The partnership between Intel and nVidia could indicate upcoming desktop APUs.
The partnership between Intel and nVidia could indicate upcoming desktop APUs.
EPYC features a single IOD with several CCDs. A higher TDP serves no benefit without additional CPU cores, GPU, and an AM5 configuration that still allows ample power for larger GPUs and CPUs in a single package. The primary challenge remains the expense of combining multiple silicon dies efficiently. The existing IOD-CCD wiring is straightforward and economical, offering good value despite its technical limitations. To achieve what you're seeking, you'd need cutting-edge silicon interconnects from TSMC, similar to those Apple employs in their M Ultra series. This would enable APUs to handle such performance without relying on expensive HBM or soldered LPDDRX solutions. Nvidia delivers this level of technology at scale, making it feasible for mainstream APUs. Even scaled down, the performance-to-cost ratio remains poor. The A100 (2020), 1.5TB/s H100 PCIe, 2TB/s H100 SXM, 3.3TB/s B100/B200 SXM, and 2x 4.1TB/s B100 variants showcase the capabilities of current offerings.
Under a low defect assumption, AMD experiences around 700 X3D cache failures per 300mm wafer. With TSMC pricing wafers at $15,000, each cache die is valued at about $21.43, rounded to $20 for simplicity. The 7800X3D sells for $360, nearly $100 higher than the 7700X. Assuming no extra costs beyond raw silicon and no packaging complications, AMD gains $80 per 7800X3D compared to 7700X. This implies an extra profit of roughly 25%, which may seem significant.
I've built quite a few CPUs, yet it feels like I'm stuck with limited results. People think I understand how CPUs function, but they don't grasp the process of designing one, ensuring it functions properly, and scaling it to a multi-billion dollar enterprise. Just sitting there watching from afar.
It's risky to judge component prices by retail figures alone. It's like comparing the cost of raw materials to eating a meal at a fancy restaurant—there are many factors involved. Perhaps it's better to assess silicon expenses only (CCD, IOD, cache) versus the final product costs (7700X vs 7800X3D) and check if the scaling makes sense.