Dicehunter
Resident Newb
I'm just curious but if cost wasn't an issue, Would HBM2 be the better option to have on GPU's rather than GDDR6 due to the much higher bandwidth of HBM2 ?
HBM2 vs GDDR6 ?
- Thread starter Dicehunter
- Start date
HBM is more compact and lower power, more efficient in many ways so yeah. Bandwidth now is in the same ballpark atm, 256bit GDDR6(Navi, RTX2070/2080) will soon have the same max bandwidth as 2-stacks of HBM2(Same 512GB/s as Vega1) . 384bit GDDR6 (2080Ti/Titan) will have the same max bandwidth as 3 stacks of HBM. 4 stacks of HBM2 as on RVII has a max of 1TB/s vs 384bit GDDR6's 0.75TB/s.
Theoretically you could make a 512-bit GDDR6 bus with the same bandwidth as 4-stacks of HBM2, it'd just be large, complex and very high power, so probably more expensive in the end because of the memory controller size.
So basically, if you want more than 768GB/s of bandwidth(Not very useful for games but great for many other applications) you realistically need HBM2 atm, otherwise you don't and it's probably not worth the cost, but it's still got other small benefits(Efficiency & size).
Theoretically you could make a 512-bit GDDR6 bus with the same bandwidth as 4-stacks of HBM2, it'd just be large, complex and very high power, so probably more expensive in the end because of the memory controller size.
So basically, if you want more than 768GB/s of bandwidth(Not very useful for games but great for many other applications) you realistically need HBM2 atm, otherwise you don't and it's probably not worth the cost, but it's still got other small benefits(Efficiency & size).
Last edited:
WYP
News Guru
If pricing wasn't an issue HBM2 had a lot of advantages. It is widely considered to use less power than GDDR6, and the latest HBM2 chips from Samsung can run a lot faster than the HBM2 that we have seen on GPUs like the RX Vega 56.
Samsung's latest HBM2 is actually 2x faster than the HBM2 uses on the RX Vega 56 per chip and can come in stacks of up to 16GB per chip.
https://www.overclock3d.net/news/me...s_hbm2e_memory_packing_a_33_bandwidth_boost/1
So basically, an RX Vega 56, if made today with the best HBM2 memory, could offer 820GB/s of bandwidth and 32GB of total memory over two chips. For context, the RTX 2080 Ti offers 616GB/s of memory bandwidth.
Other advantages of HBM2 include the lower PCB area that's required for the chips and the lower memory latencies offered by HBM, due to their more direct connection to the GPU core.
There is a clear reason why many workstation/enterprise-grade graphics cards use HBM2, the problem is that for a consumer product that extra cost and manufacturing complexities aren't worth it.
Yeah if assuming 16Gbps dies all round (Though HBM2E stacks will likely come to market a bit later than GDDR6 dies near imminent availability) it's basically x2 the practical speed for the largest GDDR6 bus created today(384-bit/12 chips or multiples of) vs the largest HBM2 bus(4096-bit/4-stack).
Last edited:
NeverBackDown
AMD Enthusiast
Basically this. Mark nailed it on the head.
However, taking this question into context it still depends. If GDDR6 is fast enough for the architecture(as in the GPU die) then there are no performance differences. The only benefit is die area really. Depends on the architecture its being used with. Though like Mark said the Enterprise/WS envinroments are willing to pay the extra money for the small benefits so they can cram more GPUs in a single space.
That's nothing to do with the memory though, those two cars have essentially the same bandwidth & equal memory performance.
Kaapstad
Active member
Its got everything to do with memory performance and clockspeed, HBM2 is just too slow at low resolutions and high bandwidth won't help.
There is lots written about how great HBM memory is on the forums mostly by people who have never used the cards side by side with their GDDR6 counterparts.
Those aren't really side by side comparisons though, they're different chips with different archs and quite different memory compression schemes, as well as cache sizes.
HBM2 and GDDR6 both use very similar memory dies, they're just packaged in different ways and with different interface specs. HBM's magic is all in the reduced path length to the processor and wide bus width enabled by the ability to connect the chips straight via an interposer.
HBM2 is lower latency than GDDR6 regardless of load and actually more consistent at lower loads, it definitely has fewer benefits at lower load/resolution but to says it's slower doesn't really make sense, they both perform more or less equally at lower loads.
What you're seeing is much more likely the RTX Titan being a better baseline performer but taking a larger hit at higher loads, than the Titan V somehow becoming slower under lower load. You've no way of working out which is true simply from observing performance between the two cards, but by observing/understanding their memories differences it's fairly easy to work out which of the two otherwise equally valid explanations for what you're observing is the real one I'd say, because only one explanation is logical or has a viable cause(But both would exhibit exactly the same external symptoms).
HBM2 and GDDR6 both use very similar memory dies, they're just packaged in different ways and with different interface specs. HBM's magic is all in the reduced path length to the processor and wide bus width enabled by the ability to connect the chips straight via an interposer.
HBM2 is lower latency than GDDR6 regardless of load and actually more consistent at lower loads, it definitely has fewer benefits at lower load/resolution but to says it's slower doesn't really make sense, they both perform more or less equally at lower loads.
What you're seeing is much more likely the RTX Titan being a better baseline performer but taking a larger hit at higher loads, than the Titan V somehow becoming slower under lower load. You've no way of working out which is true simply from observing performance between the two cards, but by observing/understanding their memories differences it's fairly easy to work out which of the two otherwise equally valid explanations for what you're observing is the real one I'd say, because only one explanation is logical or has a viable cause(But both would exhibit exactly the same external symptoms).
Last edited:
Dicehunter
Resident Newb
It'll be interesting to see if the next big Titan card uses HBM2E or GDDR6 considering, As WYP posted, HBM2E can get up to around 1.6TB/s bandwidth.
