Two Sparks or Four: DeepSeek-1017 Builds a World in 73.7 or 49.7 Minutes
@niklaslenz_ai ran @bertholomusai’s TensorFold recipes; different quants mean the test does not isolate Spark count.
Four Sparks finished about 24 minutes sooner by our arithmetic, but the quant changed too.
@niklaslenz_ai ran @bertholomusai’s DeepSeek-V4.1-Flash-1017 recipes on two DGX Sparks and on four, with the same voxel-world prompt. Both ran in DeepSeek Harness 0.2.1-rc.2 with the recipe’s own sampling, thinking on, one prompt and one try.
| Measure | Two Sparks | Four Sparks |
|---|---|---|
| Quant and split | EXL3 2.5 bpw, TP2 | EXL3 4.0 bpw, TP4 |
| Recipe | v0.7.2 (commit 30a5bf9), Oct 24 | commit 8f84e7a, Oct 22 |
| Agent time | 73.7 min | 49.7 min |
| Tokens written | 138,675 | 121,707 |
| Browser check | 8.0/10 | 9.0/10 (older capture method) |
| His rating of the world | 5.5/10 | 7.5/10 |
| Decode, code | 70.5 tok/s | 82.8 tok/s |
| Decode, prose | 70.6 tok/s | 80.8 tok/s |
| Eight requests at once, total | 106.3 tok/s | 141.2 tok/s |
He flags the limits himself: “That is a different quant and a different recipe version, so this does not isolate the Spark count.” His four Sparks are cabled as a switchless ring. The two-Spark recipe wants RDMA GID index 5, which his Sparks lacked, so he gave the link ports IPv4 addresses.
On two Sparks he also measured 1,830 tok/s reading a 32K-token prompt and 17.9 s to the first token. The recipe runs 4 slots, so with eight requests the rest wait. In Mia’s sparkDash 2.50.0 it scored 93/100 on the tool eval.
He frames the question as 2 Sparks vs 4 Sparks, 12,000$ vs 24,000$, and asks: “Are the extra 12,000$ worth it?”
Sources
- @niklaslenz_ai on X · Oct 24, 2:30 PM PT
- @niklaslenz_ai on X · Oct 24, 1:39 PM PT
- @niklaslenz_ai on X · Oct 24, 2:00 PM PT