Against DFloat11 and ZipServ on an RTX 4080 SUPER

Glyd is not the only lossless format for model weights. DFloat11 and ZipServ also keep every weight bit for bit at about 11 bits instead of 16. This report runs them on the same 16 GB RTX 4080 SUPER; the wider picture is on related work.

Three lossless formats

Method	Bits a weight	Same model?
Glyd (2026)	10.9 / 12.0	Yes, bit for bit
DFloat11 (NeurIPS 2025)	10.9	Yes, bit for bit
ZipServ (ASPLOS 2026)	11.4	Yes, bit for bit
4-bit quantization (GGUF Q4, AWQ, GPTQ)	4 + scales	No, rounded: smallest by far, and its answers change

Bits a weight on Qwen3-8B: the whole model for Glyd’s two layouts and DFloat11, the matrices for ZipServ.

End to end against DFloat11

Qwen3-8B takes 16.38 GB in bf16, which leaves no room to run on a 16 GB card. Both DFloat11 and Glyd hold its weights in about 11.2 GB.

Sequences a step	DFloat11, tokens/s	Glyd, tokens/s	Glyd against DFloat11, ×
1	13.8	47.2	3.4
8	104.7	360.2	3.4
32	387.5	1,268.6	3.3
64	746.2	1,897.2	2.5

DFloat11’s log · Glyd’s log

Tokens per second at 1, 8, 32 and 64 sequences: DFloat11 13.8, 104.7, 387.5 and 746.2; Glyd 47.2, 360.2, 1,268.6 and 1,897.2.
Sequences a step	DFloat11	Glyd
1	13.8	47.2
8	104.7	360.2
32	387.5	1,268.6
64	746.2	1,897.2

Tokens per second, Qwen3-8B on an RTX 4080 SUPER, DFloat11 against Glyd at the same size. Higher is better; log scale. Sep 26, 2026.

The matrix multiply against ZipServ and cuBLAS

The comparison with ZipServ is product against product: Qwen3-8B’s layer 18 matrices, microseconds a call with the L2 cache flushed, as ZipServ times itself. Lower is better; Glyd’s time in bold where it is the fastest.

	Bits a weight	1 token, µs	16 tokens, µs	32 tokens, µs	64 tokens, µs
q_proj, o_proj (4096 × 4096)
cuBLAS, bf16	16.00	73	73	74	77
ZipServ	11.35	57	57	58	69
Glyd	10.81	56	58	62	73
k_proj (1024 × 4096)
cuBLAS, bf16	16.00	23	27	26	28
ZipServ	11.43	22	23	23	25
Glyd	10.83	19	21	24	30
gate_proj (12288 × 4096)
cuBLAS, bf16	16.00	175	204	235	225
ZipServ	11.35	151	152	155	162
Glyd	10.76	144	148	153	172
down_proj (4096 × 12288)
cuBLAS, bf16	16.00	176	208	228	214
ZipServ	11.35	153	154	155	187
Glyd	10.75	147	150	153	181

Glyd is the fastest of the three in 10 of the 16 cells, and ZipServ in the other 6. Glyd ran its smallest layout at 1 to 32 tokens and its 12-bit layout at 64. ZipServ’s log · Glyd’s log

Against bf16 on the same card

For scale, Glyd against bf16 itself on this GPU: Qwen2.5-7B-Instruct, GPU time a token after the prompt, with q, k, v and gate, up merged as vLLM runs them.

Sequences a step	bf16, ms	Glyd, ms	Change, %
1	21.97	16.52	−25
8	22.85	17.37	−24
32	26.19	18.80	−28
64	27.67	25.14	−9

The log

Method

End to end: tokens per second generating with Qwen3-8B, DFloat11 and Glyd on the same RTX 4080 SUPER.
Kernels: microseconds a call of each matrix product with the L2 cache flushed, as ZipServ times itself.
Against bf16: GPU time a token after the prompt, over 16 steps; q, k, v as one product and gate, up as another, for bf16 and Glyd alike.
Exactness: every matrix packed, unpacked and compared bit for bit.

The logs of every run are in the repository, under benchmarks/gpu; how the harness is run: the GPU docs.