VRAM Performance Differences between GDDR5, GDDR5X and HBM2?

[Kordanor]

Hey there,

 

I am wondering about the performance differences between the different RAM Types. Now of course dependent on what you do, "performance" is defined pretty differently. So for the purpose of this posting I will use performance as equal to "performance in the standard Firestrike Benchmark of 3DMark".

 

The reason for my quesion is mostly because I'd like to be able to have a sense of how much of a step GGDR5 to GDDR5X is (in the upcoming GTX 1070 and 1080) and also how much of a boost HBM2 will be in the follow up generation.

 

Mostly "better" VRAM also comes with an increase of other specs. So the GTX 1070 and 1080 will have more differences besides of the VRAM. But can anyone estimate on what difference it would make if you just exchanged upgraded the VRAM (and it's connections that it could actually be used). And what about HBM? Well, HBM2 is not used yet from what I know. But what is the estimated boost from HBM2 over HBM? If you leave out that you could put more RAM in the same space.

In this instance, would you compare the R9 380X (which has 4GB of RAM) to the R9 fury for example and say that the difference is basically the boost due to HBM Memory?

 

Of course there is hardly an exact answer. But I'd appreciate your best guesses and extrapolations from the know examples we have with current cards and data.

 

Will a cheap HBM2 card blow a top GDDR5X card away just because of the VRAM? Or is the boost and the role of this step rather limited and only plays a small role in overall performance boost?

[Fulgrim]

about tree fiddy

[Nacho Marco Segui]

 This big

|---------------------------------------|

[NumLock21]

I want to know how much of a gaming performance do I get, when I upgrade to Windows 11.

[Bajantechnician]

1 hour ago, Kordanor said:

Hey there,

 

I am wondering about the performance differences between the different RAM Types. Now of course dependent on what you do, "performance" is defined pretty differently. So for the purpose of this posting I will use performance as equal to "performance in the standard Firestrike Benchmark of 3DMark".

 

The reason for my quesion is mostly because I'd like to be able to have a sense of how much of a step GGDR5 to GDDR5X is (in the upcoming GTX 1070 and 1080) and also how much of a boost HBM2 will be in the follow up generation.

 

Mostly "better" VRAM also comes with an increase of other specs. So the GTX 1070 and 1080 will have more differences besides of the VRAM. But can anyone estimate on what difference it would make if you just exchanged upgraded the VRAM (and it's connections that it could actually be used). And what about HBM? Well, HBM2 is not used yet from what I know. But what is the estimated boost from HBM2 over HBM? If you leave out that you could put more RAM in the same space.

In this instance, would you compare the R9 380X (which has 4GB of RAM) to the R9 fury for example and say that the difference is basically the boost due to HBM Memory?

 

Of course there is hardly an exact answer. But I'd appreciate your best guesses and extrapolations from the know examples we have with current cards and data.

 

Will a cheap HBM2 card blow a top GDDR5X card away just because of the VRAM? Or is the boost and the role of this step rather limited and only plays a small role in overall performance boost?

unless you can tell the futuree, we do not know until the card comes out.

 

its like saying, if i eat cake now, will i get fatter in 5 years.

 

 

but yeah, i believe hbm2>hbm> gddr5x>gddr5

you don know.

you have to wait those 5 years

 

also, gpu die architecture makes a difference

[Bajantechnician]

1 hour ago, Kordanor said:

Hey there,

 

I am wondering about the performance differences between the different RAM Types. Now of course dependent on what you do, "performance" is defined pretty differently. So for the purpose of this posting I will use performance as equal to "performance in the standard Firestrike Benchmark of 3DMark".

 

The reason for my quesion is mostly because I'd like to be able to have a sense of how much of a step GGDR5 to GDDR5X is (in the upcoming GTX 1070 and 1080) and also how much of a boost HBM2 will be in the follow up generation.

 

Mostly "better" VRAM also comes with an increase of other specs. So the GTX 1070 and 1080 will have more differences besides of the VRAM. But can anyone estimate on what difference it would make if you just exchanged upgraded the VRAM (and it's connections that it could actually be used). And what about HBM? Well, HBM2 is not used yet from what I know. But what is the estimated boost from HBM2 over HBM? If you leave out that you could put more RAM in the same space.

In this instance, would you compare the R9 380X (which has 4GB of RAM) to the R9 fury for example and say that the difference is basically the boost due to HBM Memory?

 

Of course there is hardly an exact answer. But I'd appreciate your best guesses and extrapolations from the know examples we have with current cards and data.

