AMD Radeon RX 9070 XT 16GB

Reported on Blog by Ataa Aldaghstani / ataadevs — posted 2026-08-17 Original report

Not reproduced by Bitworks

RX 9070 XT Windows Qwen3.8 coding-task MTP comparison

Method (as reported): Same repository-explanation task; q4_0 K/V, FA on, fit off; author averages with timing denominator unspecified.

• Reported GPU setup RX 9070 XT 16GB
• GPU count 1
• Operating system Windows (Windows)
• Engine llama.cpp
• Backend ROCm/HIP
• Model Qwen3.8-27B Q4_K_M and AD-IQ4_XS-IQ3_S (separate arms)
• Actual input input length not reported
• Full window tested no
• Reported context 32,096 Q4 arm; 64,096 mixed-quant arm; occupied input unreported.
• Simultaneous clients 1
• Server slots 1
• Topology single
• Speculative decoding MTP2
• CPU Ryzen 7 7800X3D
• Memory configuration DDR5; capacity unknown
• Host notes Display on another GPU.

• 24 tok/s — other (scope not stated) · agent session · 1 client
  Compared with 17 tok/s (Q4_K_M no MTP vs MTP2); reported difference: speculative decoding
  Author average; timing denominator unspecified; Q4_K_M at configured 32,096; repository-explanation task.
  This row: configured context: 32,096 tokens · quantisation: Unsloth Q4_K_M · speculative decoding: MTP2
• 42 tok/s — other (scope not stated) · agent session · 1 client
  Compared with 27.5 tok/s (AtomicChat mixed quant no MTP vs MTP2); reported difference: speculative decoding
  Author average; timing denominator unspecified; mixed quant at configured 64,096; repository-explanation task.
  This row: configured context: 64,096 tokens · quantisation: AtomicChat AD-IQ4_XS-IQ3_S · speculative decoding: MTP2

Not reported: model file checksum, system memory, Artifact/build digests, Occupied inputs and output counts, Timing denominator, Repetitions/variance, Objective task scoring

• Q4 arms required mid-task compaction.
• Mixed-quant MTP quality was judged weaker subjectively.
• Cross-quant comparison also changes context.
• Linked setup advice is not adopted or qualified.

Setup guide (external)

Reported on GitHub by 7269827-rgb Original report

Not reproduced by Bitworks

RX 9070 XT custom Qwen3.8 long-context state follow-up

Method (as reported): Post-publication shallow-context empty-/loaded-cache checks; q4_0 K/V and MTP2; state UNCLASSIFIED.

HEAD seen at retrieval: c7abad32e4f2 (not necessarily the commit the author measured)

• Reported GPU setup RX 9070 XT 16GB
• GPU count 1
• Operating system Windows (Windows)
• Engine patched llama.cpp bf0040e15 lineage; exact archived patch missing
• Backend Vulkan
• Model Qwen3.8-27B custom 2.72 BPW tensor mix (7269827-rgb)
• Configured context 262,144 tokens
• Actual input input length not reported
• Full window tested no
• Reported context 256K is capacity; reported samples were shallow-context, not full-window.
• Simultaneous clients 1
• Topology single
• Speculative decoding MTP2

• 24 tok/s — generation speed (scope not stated) · 1 client
  Shallow-context empty-cache sample 1/3; state UNCLASSIFIED; 256K configured capacity not occupied.
  This row: configured context: 262,144 tokens · quantisation: custom 2.72 BPW tensor mix · speculative decoding: MTP2
• 23 tok/s — generation speed (scope not stated) · 1 client
  Shallow-context empty-cache sample 2/3; state UNCLASSIFIED; 256K configured capacity not occupied.
  This row: configured context: 262,144 tokens · quantisation: custom 2.72 BPW tensor mix · speculative decoding: MTP2
• 25.8 tok/s — generation speed (scope not stated) · 1 client
  Shallow-context empty-cache sample 3/3; state UNCLASSIFIED; 256K configured capacity not occupied.
  This row: configured context: 262,144 tokens · quantisation: custom 2.72 BPW tensor mix · speculative decoding: MTP2
• 26 tok/s — generation speed (scope not stated) · 1 client
  Shallow-context loaded-cache sample 1/3; state UNCLASSIFIED; 256K configured capacity not occupied.
  This row: configured context: 262,144 tokens · quantisation: custom 2.72 BPW tensor mix · speculative decoding: MTP2
• 29.3 tok/s — generation speed (scope not stated) · 1 client
  Shallow-context loaded-cache sample 2/3; state UNCLASSIFIED; 256K configured capacity not occupied.
  This row: configured context: 262,144 tokens · quantisation: custom 2.72 BPW tensor mix · speculative decoding: MTP2
• 26.9 tok/s — generation speed (scope not stated) · 1 client
  Shallow-context loaded-cache sample 3/3; state UNCLASSIFIED; 256K configured capacity not occupied.
  This row: configured context: 262,144 tokens · quantisation: custom 2.72 BPW tensor mix · speculative decoding: MTP2

Not reported: CPU, system memory, Cause of unreproduced fast state, Exact original patch, General quality across tasks, Concurrency

• Faster campaign headlines did not reproduce reliably.
• 40K gate speeds used a different binary.
• 200K sustained figure is a per-problem proxy, not a timed floor.
• Do not transfer custom quant/attention quality to stock artifacts.

