The RX 9070 XT also advertises 16 GB of GDDR6 memory, but it is not interchangeable with the non-XT RX 9070 or the RDNA 3 RX 7800 XT. The same pinned Qwen3.5-9B Q8 candidate across those pages gives a way to isolate hardware differences in later matched tests, provided driver, prompts, context and active requests are recorded.
This candidate’s 32K context and full-GPU placement are requested settings, not observed fit. Check current free memory and sustained behavior, not only a successful file download or model-load message. Windows and Linux require their own evidence, and no card-specific result is published here.
System setup
PCIe link. AMD lists PCIe 5.0 x16 (AMD specifications). In a PCIe 4.0 slot it runs at PCIe 4.0, still with all 16 lanes. On a riser rated for PCIe 4.0, set the slot to Gen 4 (our recommendation). If the link reads slow under load, see the guide's troubleshooting order.
Loading (calculated). This card's FastLLM model, Qwen3.5-9B Q8_0 (9.53 GB), loads in at least 1.4 s from a PCIe 4.0 NVMe drive and 17.3 s from a SATA SSD; the drive sets that, not the PCIe generation. A reload from system memory takes 0.15 s at PCIe 5.0 x16 (0.30 s at 4.0).
Engine. FastLLM, in development, uses llama.cpp's Vulkan engine on this card. Bitworks has no lab observation for this card yet. AMD lists this card as RDNA4, gfx1201, in its ROCm GPU specifications. FastLLM's HIP engine currently covers only the RX 7900 XT and XTX.
Resizable BAR. Turn it on; it matters most for the Vulkan engine this card uses. If FastLLM detects it off, it applies its workaround and shows how to turn it on. The workaround is a fallback, not a replacement.
Driver. Use AMD Software: Adrenalin Edition, the current WHQL Recommended release. Our test PCs use the Minimal install option.
Power and cooling. AMD lists a typical board power of 304 W and a minimum power supply recommendation of 750 W (AMD specifications). Plan airflow for sustained load, not short bursts.
Reported by the community and linked to each original. Results can differ on your system.
Related reports from the same author or project — not independent replication
IQ4_S256 · 1 card · Linux · rainference HIP · 759 tok/s prompt (aggregate) · by rainroar on Blog
Reported on Blog by rainroar — posted 2026-09-12 · measured 2026-09-11 Original report
RX 9070 XT rainference C4 Qwen3.8-27B — DFlash2
Method (as reported): One recorded C4 SSE burst after a different-prompt warmup; cold prefill/no KV reuse; server restarted between warmup and recorded burst.
Reported GPU setup RX 9070 XT 16GB (gfx1201)
GPU count 1
Operating system Linux (Arch Linux; ROCm 7.2.4)
Engine rainference working tree; manifests 054/055 applied
Backend HIP
Model Qwen3.8-27B IQ4_S256
Configured context 16,384 tokens
Actual input 3,400 tokens
Output 512 tokens
Full window tested no
Reported context Shared 16K paged pool; four slots labelled 4K; 3,400+512 tokens per request.
Simultaneous clients 4
Server slots 4
Workers 1
Topology single (One RX 9070 XT; engine compiled for a maximum of four sequences.)
Offload other (Input embedding on host (--emb-host); other placement unreported.)
CPU Ryzen 9 9900X
Memory configuration 60 GB RAM as reported; byte convention unknown
Host notes Desktop compositor shared the GPU during the reported run.
98.3 tok/s — other (aggregate across clients) · API server · 4 clients · input 3,400 tokens Steady window: 1,947 output tokens over 19.8 s, from last stream's first token to burst end; no prefill outstanding.
81.6 tok/s — other (aggregate across clients) · API server · 4 clients · input 3,400 tokens Decode-phase aggregate begins at each stream's first token and includes overlap with unfinished prompt processing.
54.3 tok/s — other (aggregate across clients) · API server · 4 clients · input 3,400 tokens Full-burst aggregate uses all 2,048 generated tokens and includes four cold prompt-processing phases.
24.6 tok/s — other (per request) · API server · 4 clients · input 3,400 tokens Author-reported per-stream rate for the C4 burst; not a latency percentile.
