Samsung Unveils Shinebolt HBM3E Memory At Nearly 10Gbps And Blistering 32Gbps GDDR7

With very high per-chip memory bandwidth and stacking capabilities, HBM has become popular for GPUs and even high-end CPUs since JEDEC finalized the specification in 2013. NVIDIA Grace Hopper GH200 received an HBM3E variant, although the original version used HBM3. Intel also uses HBM2E for its Xeon Max Sapphire Rapids processors.

SK Hynix introduced the HBM3E to its partners in May, and Micron did the same in July. Samsung distinguishes its Shinebolt chips with significantly improved performance; While SK Hynix and Micron's HBM3E can reach speeds of 8 Gbps and 9.2 Gbps per pin, respectively, Samsung's HBM3E can reach 9.8 Gbps per pin. This is a huge overall throughput and a significant advantage over competitors.

While Samsung's announcement was clearly focused on the HBM3E, the company also announced its upcoming GDDR7 memory chips, which will reach speeds of up to 32Gbps on each chip. For reference, GDDR6X (made by Micron for NVIDIA) provides speeds of about 22 Gbps, while standard GDDR6 is 16 Gbps. GDDR6 has been around for a long time, debuting in 2018 on the NVIDIA GeForce RTX 20 series, and GDDR6X was released on the RTX 30 series. However, GDDR6X is three years old, so the next generation will offer an upgrade to GDDR7. GPUs will be important.

Samsung previously detailed additional technical specifications for its GDDR7 memory: it should be 20% more efficient than GDDR6 and provide memory bandwidth of about 1.5 terabytes per second in certain configurations. Similar speeds on consumer GPUs were previously only possible with HBM2, which could deliver over 1TB/s of throughput on Radeon VII in 2019. Even the GeForce RTX 4090 has slightly less memory bandwidth than the Radeon VII, but with GDDR7 it looks like the GPU can take advantage of previously offered HBM2 speeds.