Setup guide (external)

Related link (external)

Related link (external)

Related link (external)

Reported on GitHub by MrLordCat · measured 2026-09-26 Original report

Not reproduced by Bitworks

Dual RX 9070 XT Qwen3.8 Windows custom-fork MTP report

Method (as reported): One cold run per cell; FP8 K/V, FA, batch 8192 / ubatch 1024; no warmup or cache reuse.

HEAD seen at retrieval: e648079e8656 (not necessarily the commit the author measured)

• Reported GPU setup 2 × RX 9070 XT 16GB
• GPU count 2
• Operating system Windows (Windows 11)
• Engine llama.cpp-rdna-lab fea1c3180 plus uncommitted FA change
• Backend ROCm/HIP 7.2
• Model Qwen3.8-27B UD-Q4_K_M
• Configured context 49,152 tokens
• Actual input 32,996 tokens
• Output 256 tokens
• Full window tested no
• Reported context One cold run per cell; configured window not fully occupied.
• Simultaneous clients 1
• Server slots 1
• Topology dual-split (ROCm1,ROCm0; layer split 1:1; no P2P; PCIe placement unknown.)
• Speculative decoding MTP n3
• CPU Ryzen 7 5800X3D
• System memory 64 GiB

• 43.75 tok/s — generation speed (per request) · API server · 1 client · input 32,996 tokens
  Compared with 25.74 tok/s (Same custom fork no MTP vs MTP n3); reported difference: speculative decoding
  Single cold L2 Windows HIP cell; 32,996 input / 256 output; custom fork and uncommitted FA change.
  This row: configured context: 49,152 tokens · speculative decoding: MTP n3
• 1799.29 tok/s — prompt speed (per request) · API server · 1 client · input 32,996 tokens
  Compared with 1879.06 tok/s (Same custom fork no MTP vs MTP n3); reported difference: speculative decoding
  Same one-run L2 Windows HIP comparison; reported prompt processing declined.
  This row: configured context: 49,152 tokens · speculative decoding: MTP n3

Not reported: model file checksum, Exact executable/artifact hashes, Uncommitted change identity, Windows raw receipts unavailable in pinned tree, Repeated-run variance, Concurrent throughput

• Specialized fork, not stock FastLLM.
• Do not infer dual-card scaling without a single-card control.
• Windows/Linux comparisons also change driver and P2P.
• W24/W26 kernel gains belong to separate Linux research lanes.

Setup guide (external)

Related link (external)

Reported on GitHub by kr4ckhe4d · measured 2026-10-02 Original report

Not reproduced by Bitworks

Qwen3.8 llama.cpp build comparison on RX 9070 XT

Method (as reported): ABBA build order; two server loads per build, two deep measures/load plus discarded warmup; Q8 KV, batch 2048, microbatch 1024, FA on.

HEAD seen at retrieval: c18f1bc96ce3 (not necessarily the commit the author measured)

• Reported GPU setup RX 9070 XT
• GPU count 1
• Operating system Linux (CachyOS)
• Engine llama.cpp HIP build 11345 (a868c3e3c)
• Backend ROCm
• Model Qwen3.8-27B IQ4_XS-v3
• Configured context 32,768 tokens
• Actual input 16,852 tokens
• Output 700 tokens
• Full window tested no
• Simultaneous clients 1
• Server slots 1
• Workers 1
• Topology single
• Prefix cache off
• CPU Ryzen 7 9800X3D
• Memory configuration 2 x 16 GiB DDR5-6000
• System memory 32 GiB
• Memory speed 6000 MT/s

• 28.41 tok/s — generation speed (per request) · 1 client · input 16,852 tokens
  Compared with 26.84 tok/s (baseline build 7c35571e5; reported a868c3e3c; no MTP); reported difference: engine build
  Deep-input build comparison; Q8_0 KV, flash attention on.
  This row: configured context: 32,768 tokens
• 1193 tok/s — prompt speed (per request) · 1 client · input 16,852 tokens
  Compared with 959 tok/s (build 7c35571e5 vs a868c3e3c; no MTP); reported difference: engine build
  Same source comparison; prompt cache off.
  This row: configured context: 32,768 tokens

Not reported: exact model digest, serving driver identity, concurrent behavior

• Single Linux host; not Windows or non-XT RX 9070 evidence.
• The source's Q3 MTP row uses a different quant and is not included.
• Build changes were not bisected to an individual kernel.

Related link (external)

Reported on GitHub by Niko1221 / Strata maintainers Original report

Not reproduced by Bitworks Project documentation — capability, not a measurement

Strata AMD HIP guide

Method (as reported): Maintainer documents Windows bundled HIP runtime, per-card checks and Linux layer split; no discrete AMD Windows model-load validation.

HEAD seen at retrieval: 6f32ec070f23 (not necessarily the commit the author measured)

• Reported GPU setup Windows: one GPU per model; Linux: selectable layer split
• GPU count not reported
• Operating system Windows and Linux (Windows HIP and Linux HIP)
• Engine Strata HIP
• Backend HIP
• Model Qwen3.8-Flash-Next pack
• Actual input input length not reported
• Topology not reported — Windows one GPU per model; Linux selectable layer split

Not reported: model quantisation, configured context, actual input length, simultaneous clients, topology, CPU, system memory, release archive digest and revision, Windows discrete-card load, Windows dual-card execution, quality and memory curves

• Maintainer says ready-made Windows archive has not run a model on a discrete AMD card in documented validation.
• Capability documentation, not a measured compatibility or performance result.
• RX 9060 XT memory variant is unspecified in the guide, so neither variant is mapped.

Setup guide (external)