HBM should represent quite a significant saving in power draw, compared to 512-bits worth of GDDR5.
AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread
Silent_Buddha
Legend
Does anyone know the power consumption of HBM? We already know that GDDR is relatively power hungry. So if they are using HBM and it uses less power than GDDR (either inherently or via less complex board design/traces), that could also provide some theoretical power savings.
As well, like silent_guy said, there's no reason that AMD wouldn't be able to eek out some more power savings from the chip itself similar to Nvidia. After all, IF there's any truth to those graphs, the power/perf ratio is pretty similar to GTX 980.
But yeah, BIG IFs and large piles of salt WRT whether there's any validity to those graphs.
Regards,
SB
As well, like silent_guy said, there's no reason that AMD wouldn't be able to eek out some more power savings from the chip itself similar to Nvidia. After all, IF there's any truth to those graphs, the power/perf ratio is pretty similar to GTX 980.
But yeah, BIG IFs and large piles of salt WRT whether there's any validity to those graphs.
Regards,
SB
Would that tell us very much though? You'd also need to factor in the power consumption of the (presumably different) on-die memory controllers. FWIW (not much IMO), the hynix graphics memory webpage at https://www.skhynix.com/gl/products/graphics/graphics_info.jsp makes very vague claims of 65% system performance improvement (what system? what workload?) for 40% less memory power consumption than GDDR5.
Logic dictate it has to be very low, due to it being a stack of dies. You can't cool the lowest and intermediate dies directly, which means they must not dissipate very much power, or the stack would either kill itself or become unreliable.
I think Hynix once said it draws about 50% less power than GDDR5 for twice the bandwidth or something like that. For what it's worth, the R9 290X has a TDP of 294W (I think, or very close) and the Hawaii GPU itself is limited to 208W, which leaves some 86W for everything else, including the memory.
But I suppose HBM might also allow for power savings on the GPU itself, e.g. the memory PHYs.
But I suppose HBM might also allow for power savings on the GPU itself, e.g. the memory PHYs.
AMD will manufacture GPU's at Global Fundries:
Source AMD via Guru3D
In our previous post about 16nm TCSM we've been wondering as to why AMD was not on the 16nm fab list, whereas other players like Nvidia have been named. The answer just came to us though, AMD will manufacture their GPU at Global Foundries. Global Foundries has made great progress claims Devinder Kumar from AMD at an investor meeting.
AMD sold Global Foundries in 2008 to raise cash in difficult times, Kumar stated that they will still fab 'some' products at 20nm and from there on-wards it will move to Finfet-chips. Gobalfoundries is working together with Samsung on 14nm already, Samsung will launch with SoCs based on that fab node next year already.
Finfets have a bigger surface area which helps switching faster for the transistors. Finfet chips are more power efficient, smaller and cheaper to manufacture.
It's unclear of AMD will remain partner of TSMC.
Source AMD via Guru3D
In our previous post about 16nm TCSM we've been wondering as to why AMD was not on the 16nm fab list, whereas other players like Nvidia have been named. The answer just came to us though, AMD will manufacture their GPU at Global Foundries. Global Foundries has made great progress claims Devinder Kumar from AMD at an investor meeting.
AMD sold Global Foundries in 2008 to raise cash in difficult times, Kumar stated that they will still fab 'some' products at 20nm and from there on-wards it will move to Finfet-chips. Gobalfoundries is working together with Samsung on 14nm already, Samsung will launch with SoCs based on that fab node next year already.
Finfets have a bigger surface area which helps switching faster for the transistors. Finfet chips are more power efficient, smaller and cheaper to manufacture.
It's unclear of AMD will remain partner of TSMC.
Blazkowicz
Legend
tiffany is a really cute word or female first name but I don't understand the cultural meaning or reference.
AMD already has the experience of translating a product from TSMC to GloFo, they did that with Kabini (AM1 socket versions)
AMD already has the experience of translating a product from TSMC to GloFo, they did that with Kabini (AM1 socket versions)
They do, but I'm not sure it's a good fit for a GPU.
http://www.globalfoundries.com/technology-solutions/leading-edge-technology/20-lpm
According to quite respected guy on AMD business here in Finland, 20nm will be only used for ARM-based products
Someone from Muropaketti?
Anyway, I figured as much. Carrizo is known to be on 28nm, ditto for Carizo-L (which seems to be a renamed Beema) and there are persistent rumors that AMD is skipping 20nm for discrete GPUs. That only leaves ARM stuff.
Edit: Oh, and the x86 variant of Skybridge, if that's still happening.
Last edited:
Doesn't work with us at Muropaketti, but posts at our forums among others, maybe here too, not sure (The Stilt)
Tiffanys is a boutique jewelry store in New York City. The implication is exclusive and expensive.
D
Deleted member 13524
Guest
Until I went to NY last year, I thought Tiffany's was a caffé.
You know, some place where you'd have breakfast
Babel-17
Veteran
It's the Cadillac of jewelry stores.