Samsungs New GDDR6W Graphics Memory Rivals HBM2
In the past, chip companies like AMD used High Bandwidth Memory (HBM) instead of GDDR to increase memory bandwidth for GPUs. This vertically stacked memory offers impressive bandwidth, but is expensive. AMD dropped it in favor of GDDR memory after the ill-fated R9 Fury and Vega GPUs. Now Samsung has developed a new type of GDDR6 memory, which is said to be faster than HBM without the need for an interprocessor. Samsung claims that GDDR6W is the first "next generation" DRAM technology and enables more realistic MetaVas experiences.
Samsung took the existing GDDR6 platform and built it with Fan-Out Wafer-Level Packaging (FOWLP). In this technology, the memory chip is installed on a silicon wafer instead of a printed circuit board (PCB). Redistribution layers are spread around the chip, allowing more contact and better heat dissipation. Memory chips are also dual. Samsung says this allows it to increase bandwidth and capacity like never before. Because there is no increase in model size, the partners can put GDDR6W into existing and future models without any modifications. This theoretically reduces production time and costs.
Samsung's wafer-level, fan-assisted package allows for smaller packages due to the absence of PCBs. (Credit: Samsung)
The new memory offers twice the GDDR6 I/O and bandwidth. Using the existing 24 Gb/s GDDR6 as an example, Samsung says the GDDR6W version has double the I/O because it has more interface points. It also doubles the capacity from 16GB to 32GB per chip. As mentioned above, the height of the FOWLP model is only 0.7mm, which is 36% smaller than the DDR case. Despite doubling the I/O and bandwidth, it claims to have the same thermal characteristics as current DDR6 models.
Samsung claims that these improvements allow the GDDR6W design to compete with HBM2. The second generation HBM2 has a bandwidth of 1.6 TB/s, while the GDDR6W is closer to 1.4 TB/s. However, this Samsung edition uses a 512-bit memory bus that uses 32 GB of memory, which is not available in current GPUs. The Nvidia RTX 4090 and Radeon RX 7900 XTX both have a 384-bit memory bus and only offer 24 GB of GDDR6 memory. While AMD uses GDDR6, Nvidia opted for the Micron-developed G6X variant. Both cards have 1 TB/s of memory bandwidth, so Samsung's offering is superior.
The big news here is that thanks to Samsung's chipset stack, you need half as many memory chips to get the same amount of memory as the current package. This can lead to a reduction in production costs. Overall, the maximum pen throughput of 22 Gb/s is very close to the GDDR6X's 21 Gb/s. So future gains are likely to be in memory rather than peak performance. You could say that no one needs a GPU with 48GB of memory, but maybe that will change when we play at 16K.
In terms of products, Samsung has announced that it will soon introduce GDDR6W in small devices such as laptops. It works with partners to embed it in AI and other accelerators. It's unclear whether AMD or Nvidia will adopt it, but if they do, it's likely in the future. This is simply because both companies make their current cards with GDDR6/X models, so we doubt they will trade until a new architecture arrives.
Now to read :
