DFSX rolled out a different take on AI accelerators: rather than chase ever-smaller process nodes, they pushed memory bandwidth instead. Their TY64 supernode, built around forthcoming DF2000 chips, is claimed to reach up to 960 TB/s of memory bandwidth—compare that to NVIDIA’s GB200 NVL72 at 576 TB/s.
They stick with 14nm production but change geometry. Memory sits above compute blocks in a stacked, 3D arrangement, linked through a hybrid 3D interconnect that forgoes microbumps. The result, according to DFSX, is lower latency and faster data movement (i.e., less idle time waiting on memory).
Why this focus? The company points to the “memory wall” as the real choke point: accelerators often wait for data, not for raw math. So DF2000 prioritizes bandwidth. That comes with trade-offs—BF16 throughput is much lower than NVIDIA’s: 64 PFLOPS vs. 360 PFLOPS on the GB200 NVL72.
Looking forward, DFSX teases a DF3000. Target figures: 20 TB/s per processor and 1280 TB/s in a TY64 SuperNode. If achieved, that would bring them close to NVIDIA’s Vera Rubin NVL72 numbers, despite using the older 14nm node.
A small aside: this is a hardware gamble—prioritizing memory paths over die shrinks. It might pay off in certain workloads, or it might reveal limits in others. Your mileage will vary depending on the models and data patterns you run.