NVIDIA Makes Eight-Spark Clusters Official; Full GLM-5.3 Runs at 31 tok/s

One terabyte of pooled memory, a faster all-reduce for switched clusters, and a $4,999 64 GB Spark today.

Eight boxes, 1 TB, and the first official numbers for the full GLM-5.3 on Sparks.

NVIDIA announced official support for eight-node DGX Spark clusters at GTC Berlin, its technical blog says. A Spark has 128 GB of unified memory, four reach 512 GB and eight reach 1 TB. That opens models that would not start on four nodes, NVIDIA writes, including the full 753B GLM-5.3. @eTeknix calls it a strong boost for local AI. Support for eight nodes in NVIDIA Sync comes by the end of the year.

Users	Prefill total tok/s	Decode total tok/s	Decode per user tok/s
1	1,600	31	31
4	1,600	51	21
8	1,600	67	18

nvidia/GLM-5.3-NVFP4, about 470 GB, with MTP3; prefill at 8K input. Source: NVIDIA Technical Blog. Swipe the table sideways to see every column.

NVIDIA also ran GLM-5.3-Flash across eight nodes at roughly 80 tok/s in generation. To test the network alone, it timed a dense Qwen3.8-27B in NVFP4 without speculative decoding.

Sparks	Topology	Decode tok/s	Scaling
1	None	11.86	1.00x
2	Direct cable	21.71	1.838x
4	Ring	32.82	2.77x
4	Switch, RoCEnante	36.34	3.06x
8	Ring	40.11	3.38x
8	Switch, RoCEnante	52.58	4.43x

8,192 tokens in, 256 out. Source: NVIDIA Technical Blog. Swipe the table sideways to see every column.

RoCEnante is a Spark-tuned all-reduce, merged into FlashInfer, that works only on switch-connected clusters. It falls back to NCCL if its startup test fails. A four-box ring with no switch is the cheapest four-node setup, NVIDIA says.

Separately, DGX Spark 64GB goes on sale Friday, October 23, from Acer, ASUS, Dell, Gigabyte, HP and MSI, starting at $4,999, NVIDIA’s blog says. It keeps the GB10 chip and supports models up to 100 billion parameters; two units connect over ConnectX-7.

NVIDIA Makes Eight-Spark Clusters Official; Full GLM-5.3 Runs at 31 tok/s

Sources