HBM2 advantages

[Chaos27]

Im a very curious person and find these things very interesting hehe.

I know that HBM and HBM2 is new technology and expensive compared to GDDR5 and GDDR5X.

I was just wondering is HBM2 better in every way to GDDR5X?

And is it true that HBM2 is more power efficient then GDDR5X and HBM and GDDR5? and if it is does it have an advantage in energy efficiency.

Is it that HBM2 is bigger (GB) and faster?

Also is HBM compared more to GDDR5/X or more with DDR4?

As I have read that DDR4 doesn't have much of an advantage compared to DDR3 and also DDR4 apparently uses 15w less then DDR3.

I haven't read about LPDDR4 but i know mobiles, tablets etc uses LPDDR3.

Was wondering if HBM2 is better for mobiles compared to LPDDR3 or LPDDR4.

Apparently LPDDR4x uses 40% less energy.

[kalniel]

I know that HBM and HBM2 is new technology and expensive compared to GDDR5 and GDDR5X.

I was just wondering is HBM2 better in every way to GDDR5X?

You just answered that above. No, it's much more expensive. Also it isn't pin/controller compatible with GDDR5, so significant design and PCB work has to be carried out.

And is it true that HBM2 is more power efficient then GDDR5X and HBM and GDDR5? and if it is does it have an advantage in energy efficiency.

Yes, energy efficiency is the same thing as power efficiency.

Is it that HBM2 is bigger (GB) and faster?

It's a technology, not a memory size unit, so the bigger question doesn't quite make sense. In practical terms HBM2 is going to be available in more or less the same sizes that GDDR5(X) is. There is a bit less flexibility in how those sizes are achieved though. Yes it's faster.

Also is HBM compared more to GDDR5/X or more with DDR4?

GDDR5/X - it's a graphics RAM technology at this point in time.

As I have read that DDR4 doesn't have much of an advantage compared to DDR3 and also DDR4 apparently uses 15w less then DDR3.

DDR4 has quite an advantage over DDR3, but don't forget you might be comparing a technology that has been on the market for a long time thus is well optimised to a product that is still fairly new.

I haven't read about LPDDR4 but i know mobiles, tablets etc uses LPDDR3.

Was wondering if HBM2 is better for mobiles compared to LPDDR3 or LPDDR4.

Apparently LPDDR4x uses 40% less energy.

Different uses. As mentioned HBM2 is a high end graphics RAM. LPDDR is a low power system RAM component. In the far future there may be an HBM that is flexible and cheap enough for use in system RAM. Many of us on these forums have already speculated that HBM would make a good choice for a system on chip type design where the CPU and GPU are part of the same package. I wonder if the Xbox scorpio might go down this route.

[DanceswithUnix]

It isn't that the memory is lower power, it is more that the interface to drive it requires less power because the memory has to be in the same package as the GPU it is connected to.

Driving pins outside a chip requires by CPU/GPU standards some huge transistors and they require both space and power. By going to an easier to drive memory bus on really short wiring you can use some of that silicon area for shaders and use the power to clock the whole thing faster.

[Chaos27]

So HBM right now is a graphics ram as well.

So In the future when HBM is cheaper then it will take over GDDR5 and GDDR5 will be a thing of the past.

And if I read correctly HBM is more efficient and faster at the same time. So basically more efficient per watt.

Would be good to see HBM 3 and how much better it would be.

Don't know if HBM 2 is a big difference to HBM in power efficiency and speed etc.

Thanks for the answers. I do find it very interesting.

[watercooled]

And if I read correctly HBM is more efficient and faster at the same time. So basically more efficient per watt.

In computing terms, performance per watt is referred to as efficiency. For example, some hypothetical memory which uses half the energy but is a quarter of the speed of a another memory standard would be considered less efficient - sure it uses less power but overall it uses twice as much per unit of work.

Don't confuse this with thermodynamic efficiency - for every electrical watt into something like a processor, you'll get near-as-makes-no-difference a watt of heat out* - obeying the laws of physics, none of this energy is consumed/destroyed, it just changes from one form to another, doing some useful work in the process.

*You may get negligible amounts of energy radiated as e.g. radio waves if you're being extremely pedantic.

[Chaos27]

Ok I think I understand. So in simple terms if HBM does the same speed as GDDR5 for less watts/energy? and it does more as it seems to be the new better technology?

[scaryjim]

Ok I think I understand. So in simple terms if HBM does the same speed as GDDR5 for less watts/energy? and it does more as it seems to be the new better technology?

Hmm, not exactly. The technologies are very different and you can't really compare them directly like that - it's not simple, so trying to put it in simple terms won't really work!

Basically, HBM generally provides more bandwidth whilst using less energy, but (currently) at greater cost. But bandwidth doesn't necessarily equate directly to speed, as bandwidth is only one aspect of memory performance, and more bandwidth isn't always beneficial.

It's probably worth taking some time to read up more on memory technology in general, and the differences between (G)DDR3/4/5 (which are all superficially similar) and HBM (which is quite different), as it's rather difficult to explain/discuss these topics in very simple terms.

[watercooled]

HBM uses a wider but less highly-clocked bus than GDDR5. Since power grows faster than linearly with clock speed for things like this, you can save power by using more but individually slower data paths, if that makes sense? But like scaryjim says it's not always as simple as that and there can be trade-offs with latency.

This might be a good place to start reading: http://www.anandtech.com/show/9266/amd-hbm-deep-dive

[TooNice]

Going by how often HBM2 comes up in comments though, I feel like it's been heralded as the next biiiiiig thing with many refusing to buy a new card with it. I'd be amused (*) if, after it comes out, it turns out that on cards with the same/similar architecture and specification, that the one with HBM2 yields an average improvement under 10%
(*) Okay, that might also depend a bit on whether I hold out for it myself

[watercooled]

It's a technology driven by demand; GPUs continue to need more and more memory bandwidth and ideally they want it without driving memory controller power, die space and pin count through the roof - HBM is one answer to that. If you can satisfy the memory demands with GDDR5/X then you'll see likely zero improvement with HBM - I think many people are seeing it as something it's not; it's not something that's going to magically improve performance simply by being present, although the power and die area savings it brings allow more of those budgets to be spent on other parts of the processor.

It's like FinFETs; the properties of planar transistors were getting problematic on smaller nodes and FinFETs were a way (and not the only way) of alleviating many of those problems. Hence why Sandy Bridge>Ivy Bridge didn't magically bring substantial performance improvements, just the usual micro-architectural improvements usually seen on a tick.

[Chaos27]

So its not a technology that going to magically improve on everything. Though it does use less power and it has more memory. But performance you might not notice it so much. Maybe not untill HBM 3 hehe.

Processors as you say are different. As the smaller the node (14nm we have now) the smaller it is the less heat it has and hence the less power it uses and the more space it has to be faster.

[kalniel]

So its not a technology that going to magically improve on everything. Though it does use less power and it has more memory. But performance you might not notice it so much. Maybe not untill HBM 3 hehe.

No difference between HBM2 or 3 with regard noticing - you will only notice an improvement in performance if you are either limited by your current bandwidth or you can use a space/power saving for something else that will increase performance. That is already the case in some market segments, but isn't the case in others.

Processors as you say are different. As the smaller the node (14nm we have now) the smaller it is the less heat it has and hence the less power it uses and the more space it has to be faster.

And yet processor speeds have hardly improved at all over the years.