JEDEC Updates The High Bandwidth Memory Standard (HBM) up to 24 GB per stack
Click here to post a comment for JEDEC Updates The High Bandwidth Memory Standard (HBM) up to 24 GB per stack on our message forum
DeskStar
Why is HBM still something...?? It obviously isn't for consumers as in our level of use....is it??
I just keep thinking about AMD's sorry offering of 4gb's when 6 was the standard at the time. Hence those cards got squeezed out of relativity faster than you can fart!!
Denial
It's still "something" because as you hint in the second sentence it's still useful in non-consumer workloads.
AMD used it on consumer cards because splitting an architecture has a massive upfront cost and AMD has relatively little cashflow. So they designed one card to target two markets and that card has to be competitive in both so HBM is required.
rl66
It is the same as any new tech, i remember all the "why there is still DDR3 motherboard as DDR2 are faster", "is GDDR5 still alive? DDR3 is less expensive and do the same work"...
HBM is a great idea, the main problem is the price to have it on chip... It's for that we get GDDR6 on most card.
HBM on NVidia and AMD GPU isn't clearly well exploited yet, but maybe in the future... 🙂
TLD LARS
HBM is superior at almost everything compared to GDDR, only the price is hurting the adoption.
The Fury and Vega cards really shine at some workloads and HBM makes allot of interesting form factors possible, because the packaging is much smaller and easier to cool.
Fox2232
It is future. You may not like it or understand, but it is future. Take memory bandwidth limitation in cellphone. How many and how fast LPDDR chips you can fit in cellphone?
One stack of HBM... Win at Capacity, Power consumption, bandwidth per mm^2. And price is not worst too if right balance of parameters is used as cellphone does not need top memory density, nor top bandwidth. Today 4GB stack at 128GB/s is more than enough, and that is perfectly fine priced when you consider pricing of high end cellphones.
Next are high performance tablets, Notebooks. One day... small form factor APUs.
As for Fury X. In all that time I had it, it never had VRAM related stutter. I had a lot of stutter in last year of having i5@4.5GHz as CPU was very often 100% under load (had to use fps limiter in RTSS to get it under control). But moment I moved to 2700X, everything was flawless.
Game VRAM allocation does not mean it needs all that data per frame.
Take GTX Titan X with 12GB of VRAM and 336GB/s bandwidth. It takes 0.036s to load all that data through that bus. If GPU processing took absolutely no time that card would render 28 frames per second if each frame required all 12 GB. If 8GB was required for each frame, then you would end up with 42fps.
6GB data loaded in 0.018s that would result in 56fps.
Actually loading 4GB of data would take 0.012s, 84fps at best while ignoring that GPU has to actually process that data. Most of data in VRAM is there just to prevent caching based micro freezes and not actually needed.
Ricepudding
It's a shame that the top end cards don't at least use it, then again the Titan RTX won't be using it this time around either, figured it would have considering the cost of the damn thing!
HBM sounds lovely as it always did, but it's been years now since it came out and still waiting for it to hit other parts of the market beyond the industrial non-consumer level.
As someone else stated this inside tablets and phones would be amazing due to the form factor