759 tok/s — prompt speed (aggregate across clients) · API server · 4 clients · input 3,400 tokens Aggregate cold prefill across four distinct-prefix prompts; 256-token server-side chunks and no KV reuse.
4.2–17.92 s — time to first token (per request) · API server · 4 clients · input 3,400 tokens First-to-last stream TTFT range after simultaneous release; not a percentile.
15.91 GB — memory used (server-wide) · API server · 4 clients · input 3,400 tokens Peak VRAM from the author's periodic sampler on the single card.
Not reported: model file checksum, system memory, Public rainference repository, source commit and license, GPU driver version, Main and draft artifact digests and download provenance, Repeated samples, variance and run-order effects, Raw prompt/output bundles behind author-reported digests, Output grading, cancellation and sustained stability, Remaining model and operation placement
One self-published author and host; no statistical winner follows from the three arms.
Linux custom rainference/HIP evidence only; it does not establish native-Windows or FastLLM behavior.
The custom .raif artifacts are not FastLLM catalog GGUFs; engine source, build, license and artifact digests are unverified.
Input embeddings are on the host; remaining model and operation placement and dedicated-memory residency are not proven.
The engine was compiled for at most four sequences; this is no evidence for C8 behavior.
Method (as reported): C4 SSE comparison burst under the same stated workload; cold prefill/no KV reuse; server restart required; repetition count unreported.
Reported GPU setup RX 9070 XT 16GB (gfx1201)
GPU count 1
Operating system Linux (Arch Linux; ROCm 7.2.4)
Engine rainference working tree; manifests 054/055 applied
Backend HIP
Model Qwen3.8-27B IQ4_S256
Configured context 16,384 tokens
Actual input 3,400 tokens
Output 512 tokens
Full window tested no
Reported context Shared 16K paged pool; four slots labelled 4K; 3,400+512 tokens per request.
Simultaneous clients 4
Server slots 4
Workers 1
Topology single (One RX 9070 XT; engine compiled for a maximum of four sequences.)
Speculative decoding in-file MTP head
Offload other (Input embedding on host (--emb-host); other placement unreported.)
CPU Ryzen 9 9900X
Memory configuration 60 GB RAM as reported; byte convention unknown
Host notes Desktop compositor shared the GPU during the reported run.
96 tok/s — other (aggregate across clients) · API server · 4 clients · input 3,400 tokens Steady window starts at the last stream's first token and ends at burst completion; no prefill outstanding.
83 tok/s — other (aggregate across clients) · API server · 4 clients · input 3,400 tokens Decode-phase aggregate begins at each stream's first token and includes overlap with unfinished prompt processing.
56.1 tok/s — other (aggregate across clients) · API server · 4 clients · input 3,400 tokens Full-burst aggregate uses all 2,048 generated tokens and includes four cold prompt-processing phases.
24 tok/s — other (per request) · API server · 4 clients · input 3,400 tokens Author-reported per-stream rate for the C4 burst; not a latency percentile.
835 tok/s — prompt speed (aggregate across clients) · API server · 4 clients · input 3,400 tokens Aggregate cold prefill across four distinct-prefix prompts; 256-token server-side chunks and no KV reuse.
3.8–16.29 s — time to first token (per request) · API server · 4 clients · input 3,400 tokens First-to-last stream TTFT range after simultaneous release; not a percentile.
15.02 GB — memory used (server-wide) · API server · 4 clients · input 3,400 tokens Peak VRAM from the author's periodic sampler on the single card.
Not reported: system memory, Public rainference repository, source commit and license, GPU driver version, Main artifact digest and download provenance, Repeated samples, variance and run-order effects, Raw prompt/output bundles behind author-reported digests, Output grading, cancellation and sustained stability, Remaining model and operation placement
One self-published author and host; no statistical winner follows from the three arms.
Linux custom rainference/HIP evidence only; it does not establish native-Windows or FastLLM behavior.
The custom .raif artifact is not a FastLLM catalog GGUF; engine source, build, license and artifact digest are unverified.
Input embeddings are on the host; remaining model and operation placement and dedicated-memory residency are not proven.
