Vega 8-GPU for laptops uses DDR4, not HBM2

Silva

2017-11-25 13:37

kruno:

Power consumption, GDDR5 requires more power (significantly ) then HBM2,and considering that Vega is already power hungry using GDDR5 memory would make things even more worse then it is.And second thing is as seen in different reviews Vega is memory constrained,so you would need to put boatload of GDDR5 chips to get HBM2 throughput.

Vega isn't power hungry, just because AMD pushed the architecture to it's limits doesn't mean it's bad. They tried to beat the 1080Ti and failed. Vega cards come with 3 power modes and if you enable the power saving one, you get a card that performs 5/10% less using 100 less watts. I'm sure a Vega 32 with 4GB DDR5 not pushed to the brim would happily sit below 200W and I would buy it at the right price.

#5495444

kruno

2017-11-25 14:12

Silva:

Vega isn't power hungry, just because AMD pushed the architecture to it's limits doesn't mean it's bad. They tried to beat the 1080Ti and failed. Vega cards come with 3 power modes and if you enable the power saving one, you get a card that performs 5/10% less using 100 less watts. I'm sure a Vega 32 with 4GB DDR5 not pushed to the brim would happily sit below 200W and I would buy it at the right price.

Yeah, i know see my post above yours, post was already long enough so i didn't go further with it to explain why Vega was "bad" ,but i sad that power hungry in the context to explain why did they go with HBM2 instead of GDDR5 , and that is GDDR5 would even further increase power draw

#5495479

Loophole35

2017-11-25 19:54

Kaotik:

No, it's not a good question. It's quite simple really, AMDs GPU architectures are more or less modular these days, you can mix'n'match different IP blocks -together (with some possible limitations, but now that they have Infinity Fabric it's more simple than before, examples of different IP blocks include the "GPU-block", VCE, UVD, memory controller etc). Vega 10 has HBCC, which is HBM2-memory controller. It can't use GDDR or DDR-memory. That, however, doesn't mean that every Vega has to have HBCC, like in case of Raven Ridge, they're sharing the same DDR4-memory controller with the CPU. They could make a Vega with GDDR-memory controller too, but it would be completely separate project and chip. This is also why we're currently not sure whether the dGPU on that AMD+Intel-combochip is actually Vega or Polaris. HBM and timing hints Vega, but old leaks suggest Polaris, and nothing would prevent AMD from making a Polaris with HBM-memory controller (whether it's same as in Fiji or same as in Vega is another thing). Oh, and AMD said somewhere around Polaris launch outright that they have the capability to use which ever memory technology makes most sense for each chip, referring to the modularity of the architecture. This of course has to be decided before you start laying out your chip, but it can be done.

I think you are misunderstanding what his question was. It was why if they could, didn’t they make Vega with GDDR almost two years ago when they KNEW HBM2 would not be available?

#5495491

Kaotik

2017-11-25 20:38

Loophole35:

I think you are misunderstanding what his question was. It was why if they could, didn’t they make Vega with GDDR almost two years ago when they KNEW HBM2 would not be available?

We don't actually still know if HBM2 availability has been deciding factor on Vega availability or not, it's just speculation. We know for a fact that they've been already using 8GB HBM at the time when everyone said such chips don't even exist and aren't even coming for a long time edit: And also because AMD has very limited R&D budget compared to the competition they kinda need to make their chips last for 2 gens, upgrading midrange/lowend now, highend next, midrange/low then etc etc late edit: and Vega GPU portion definitely wasn't ready 2 years ago

#5495564

icedman

2017-11-26 07:02

This news is somewhat disapointing since Richland with ddr3 overclocked to 2400mhz( yes some loose timings where used) was still scaling up most laptop configs only use 2400mhz ddr4.

#5495659

tunejunky

2017-11-26 17:26

Yawn. i'm waiting for mobile ryzen with a discrete gpu like a vega 56.

#5495701

Alessio1989

2017-11-26 21:17

Let's hope this time the new MC will shine, after all the swizzle's now 64k standard......

#5495829

rl66

2017-11-27 14:34

SweenJM:

Back to the topic, i think this is pretty interesting. So this tells me that you can either use the normal chip interconnects, or can use hbm w interposer and substrate as is required for the specs of chips. pretty neat. dont have to alter the chip radically to use hbm.

yes and no... this is before foundery process, and so it is still difficult, just a bit more easy just not have to do less research.

#5495873

Kaotik

2017-11-27 16:43

SweenJM:

Back to the topic, i think this is pretty interesting. So this tells me that you can either use the normal chip interconnects, or can use hbm w interposer and substrate as is required for the specs of chips. pretty neat. dont have to alter the chip radically to use hbm.

No, you understood it completely wrong. You still have to design the whole chip (layout) from scratch even when the blocks can be considered modular. GDDR and HBM memory controllers are radically different on the chip (see picture in the end, they're different chips but you can easily see what I mean anyway) Oh, and HBM doesn't necessarily need interposer if you have another way to get wide enough bus through, the usual substrate won't work but for example the AMD+Intel-chip uses Intel's EMIB instead of interposer. (edit: and just for the sake of it, Tonga does indeed have 384-bit (6x64bit) memory controller as seen in the image, none of the products use more than 256-bit though) https://cdn.wccftech.com/wp-content/uploads/2015/11/AMD-Fiji-Tonga-Die-shots-2.jpg

#5495896

tsunami231

2017-11-27 17:13

am i reading that right ? DDR4

#5495926

Kaotik

2017-11-27 19:12

tsunami231:

am i reading that right ? DDR4

Of course, just like previous APUs it's sharing system RAM with the CPU

#5495956

icedman

2017-11-27 20:41

I hope for the sake of these new apu's they use higher speed laptop ram or they will only perform marginally better than previous apu's unless they have crazy texture compression or something.

#5495959

kruno

2017-11-27 20:49

icedman:

I hope for the sake of these new apu's they use higher speed laptop ram or they will only perform marginally better than previous apu's unless they have crazy texture compression or something.

You know already how that is going to end, even at presentation show (when AMD introduced mobile chips) there was OEM's that equipped Ryzen laptops with single channel memory (Lenovo if i remember correctly),and there was some with 2133MHz memory. So don't get your hopes high because it looks like good old game that OEM's like to play and it is called #Let's f**k AMD"

#5495993

Agent-A01

2017-11-28 01:34

tunejunky:

Yawn. i'm waiting for mobile ryzen with a discrete gpu like a vega 56.

That will never happen. Way too big of a chip

#5496193

tunejunky

2017-11-28 17:56

Agent-A01:

That will never happen. Way too big of a chip

no need for apu, discrete cpu + gpu, like high end laptops now.

#5496238

Agent-A01

2017-11-28 19:57

tunejunky:

no need for apu, discrete cpu + gpu, like high end laptops now.

I'm not talking about APU, discrete GPU won't happen. Die size of vega is huge, 50% bigger than 1080mobile. It's also bigger than a 1080Ti which is already 40% faster. So logistics prevent that from ever happening.

#5496263

tunejunky

2017-11-28 21:19

Agent-A01:

I'm not talking about APU, discrete GPU won't happen. Die size of vega is huge, 50% bigger than 1080mobile. It's also bigger than a 1080Ti which is already 40% faster. So logistics prevent that from ever happening.

yes, i suppose you're right. a boy can dream can't he 😳 anyway, maybe 7nm will bring that within reason... the power needs that anyways to be practical since vega is a piggy little beast... but in all practicality Nvidia will kill in mgpu (again).