How does GDDR6 compare with HBM for these "non-graphics" applications? Suppose I was designing a high-end router, why would I pick GDDR6 over HBM and vice versa?
For a high end switch/router, HBM would be better suited due to lower latency. While HBM is more expensive due to interposer/EMIB, the high end market carries such a premium this isn't much of an issue.
I'd say it's unlikely except perhaps with future generations of Intel+Vega type products (and for most of those applications HBMs smaller footprint is probably a design winner). On the PS4 it was done because of sharing ram between the CPU and GPU vs separate address spaces for PC. Between the architectural differences that the OS would need, it being a soldered only product, and not needed for general CPU use (CPUs haven't been ram speed bottlenecked since P4) I don't see it happening.
(Ryzen/Epyc are faster with high speed ram, but only because an internal bus runs at ram speed; the faster ram itself isn't doing anything direct for it.)
Well we're always limited by memory throughput. Just as many transistors are spent on cache as there are spent on the core. If main memory was faster and lower latency that wouldn't be necessary. This is particularly useful for many core designs where keeping memory synchronized adds a lot of control complexity.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
14 Comments
Back to Article
r3loaded - Tuesday, January 23, 2018 - link
How does GDDR6 compare with HBM for these "non-graphics" applications? Suppose I was designing a high-end router, why would I pick GDDR6 over HBM and vice versa?Kevin G - Tuesday, January 23, 2018 - link
For a high end switch/router, HBM would be better suited due to lower latency. While HBM is more expensive due to interposer/EMIB, the high end market carries such a premium this isn't much of an issue.A5 - Tuesday, January 23, 2018 - link
Eh, networking vendors are looking to keep margins high whenever possible too. GDDR6 will have the key benefit of being cheaper than HBM.Pork@III - Tuesday, January 23, 2018 - link
I want GDDR6 will use for PC memory? Like GDDR5 is PS4... GDDR6 is just from several to several dozen times faster than ordinary DDR4/5.Pork@III - Tuesday, January 23, 2018 - link
Ouch! :D I hope GDDR6 to be used for PC memory!DanNeely - Tuesday, January 23, 2018 - link
I'd say it's unlikely except perhaps with future generations of Intel+Vega type products (and for most of those applications HBMs smaller footprint is probably a design winner). On the PS4 it was done because of sharing ram between the CPU and GPU vs separate address spaces for PC. Between the architectural differences that the OS would need, it being a soldered only product, and not needed for general CPU use (CPUs haven't been ram speed bottlenecked since P4) I don't see it happening.(Ryzen/Epyc are faster with high speed ram, but only because an internal bus runs at ram speed; the faster ram itself isn't doing anything direct for it.)
willis936 - Tuesday, January 23, 2018 - link
Well we're always limited by memory throughput. Just as many transistors are spent on cache as there are spent on the core. If main memory was faster and lower latency that wouldn't be necessary. This is particularly useful for many core designs where keeping memory synchronized adds a lot of control complexity.Lolimaster - Tuesday, January 23, 2018 - link
Is faster in BW but with higher latency, no that important for gpu's.Lolimaster - Tuesday, January 23, 2018 - link
For actual PC's is more important, low latency, density than BW.nathanddrews - Tuesday, January 23, 2018 - link
We'll never be able to buy GPUs again at this rate.webdoctors - Thursday, January 25, 2018 - link
GDDR6 and 7 and future memory will all be reserved for miners.Non-miners will be stuck with G5 for life.
Gastec - Thursday, January 25, 2018 - link
True that, especially if we die sometime in the next 5 years.