The engine was compiled for at most four sequences; this is no evidence for C8 behavior.
IQ4_S256 · 1 card · Linux · rainference HIP · 871 tok/s prompt (aggregate) · by rainroar on Blog
Reported on Blog by rainroar — posted 2026-09-12 · measured 2026-09-11 Original report
RX 9070 XT rainference C4 Qwen3.8-27B — no speculation
Method (as reported): C4 SSE comparison burst under the same stated workload; cold prefill/no KV reuse; server restart required; repetition count unreported.
Reported GPU setup RX 9070 XT 16GB (gfx1201)
GPU count 1
Operating system Linux (Arch Linux; ROCm 7.2.4)
Engine rainference working tree; manifests 054/055 applied
Backend HIP
Model Qwen3.8-27B IQ4_S256
Configured context 16,384 tokens
Actual input 3,400 tokens
Output 512 tokens
Full window tested no
Reported context Shared 16K paged pool; four slots labelled 4K; 3,400+512 tokens per request.
Simultaneous clients 4
Server slots 4
Workers 1
Topology single (One RX 9070 XT; engine compiled for a maximum of four sequences.)
Speculative decoding none
Offload other (Input embedding on host (--emb-host); other placement unreported.)
CPU Ryzen 9 9900X
Memory configuration 60 GB RAM as reported; byte convention unknown
Host notes Desktop compositor shared the GPU during the reported run.
98.2 tok/s — other (aggregate across clients) · API server · 4 clients · input 3,400 tokens Steady window starts at the last stream's first token and ends at burst completion; no prefill outstanding.
83.2 tok/s — other (aggregate across clients) · API server · 4 clients · input 3,400 tokens Decode-phase aggregate begins at each stream's first token and includes overlap with unfinished prompt processing.
57.6 tok/s — other (aggregate across clients) · API server · 4 clients · input 3,400 tokens Full-burst aggregate uses all 2,048 generated tokens and includes four cold prompt-processing phases.
24.6 tok/s — other (per request) · API server · 4 clients · input 3,400 tokens Author-reported per-stream rate for the C4 burst; not a latency percentile.
871 tok/s — prompt speed (aggregate across clients) · API server · 4 clients · input 3,400 tokens Aggregate cold prefill across four distinct-prefix prompts; 256-token server-side chunks and no KV reuse.
3.72–15.61 s — time to first token (per request) · API server · 4 clients · input 3,400 tokens First-to-last stream TTFT range after simultaneous release; not a percentile.
14.26 GB — memory used (server-wide) · API server · 4 clients · input 3,400 tokens Peak VRAM from the author's periodic sampler on the single card.
Not reported: system memory, Public rainference repository, source commit and license, GPU driver version, Main artifact digest and download provenance, Repeated samples, variance and run-order effects, Raw prompt/output bundles behind author-reported digests, Output grading, cancellation and sustained stability, Remaining model and operation placement
One self-published author and host; no statistical winner follows from the three arms.
Linux custom rainference/HIP evidence only; it does not establish native-Windows or FastLLM behavior.
The custom .raif artifact is not a FastLLM catalog GGUF; engine source, build, license and artifact digest are unverified.
Input embeddings are on the host; remaining model and operation placement and dedicated-memory residency are not proven.
The engine was compiled for at most four sequences; this is no evidence for C8 behavior.
Q4_K_M and AD-IQ4_XS-IQ3_S (separate arms) · 1 card · Windows · llama.cpp ROCm/HIP · 24 tok/s (other measure, see details) · 42 tok/s (other measure, see details) · by Ataa Aldaghstani / ataadevs on Blog
Reported on Blog by Ataa Aldaghstani / ataadevs — posted 2026-08-17 Original report
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)
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.
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.
IQ4_XS-v3 · 1 card · Linux · llama.cpp HIP ROCm · 28.41 tok/s generation · 1193 tok/s prompt · by kr4ckhe4d on GitHub
Reported on GitHub by kr4ckhe4d · measured 2026-10-02 Original report
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)
Discussion
Share setups, questions and your own results. Posts are reviewed before they appear. Replies from Bitworks staff are marked.