You guys really need a "corrections" link so the comments section isn't full of people pointing out typos and malapropisms (I'm guilty of the latter myself, though).
496 GB/s for $700. I'm curious to see a retrospective of GPU memory bandwidth vs. cost over the last ten years. It feels like it's really sat still compared to transistor count. Are GPU caches getting bigger? Even then there is little that can be done about the main memory bandwidth requirements of SIMD workloads. We have faster interconnects yet the buses are staying the same or getting smaller.
Bandwidth matters less now than it did many years ago thanks different types of compression being used. You can fit more data into the same amount of bandwidth now than you could years ago.
Lossy compression isn't free. At some point the user will say "this looks bad". If that wasn't the case then why not compress every 64x64 tile to 1 KB? It's dependent on the data's entropy and many textures are high entropy. It's nice to have tuneable control over a soft cap, but it isn't a magic bullet that makes things better. Lossless compression would be bad in this application. No one should make a system that imposes a maximum allowed entropy on artists.
Memory bandwidth always has and remains to be the bottleneck of SIMD systems.
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
111 Comments
View All Comments
willis936 - Tuesday, July 23, 2019 - link
I think there is an error on the first page comparison table: 2080 Ti memory clock.Also first I guess.
willis936 - Tuesday, July 23, 2019 - link
Also the last paragraph of the conclusion should have "barring" rather than "baring".Ryan Smith - Tuesday, July 23, 2019 - link
This is what happens when you get overeager with copying & pasting... Thanks!extide - Tuesday, July 23, 2019 - link
and you say ending the bundle when I think you mean extendingRyan Smith - Tuesday, July 23, 2019 - link
The fault with that one lies solely with Word!boozed - Tuesday, July 23, 2019 - link
You guys really need a "corrections" link so the comments section isn't full of people pointing out typos and malapropisms (I'm guilty of the latter myself, though).Cheers for the review
RSAUser - Wednesday, July 24, 2019 - link
They rather need a grammarly subscription.willis936 - Tuesday, July 23, 2019 - link
496 GB/s for $700. I'm curious to see a retrospective of GPU memory bandwidth vs. cost over the last ten years. It feels like it's really sat still compared to transistor count. Are GPU caches getting bigger? Even then there is little that can be done about the main memory bandwidth requirements of SIMD workloads. We have faster interconnects yet the buses are staying the same or getting smaller.Stuka87 - Tuesday, July 23, 2019 - link
Bandwidth matters less now than it did many years ago thanks different types of compression being used. You can fit more data into the same amount of bandwidth now than you could years ago.willis936 - Tuesday, July 23, 2019 - link
Lossy compression isn't free. At some point the user will say "this looks bad". If that wasn't the case then why not compress every 64x64 tile to 1 KB? It's dependent on the data's entropy and many textures are high entropy. It's nice to have tuneable control over a soft cap, but it isn't a magic bullet that makes things better.Lossless compression would be bad in this application. No one should make a system that imposes a maximum allowed entropy on artists.
Memory bandwidth always has and remains to be the bottleneck of SIMD systems.