 

Will a cheap HBM2 card blow a top GDDR5X card away just because of the VRAM? Or is the boost and the role of this step rather limited and only plays a small role in overall performance boost?

also, welcome to the forum

[Kordanor]

I don't quite see why you all tell "nobody knows, nobody can estimate, we don't have a crytal ball" and then everyone was disappointed when it became clear that the new pascal cards don't actually use HBM2.

[NumLock21]

Tesla P100 is Pascal and runs on HBM2. Why most are upset with HBM2 isn't used? HBM has a higher memory bandwidth compare to GDDR5. This makes everyone automatically assume HBM2 will be faster than GDDR5X, but the number on paper does not translate to real world performance.

Ask a question on some future product that don't even exist, and you will get stupid responses. Not because they're stupid, it's because your question is stupid to being with.

Create this same thread in another few years when AMD Vega and Nvidia Volta appears, then you might get more intelligent answers.

[Kordanor]

No, I don't want a solid answer. I want extrapolation as I just wrote in my opening post. And for those who are not familiar with this word: It means your estimate according to the facts which are available.

 

I mean if I asked "How do you think your favorite sports team will do in it's next match" you would probably all have an opionion according to which players play, how they played in the past and whatnot. And yeah...if your guess happens to be correct this would be far more luck based than guessing in this thread (except if you have no idea what you are talking about).

 

For example something intersting I found on WCCFTECH:

http://wccftech.com/micron-gddr5x-memory-analysis-nvidia-pascal-graphic-card/

 

One of the examples given is DDR3 to DDR4, which also happens to be a good approximate analogy to think of the GDDR5 to GDDR5X jump.

 

(Speculated 448 GB/s on the DDR5X) is pretty close to the 512 GB/s number that HBM1 currently gives. Ofcourse, HBM2 is a whole other ball game and runs circles around the performance advantages offered by GDDR5X.

Be aware that this article is already 6 months old.

 

Now I compared the Benchmark numbers of the Radeon R9 Fury which uses HBM and tried to find a card as similar as possible using GDDR5, which is the Radeon R9 380 from what I can tell. Both using same architecture, having same amount of VRAM and the clock rate isn't that different either.

The performance difference is, that the Fury Card is 50% faster. Now the question is how much of that is by having 8900 transistors instead of 5000? I have no idea.

 

Now if we'd say that the performance advantage of HBM(1) over GDDR5 is almost as high as the boost of GDDR5X over GDDR5 - and if the transistors in the beforementioned examples have not a big impact, then the exchange could lead to a 40-50% boost. And this could be the difference of GTX 1070 (GDDR5) and 1080 (GDDR5X) (there is a small transistor difference between them as well).

 

Of course I make no claim of being correct here. Also I have far less knowledge about the topics than some experts here (well...that was what I thought). So with some decent knowledge you should be able to make much better estimates than I did.

I mean I could even ask someone with no idea about hardware and he would probably be able to tell that new memory doesn't incrase the performance of the whole card by 500%. Why? Because it never did in the past.

[DXMember]

we don't know yet

 

faster memory doesn't imply faster graphics card, it will only be faster if the GPU was being held back by the slow memory in the first place

have you tried overclocking your GPU memory? that give a good estimate for GDDR5X as it's mostly just higher clocked

and for the love of god... don't compare Fury with 380... they have like like 40 compute units apart (380 having 28 CUs, Fury having 64CUs) that's where the performance difference comes from, not the HBM

in order for the new memory technology to be helpful and offer the expected performance boost, the GPUs themselves have to get a whole lot faster

 

the reason why people were booing nVidia for the new cards not having HBM2 is because they told us in advance that the new cards will be using HBM2 and it did the whole marketing around it for like half a year and now they just backpedal on what they said, it's like the 970 with 4GB all over again,

also been waiting on that in-driver ASync support for like a year

[ivan134]

Comparing fury to 380?

Ahahahahahahahahahahahahahahahahaha

[Kordanor]

Thanks DXMember. So if you say that this is where the difference is coming from - doesn't this mean that the HBM doesn't have an impact at all?

And even if it was marketed towards HBM2..or lets say speculated - if nobody is aware of performance difference, DDR5X could theoretically be even better than HBM2, right? I mean it's clear that it isn't but as long as the actual difference isn't known I don't see a point in being upset about it. For all we know it could be just as fast.

 

@Ivan: Make a better comparison. Not helping otherwise.