Chinese GPU maker InnoSilicon has managed a remarkable feat. In November last year, it launched the Fenghua No.1 graphics card for desktops and data centers promising leadership compute performance and packaging technologies in one fell swoop. We now have more details on this SKU, confirming the specifications and various IPs used by it. The first and most important disclosure is with respect to memory. The Fenghua graphics card uses the same 21Gbps GDDR6X memory found on the NVIDIA RTX 3090 Ti and its successor, the RTX 4090. Along with the top-end Ampere parts, it’s the only GPU to feature ultra-high-speed GDDR6X memory.
Micron has confirmed on its website that it worked with Innosilicon to fine-tune the PAM4 signaling technology powering its GDDR6X memory IP. The two have co-developed the memory for bandwidth-hungry AI and neural networks:
When paired with a 256-bit bus, this results in a bandwidth of over 5TBits/s, 5x more than LPDDR4 and pretty much the same as HBM, albeit at a much lower cost. In addition to rocking cutting-edge graphics memory, this Chinese GPU also comes with a chiplet design. The individual dies are connected using the Innolink IP. Fenghua features all the latest I/O technologies, including PCIe Gen 4, Display Port 1.4, HDMI 2.1, etc.