Sadly I think you are right. While commendable AMD has always pushed higher memory capacities to the mainstream, their focus on memory bandwidth has never really paid off, and at a huge expense to die area for the larger memory controller, and obviously an energy efficiency deficit. This is why 3-channel memory was dropped in favor of reversion back to two channel with the Intel X58 chipset. It would be years before we would move beyond two channel again, and even then - quad channel never became mainstream.
The reason is simple. Even on single channel, Intel CPU’s especially show extraordinary memory performance. The controller is well optimized and cache hit rates are high. Likewise, Nvidia using excellent compression with optimized caches makes high memory bandwidth unnecessary.
SISD benefits greatly from caching and ILP. SIMD doesn’t need to run ILP to keep execution units busy so it chews through memory bandwidth by comparison. There is also quick diminishing returns on GPU cache size. GPUs have 20x the memory bandwidth of CPUs for a good reason: they use it.
Somewhat related to the subject of compression... adaptive resolution is by far the best graphics technology I have ever seen. Render at 1800p, drop down to 1400p when below the target framerate, and upscale everything to 4k. No need to buy the highest-end graphics card anymore. If we had adaptive resolution when Far Cry 1 came out, there would have been no market for the 6800, just use a 6600.
Combine with checkerboarding for console, which is impressive in its own right by NEAR-HALVING the workload. So render at half 1800p every other frame (equivalent of about 2300*1300 pixels, so 1.44x 1080p, not 4.0x) and get a generated 4k image.
Radeon VII has double the bandwidth for the same price but it doesn't really help performance at least in games. I think there has been more focus on effectively utilizing bandwidth because making the buses wider can get really expensive.
Hard to say . . . GDDR6 has a good deal of *theoretical* bandwidth on the table, there is the economical 'ghetto-HBM2' from Sammy, and HBM3 in the short-term.
We are likely to hear about Radeon **Navi-Instinct** pro cards this quarter, in addition to a Titan/Ampere 7nm HPC update. I'm thinking the trend will continue toward more efficient 'wider' bandwidth and advances in compression algorhtms, too.
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
Cellar Door - Tuesday, July 23, 2019 - link
The delta compression used by Nvidia is loseless.notashill - Tuesday, July 23, 2019 - link
If memory bandwidth was "the" bottleneck then the Radeon VII would be the fastest consumer level GPU on the market by an enormous margin.Samus - Tuesday, July 23, 2019 - link
Sadly I think you are right. While commendable AMD has always pushed higher memory capacities to the mainstream, their focus on memory bandwidth has never really paid off, and at a huge expense to die area for the larger memory controller, and obviously an energy efficiency deficit. This is why 3-channel memory was dropped in favor of reversion back to two channel with the Intel X58 chipset. It would be years before we would move beyond two channel again, and even then - quad channel never became mainstream.The reason is simple. Even on single channel, Intel CPU’s especially show extraordinary memory performance. The controller is well optimized and cache hit rates are high. Likewise, Nvidia using excellent compression with optimized caches makes high memory bandwidth unnecessary.
willis936 - Tuesday, July 23, 2019 - link
SISD benefits greatly from caching and ILP. SIMD doesn’t need to run ILP to keep execution units busy so it chews through memory bandwidth by comparison. There is also quick diminishing returns on GPU cache size. GPUs have 20x the memory bandwidth of CPUs for a good reason: they use it.flyingpants265 - Monday, July 29, 2019 - link
Somewhat related to the subject of compression... adaptive resolution is by far the best graphics technology I have ever seen. Render at 1800p, drop down to 1400p when below the target framerate, and upscale everything to 4k. No need to buy the highest-end graphics card anymore. If we had adaptive resolution when Far Cry 1 came out, there would have been no market for the 6800, just use a 6600.Combine with checkerboarding for console, which is impressive in its own right by NEAR-HALVING the workload. So render at half 1800p every other frame (equivalent of about 2300*1300 pixels, so 1.44x 1080p, not 4.0x) and get a generated 4k image.
notashill - Tuesday, July 23, 2019 - link
Radeon VII has double the bandwidth for the same price but it doesn't really help performance at least in games. I think there has been more focus on effectively utilizing bandwidth because making the buses wider can get really expensive.Smell This - Tuesday, July 23, 2019 - link
Hard to say . . .GDDR6 has a good deal of *theoretical* bandwidth on the table, there is the economical 'ghetto-HBM2' from Sammy, and HBM3 in the short-term.
We are likely to hear about Radeon **Navi-Instinct** pro cards this quarter, in addition to a Titan/Ampere 7nm HPC update. I'm thinking the trend will continue toward more efficient 'wider' bandwidth and advances in compression algorhtms, too.
wr3zzz - Tuesday, July 23, 2019 - link
How are these new cards draw so much more power than GTX980 under load yet have lower load temperature and noise? Are the new fans that good?Ryan Smith - Tuesday, July 23, 2019 - link
Blower versus open air (axial) cooler.Betonmischer - Tuesday, July 23, 2019 - link
Absolutely, if you compare against the reference blower that Nvidia used prior to RTX 20.