The development of high-performance computing today has become dependent on the ability of developers of specialized accelerators to use the fastest types of memory available on the market. According to some reports, SK hynix and NVIDIA are already working on integrating HBM4 chips directly onto the GPU.
Image source: SK hynix
Of course, such a processor is usually called “graphics” purely out of inertia, since NVIDIA has long separated its product lines for accelerating calculations and working with graphics. The use of HBM memory has become an attribute of the first type of NVIDIA products. In the case of the current range of accelerators from the latter brand, South Korean SK hynix was initially the only supplier of HBM3 memory, since only its products met NVIDIA’s strictest requirements both in terms of performance and the possibility of integration in the manufacturing process.
Now HBM3 and HBM3e chips are adjacent to the graphics processor on the same substrate, and the memory itself can have up to 12 tiers. As Tom’s Hardware reports, citing South Korean media, NVIDIA and SK hynix are already developing technology for integrating HBM4 memory chips directly on top of the GPU chip. Such proximity of the two types of these components will significantly increase the speed of information exchange, but will impose special requirements not only on the production process, but also on the cooling of the components.
Most likely, SK hynix will transfer its HBM4 chips to TSMC, which will simultaneously produce NVIDIA GPUs, and then this Taiwanese contractor will “splice” these chips without using an intermediate substrate. This technology is vaguely reminiscent of the production of AMD Ryzen processors with 3D V-Cache type memory integrated directly onto the chip with computing cores. It’s just that HBM4 memory will be slightly slower, but cheaper than the AMD cache used, and at the same time offer much greater capacity. HBM4 memory should increase the memory bus width from 1024 to 2048 bits, which further increases the feasibility of eliminating the intermediate substrate, since it would be too expensive to manufacture.
A significant problem will be the cooling of such a design, since GPUs without memory are becoming increasingly hotter, and HBM4 chips will also have to be cooled intensively, and both types of components will be located on top of each other. However, in the server segment, the use of advanced cooling methods using liquid or immersing the entire accelerator in a dielectric liquid justifies itself from the point of view of economic feasibility and complexity. It is also possible that HBM4 chips and associated graphics processors will be produced, if not using the same technical process, then at least according to similar technological standards. From this point of view, the requirements for the production process will become almost identical for both the logic part of the accelerators and the memory.
If you notice an error, select it with the mouse and press CTRL+ENTER.