Over the past decade, several well-established research directions have argued that “less is more”: Regularization theory shows that constraining models often improves generalization. Techniques such as weight decay, dropout, and early stopping deliberately reduce effective capacity—and frequentl…

Its about loss of information. FACT: you loose something that the llm was glad to have been learned from training: nuances. there is another thread concerning FP4, especially NVFP4. NVIDIA has a new star on the firmament: PPXZPXUT? ( after QAD, QAT and other quant aware techniques. And they ended u…

[image] fidecastro: Is there still hope for FP4 performance that’s commensurate with expectations? Yes, I’ve been working with NVIDIA directly on this, it’s finally getting some traction internally.

As is see, GLM 5.0 goes to A40B ! but misses 1T whats good. 750B could fit on 4 dgx spark in NVFP4 but A40B …. stresses memory bandwidth and this 4-bit recomputing in the shared memory. There are two fractions. first thinks 3B and 10B active weights are enough, others argue its too less to …

I did some preliminary work on this and as of today, here are the identified blockers for running GLM-5 in NVFP4 on 4x DGX Spark: GLM-5 NVFP4 on 4x DGX Spark — What’s Missing Estimated NVFP4 size: ~430–470 GB → fits in 4x 128 GB (~464 GB usable), tight but feasible. Current blockers: llm-compre…

[image] eugr: Yes, I’ve been working with NVIDIA directly on this, it’s finally getting some traction internally. Can you elaborate more on that? Share your vision and expectations. Thank you.

[image] kyrylo.gorbachov: Can you elaborate more on that? Share your vision and expectations. Thank you. It’s all in the early stages still. Will update once I see real progress.

[image] FP4 on DGX Spark — Why It Doesn't Scale Like You'd Expect DGX Spark / GB10 testing vllm int4 autoround Qwen3-Coder-30B-A3B — DGX Spark (SM121) Quant Spec GEMM Kernel tok/s (long) Math vs BF16 INT4 (W4A16) mma.sync EAGLE3 Marlin FP16 94.6 78% 3.…

Cost, power, heat, performance – they all go into the mix. The DGX Station if it ever arrives will have impressive memory bandwidth and a large quantity of GB but will cost much more than a brace of Sparks and ConnectX switches. I’ve been pleased that the Spark performance has been improving thank…

I have to insist that I never criticize here in a way that suggests bad motivation as the cause, but always as a constraint arising from what’s technically feasible, combined with market considerations. I think many in the forum agree that this is an interesting piece of hardware that NVIDIA has ass…

Agreed, end to end measures are the only way to go! You’re only as fast as your slowest component. People in general just discount things they are not familiar with, and granted not a lot of people have datacenter switching experience. There are also a couple gotchas here: The ASIC architecture …

Why 273 GB/s? Less Is More, Until It Isn’t

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10

flash3 March 27, 2026, 6:24am 69

TurboQuant frees up memory and bandwidth:

[https://forums.developer.nvidia.com/t/why-turboquant-saves-dgx-twice/]

Topic		Replies	Views
DGX Spark performance DGX Spark / GB10	49	6542	February 13, 2026
6x Spark setup DGX Spark / GB10	112	11162	April 25, 2026
FP4 on DGX Spark — Why It Doesn't Scale Like You'd Expect DGX Spark / GB10	213	7144	March 13, 2026
How to run GLM 4.7 on dual DGX Sparks with vLLM / mods support in spark-vllm-docker DGX Spark / GB10	27	4480	January 2, 2026
Two multi-node DGX Spark wins: RoCE 2× inference throughput + Qwen3.5-397B-A17B-NVFP4 serving (with SM121 CUTLASS patch) DGX Spark / GB10 Projects	4	972	April 16, 2026
I am EXTREMely disappointed with the current state of DGX Spark DGX Spark / GB10	91	20699	June 24, 2026
DGX Spark: The Sovereign AI Stack — Dual-Model Architecture for Local Inference DGX Spark / GB10 Projects docker , spark , llm	9	2108	February 13, 2026
NVIDIA folks -- where is this promised nvfp4 speedup? DGX Spark / GB10	27	2950	March 26, 2026
A Spark to beat M5 Ultra and a MegaSpark to beat 2x Rubin PRO 6000! DGX Spark / GB10 nemotron	50	2985	June 25, 2026
We unlocked NVFP4 on the DGX Spark: 20% faster than AWQ! DGX Spark / GB10	144	9373	March 14, 2026

Why 273 GB/s? Less Is More, Until It Isn’t

Related topics