When will AMD's GPU not memory bandwidth-starved?

[Totally Average Gameplay]

Since DDR5 is expected to come in the near future and AMD with their GPU archs being improved that will lead to future AM5 and future APUs. I wonder, what kind of mem bandwidth would satisfy the age-old idea that AMD's GPUs were limited in bandwidth. Because in desktop APUs, they were limited performance because of the RAM's bandwidth ( and also due to being iGPUs).

Navi with the latest GDDR6 memory has still.

Heck, even HBM2 could not satisfy Vega's bandwidth.

 

How can AMD solve this?

And what kind of memory would solve this problem? I still think DDR5 will still see bandwidth-starved behavior for APUs in the future socket.

[DoctorNick]

4 minutes ago, Totally Average Gameplay said:

Since DDR5 is expected to come in the near future and AMD with their GPU archs being improved that will lead to future AM5 and future APUs. I wonder, what kind of mem bandwidth would satisfy the age-old idea that AMD's GPUs were limited in bandwidth. Because in desktop APUs, they were limited performance because of the RAM's bandwidth ( and also due to being iGPUs).

Navi with the latest GDDR6 memory has still.

Heck, even HBM2 could not satisfy Vega's bandwidth.

 

How can AMD solve this?

And what kind of memory would solve this problem? I still think DDR5 will still see bandwidth-starved behavior for APUs in the future socket.

The first part may be right, that APU's get limited by bandwidth from DDR memory but Vega isn't limited by HBM2..

[GoldenLag]

never, because they allways just go for what they can feed. 

 

the 2080ti is as fast as it is because thats what its bus can feed. 

 

we are at the point where HBM might be the solution for dramatically faster GPUs and APUs. GDDR6 is limited with the nightmare that is 512-bit memmory busses and CPUs dont want a big memmory controller due to the power costs associated with them. 

 

 

mobile graphics will be rather interesting and i wouldnt be suprised if we see a 60% jump in performance just with what ddr5 gives in bandwidth 

[Totally Average Gameplay]

1 hour ago, DoctorNick said:

Vega isn't limited by HBM2..

No. When you overclock Vega, the core overclock provide only small gains but higher in power draw. Mem OC has more gains and lesser power when touching the Memory alone.

[Mira Yurizaki]

I don't know if AMD's doing the same thing or not, but NVIDIA invested a lot in fine tuning execution unit scheduling, memory organization, and memory compression.

 

Another thing that may stand out is AMD GPUs have less L2 cache than NVIDIA's GPUs. Navi and Vega had up to 4MB, Polaris had up to 2MB (implied by AMD saying Vega has " twice the size of the L2 cache in previous high-end AMD GPUs"). Turing has 6MB, whereas Pascal had up to 3MB (https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/technologies/turing-architecture/NVIDIA-Turing-Architecture-Whitepaper.pdf). I recall one of the major standout features of Maxwell was a rather large L2 cache for the time of 2MB, when the previous generation had 512KB (https://www.microway.com/download/whitepaper/NVIDIA_Maxwell_GM204_Architecture_Whitepaper.pdf).

 

So it's not that I think AMD GPUs are more hungry for data than NVIDIA's GPUs (I have said AMD GPUs are hungry in the past though), but that AMD likely hasn't invested in the same fine tuning that NVIDIA has done.