@mrwang420 wrote:
One of the major hurdles for integrated graphics on cpu dies like the X3D chips is that they have to use system memory for their graphics. Which is slower then actual video memory. These add in cards featuring just video memory could also potentially be used to give gpus access to extra video memory as well without the need of buying a whole gpu. I have been suggesting this for 8 years now.
It is true that most system memory is slower than dedicated graphics memory. But on a discrete graphics card the GPU and dedicated memory are physically on the same board which allows for very fast transfer speeds. The memory bandwidth of my 7900XT for example is 800GB/s on a 320bit bus.
However your proposed solution to rely on the PCIe bus to transfer data between an add-in graphics memory card and the IGP in the CPU is flawed. The bandwidth of the PCIe bus isn't an improvement over that of system memory, even the latest PCIe 5.0 16x slot tops out at 64GB/s, whereas dual-channel DDR5-6000 memory bandwidth is 96GB/s. Both use a 128bit bus.
Here is a somewhat dated article (2018) comparing the throughput of PCIe 3.0 16x @ 15GB/s to that of DDR4-3200 @ 25GB/s, so the suggestion that this idea was relevant 8 years ago is likewise false.
Even if you overlook the bandwidth limitations of the PCIe bus, I would argue the primary hurdle of IGPs is not memory speed, it is die space. AMD's desktop "G" series APUs give up L3 cache (compared to the "X" series) to gain enough real estate to incorporate the IGP, thus sacrificing performance. And even the top 8700G has a paltry 12CU in comparison to a RX 7600 32CU for example. Dare I mention the RX 7600 has 288GB/s memory bandwidth as well?
Ryzen R7 5700X | B550 Gaming X | 2x16GB G.Skill 3600 | Radeon RX 7900XT
Ryzen R7 5700G | B550 Gaming X | 2x8GB G.Skill 4000 | Radeon Vega 8 IGP
Ryzen R5 5600 | B550 Gaming Edge | 4x8GB G.Skill 3600 | Radeon RX 6800XT