Atlas vs BF16 Qwen3.6 27B Report

Headline

BF16 has the stronger prompt-processing lane and lower loaded memory. Atlas reports higher decode throughput in the exact-token speed sweep, but BF16 wins the tested end-to-end speed prompts because prompt processing dominates these contexts.

324.02 vs 163.694k PP tok/s, BF16 compared with Atlas.

15.67 vs 11.304k TG tok/s, Atlas compared with BF16.

64.96 vs 98.27 GiBLoaded memory, BF16 compared with Atlas.

87 vs 82HermesAgent-20 score, BF16 compared with Atlas.

Selected BF16 MTP setting is nmax6. Atlas MTP setting is num_drafts=1 with K2 verification.

Runtime Profiles

Field	Atlas	BF16
Runtime	Atlas Spark	llama.cpp ROCm
Model lane	`unsloth/Qwen3.6-27B-NVFP4`	`unsloth/Qwen3.6-27B-MTP-BF16-GGUF`
Main weight format	NVFP4	BF16 GGUF
MTP	K2, `num_drafts=1`	`draft-mtp`, selected `nmax6`
KV cache	bf16	f16/f16
Batch shape	single-stream, batch size 1	single-stream

Charts

Four-model score, wall time, and TG

Model	Hardware lane	Hermes score	Wall time	TG metric
Chadrock 27B MTP	Strix Halo	81/100	19.5 min	34.59 tok/s
Atlas	Strix Halo	82/100	39.7 min	16.63 tok/s
Qwen3.6 27B NVFP4 MTP	RTX 5090	60/100	4.5 min	190.43 tok/s
BF16	Strix Halo	87/100	42.4 min	11.30 tok/s

Fresh Chadrock values come from backend counters sampled during the official HermesAgent-20 scenario window: 184,881 prompt tokens and 14,686 generated tokens.

The Qwen3.6 27B NVFP4 MTP reference was run on an RTX 5090; its TG value is the captured context-speed gate result, not a Strix Halo result.

Chadrock metric	Value
HermesAgent-20 score	81/100
Wall time	19.5 min
Hermes-window TG	34.59 tok/s
Hermes-window PP	301.13 tok/s
Peak combined GPU memory	36.54 GiB
MTP setting	draft-mtp nmax4

PP and TG throughput

Wall time by context

BF16 4k MTP sweep

Loaded memory

Speed Side By Side

Context	Atlas PP	Atlas TG	Atlas wall	Atlas gen	BF16 PP	BF16 TG	BF16 wall	BF16 gen
4k	163.69	15.67	41.30 s	256	324.02	11.30	23.98 s	128
8k	155.44	16.63	68.04 s	256	327.02	10.27	37.52 s	128
Long	139.43 at 18k	15.26 at 18k	148.92 s	256	302.17 at 16k	9.39 at 16k	67.86 s	128

Atlas context sweep

Prompt tokens	TG tok/s	PP tok/s	TTFT ms	Wall s	Total tok/s wall	K2 acceptance sample
4096	15.67	163.69	25022.91	41.30	105.38	48.0%
8192	16.63	155.44	52702.64	68.04	124.17	49.0% / 66.0%
18432	15.26	139.43	132199.10	148.92	125.49	64.0%

BF16 context sweep

Prompt tokens	TG tok/s	PP tok/s	Wall s	Draft accepted	Draft generated	Accept rate
4096	11.30	324.02	23.98	96	103	93.2%
8192	10.27	327.02	37.52	93	113	82.3%
16384	9.39	302.17	67.86	88	92	95.7%

BF16 4k MTP Nmax Sweep

The selected BF16 setting is nmax6, which produced the best 4k wall time and the highest TG in this sweep.

Setting	TG tok/s	PP tok/s	Wall s	Draft accepted	Draft generated	Accept rate	ROCm used MiB
no MTP	4.15	362.20	42.17	-	-	-	64127
nmax1	6.37	320.05	32.89	56	58	96.6%	65249
nmax2	6.84	304.33	32.18	76	81	93.8%	65368
nmax4	9.64	329.52	25.72	89	96	92.7%	66110
nmax6	10.97	331.96	24.01	95	100	95.0%	66523
nmax8	9.20	300.91	27.54	96	113	85.0%	66711

HermesAgent-20 Side By Side

82/100Atlas score.

87/100BF16 score.

39.7 minAtlas aggregate wall time.

42.4 minBF16 aggregate wall time.

Scenario score comparison

Scenario	Atlas score	Atlas wall s	BF16 score	BF16 wall s
HA-01	100	21.843	100	23.704
HA-02	100	201.986	100	315.344
HA-03	100	10.915	100	31.812
HA-04	100	194.827	100	393.883
HA-05	100	124.944	100	135.560
HA-06	100	93.741	100	180.815
HA-07	30	292.600	100	164.992
HA-08	100	113.484	100	115.352
HA-09	30	179.254	100	222.446
HA-10	100	205.268	100	82.460
HA-11	50	49.523	100	75.768
HA-12	70	94.443	100	46.711
HA-13	100	61.880	100	235.906
HA-14	70	40.538	70	31.836
HA-15	100	100.206	100	63.399
HA-16	30	109.123	30	118.041
HA-17	70	121.259	20	84.938
HA-18	100	25.051	100	40.093
HA-19	85	213.333	20	88.681
HA-20	100	104.339	100	94.512

Evidence Basis

This public report is based on captured benchmark outputs from the Atlas exact-token speed sweep, the Atlas runtime log, the BF16 MTP speed sweep, the BF16 MTP nmax sweep, the fresh Chadrock Hermes-window MTP metrics, the RTX 5090 NVFP4 context speed gate, and HermesAgent-20 aggregate benchmark results.

Atlas exact-token speed sweepBF16 MTP speed sweepBF16 nmax sweepChadrock Hermes-window MTP metricsRTX 5090 NVFP4 context speedHermesAgent-20 aggregateROCm memory accounting