that they've shown off RR silicon with just a single die on the package. Adding a memory controller to a die if it's expected to be used in <10% of sold units would not be a good decision. As you point out, using different memory controllers across market segments is neither new or really newsworthy. High end products using expensive and new memory technologies is nothing new in the GPU world, and we've never, ever seen this trickle down to low-end or integrated solutions before the recently announced Intel+AMD collaboration. Heck, there are still entry-level dGPUs out there using DDR3!

If RR had the option for HBM2, AMD would have been shouting it from the rooftops, simply because it would be ridiculously competitive for esports and the like, but still cheap(er than a dGPU), a great sales argument, and a driver for enthusiasm and excitement.
Posted on Reply
#17
bug
Ok, this thing is the best thing since sliced bread. Us, doubters, stand ashamed.
Posted on Reply
#18
Valantar
bug said:
Ok, this thing is the best thing since sliced bread. Us, doubters, stand ashamed.
What an odd response.

I'm simply arguing to keep assumptions realistic and to base them off of what we know, not what we might hope. And, of course, to not report already-known facts as news.
Posted on Reply
#19
Kaotik
bug said:
That's the thing, they never announced anything about Raven Ridge, but all Vega we know is using HBM2. This news is significant, because is means AMD has actually produced two memory controllers for Vega (just like Nvidia did for Pascal and Maxwell). It also sets an upper limit on performance expectations.
No it's not and no they didn't.
AMD GPU architectures are modular and you can pretty much mix'n'match any IP blocks.
In case of APUs, they of course use the shared DDR4 memory controller, in case of Vega 10 they use HBCC HBM2 memory controller, in case of Polaris 10 they use GDDR5 memory controller and so on - but they could for example make a Polaris with HBM memory controller (the AMD+Intel chip actually might be Polaris+HBM instead of Vega) or Vega with GDDR memory controller, but that would mean completely new GPU design even when the architecture is the same, and in case of Vega losing certain features since the GDDR memory controllers they have (at least publicly) can't do things the HBCC can
Posted on Reply
#20
john_
Don't shoot the messenger :p

Other sites, Guru3D included, also look amazed by the fact that it doesn't use HBM2. I thought it was common knowledge from about forever. Vega architecture is a GPU architecture, not an HBM2 memory controller. Why would Vega have to be be an HBM2 only solution?

As for performance, let's wait and see some 25W TDP solutions first, judge later. And yes, it does run circles around intel iGPUs, even in this HP laptop.
Posted on Reply
#21
bug
john_ said:
Don't shoot the messenger :p

Other sites, Guru3D included, also look amazed by the fact that it doesn't use HBM2. I thought it was common knowledge from about forever. Vega architecture is a GPU architecture, not an HBM2 memory controller. Why would Vega have to be be an HBM2 only solution?

As for performance, let's wait and see some 25W TDP solutions first, judge later. And yes, it does run circles around intel iGPUs, even in this HP laptop.
You're right, technically. But as long as you manage to stick with one type of memory, you only have to design (and integrate and test and support) one SKU. For those that thought AMD was able to cut their costs and made Vega HBM only, this is the first confirmed clue that' not the case. Which in turn matters little, because it's not like Vega was cheap to begin with, so even if AMD managed to cut their costs, the end-user wasn't seeing it.
(sorry if the above isn't too clear, but the significance of Vega not being HBM2-only isn't either)
Posted on Reply
#22
john_
bug said:
You're right, technically. But as long as you manage to stick with one type of memory, you only have to design (and integrate and test and support) one SKU. For those that thought AMD was able to cut their costs and made Vega HBM only, this is the first confirmed clue that' not the case. Which in turn matters little, because it's not like Vega was cheap to begin with, so even if AMD managed to cut their costs, the end-user wasn't seeing it.
(sorry if the above isn't too clear, but the significance of Vega not being HBM2-only isn't either)
Nope. When you don't have the money in your pockets, you stick with only one design. If and when you manage to get some money, then you can create a second design with a different memory controller. Nvidia that enjoys full pockets, creates multiple designs in parallel. AMD.... one at a time.

If I am not mistaken it is rumored for many weeks/months now that AMD will create cheaper cards based on the Vega architecture to replace Polaris models and it was obvious from when those rumors got out, that those cheaper models would not be using HBM2. AMD still tries to create new products by mainly recycling what it already has. Probably one of the reasons Raja left. So people saying that AMD was sticking with HBM2 where right and still are. On an APU design or on a mainstream card, NO ONE expects you to use an expensive solution as HBM2 is. Also, HBM2's main advantage is power consumption and space. On an APU where the integrated GPU doesn't have dedicated memory anyway, why even talk about HBM2? It makes no sense at all.
Posted on Reply
#23
Aquinus
Resident Wat-man
There 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.
Posted on Reply
#24
Imsochobo
bug said:
Tell that to Intel's CPU that packs AMD graphics.
Intel doesn't pack memory on their 15watt packages.
this is a 15-25w package, hbm alone would make it a 30-40 watt package.
Posted on Reply
#25
Kaotik
bug said:
You're right, technically. But as long as you manage to stick with one type of memory, you only have to design (and integrate and test and support) one SKU. For those that thought AMD was able to cut their costs and made Vega HBM only, this is the first confirmed clue that' not the case. Which in turn matters little, because it's not like Vega was cheap to begin with, so even if AMD managed to cut their costs, the end-user wasn't seeing it.
(sorry if the above isn't too clear, but the significance of Vega not being HBM2-only isn't either)
Sigh, 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.
Posted on Reply
Add your own comment