What is the real performance benefit to HBM?

[Vogner16]

Reading a lot of information out there and it seems that nobody really has done a good comparative analysis of how HBM compares to GDDR5 in anything other than capacity physical size and total bandwidth. there are more factors to gaming memory than that alone.

I simply find it hard to believe that a gpu with equal to more floating point perf and higher bandwith gets consistently slower performance than that of the 980ti.

Does HBM inflict a really poor CL timing? Are there other side effects to using HBM that degrade performance?

I don't think there is any doubt that HBM is the future, as it gives us a path to smaller and faster, the trend that the computer industry has taken for the last decades, but was amd too quick in implementing a product not yet ready?

What do you think (or hopefully know) about this topic?

Cheers!

 

[uglyduckling81]

The difference is the chipset architecture. The Fury and the 980ti are completely different. Just because of Mhz or the amount of RAM does not equate to more performance.
It's the same reason a Intel Core chip outperformed a P4 chip despite being 2Ghz slower on release.
Maxwell is just a better and more efficient design.
That's not to say HBM is doing nothing, it may well be making up the gap somewhat so they are almost on a level pegging.

 

[Guest]

nvidia has much more experience making gpus with huge die. so amd, which is new in the area of huge gpus will need more experience before it can be as efficient in performance. i doubt it's hbm's lacking that is restricting furyX.

 

[Vogner16]

That's the conclusion i'm getting from the vast web. If only somebody had some benchmarks of a Fury X on Gddr5 to prove that the HBM isn't really helping or maybe vice versa to show that the HBM is giving amd a huge performance boost. I think there has a lot to be said for latency however. does anyone have a good breakdown of hbm vs gddr5? ive seen all the amd slides with bandwith and capacity.

Also if I look at 3dmark records I see hundreds of people using 980ti's on LN2 trying to break the record but almost none with Fury or titan. saw one guy with fury (the only guy in top 100) get place 22, and the Titan x despite having more cores is only seen at place 15 or so to a swarm of 980ti's. why try to set records with the second best gpu? isn't the titan faster? This really supports your points regarding fury is a poorly designed arch compared to Maxwell but it does come really close.

What are your thoughts to HBM in the future did amd jump the gun here or is NVidia behind and lucky that there uarch is really good.

 

[Guest]

hbm is helping with efficiency and size. i don't know about its latency. it shouldn't be bad, at least for hbm2 because that's what both the companies have chosen for their next generation. maybe fiji is a conscious decision of amd to gain experience in making huge gpu, improving efficiency and using hbm. the fact that amd managed to improve performance per watt in their first big gpu that is comparable to nvidia's biggest is an achievement in itself and a good experience. nvidia has been making huge gpus for a long time. it's no surprise that there architecture is more mature. after all if you make big gpu, you must know how to prevent it from being a heater. my view is that amd needed fiji but nvidia didn't need its equivalent. fiji is a step for amd to make artic island gpus better.