Estimate — sign in for exact rates.

 Trusted GPU & AI infrastructure since 2015 — Huntsville, AL
[email protected] Track order Support Log in
24 GB GDDR6 · RDNA 3 · gfx1100

AMD Radeon RX 7900 XTX 24GB

What this card can do

One card
27B models at 4-bit, 32K context
Two or more cards
27B at 8-bit, split across 2 cards
Windows
Windows preview, lab-measured
Linux
Linux lab alpha, unqualified

How we test and measure →

The RX 7900 XTX has 24 GB of GDDR6 memory. Its single-card candidate uses Qwen3.8-27B UD-Q4_K_M at a 32K catalog context; the 20 GB XT candidate uses another quantization and context. A raw speed comparison between those recipes would confound the cards with the model settings. A later comparison should hold artifact and workload fixed.

Two XTXs could split the Q8 candidate for capacity or run separate workers for simultaneous requests, but those are distinct configurations. Confirm each card’s free VRAM and actual PCIe connection, then measure completed work and response quality. No candidate on this page is a compatibility or performance promise.

AMD specifications

Download FastLLM Preview 2026.10.21.12

FastLLM Preview 2026.10.21.12 is a signed Windows release for AMD Radeon graphics (a full release on GitHub, not a pre-release). It currently sets up the AMD Radeon RX 7900 XTX 24GB. It is not a qualified release, and its recipes remain candidates.

Get FastLLM Preview 2026.10.21.12 on GitHub

On the release page, download FastLLM-Setup-2026.10.21.12.exe. The publisher shows as Bitworks LLC.
SHA-256 f96916b08caf5b06ecef51a688eac65621a925a2f566625aa86328a90182eb95

Test results for this release →

  • Microsoft Edge may say the file isn’t commonly downloaded while the publisher builds reputation.
  • Once installed, FastLLM shows an update banner when a newer version is available.

Models on this card

Qwen3.8-27B Model page

Bitworks

Measured by Bitworks on our lab bench with the published methodology. Not fully qualified: residency, peak VRAM, soak and reference-quality gates are not all proven for this run.

Community

Reported by the community and linked to each original. Results can differ on your system.

UD-IQ4_XS · 2 cards · Linux · llama.cpp server-rocm ROCm · 69.25 tok/s generation · 361 tok/s prompt · by ghosthand on Reddit

Reported on Reddit by ghosthand Original report

Qwen3.8-27B on dual RX 7900 XTX with tensor split and MTP

Method (as reported): Author compared single-GPU non-MTP and dual-tensor MTP server rows with f16 KV and flash attention.

  • Reported GPU setup RX 7900 XTX
  • GPU count 2
  • Operating system Linux (Docker ROCm)
  • Environment Container
  • Engine llama.cpp server-rocm b10481 (25ae3a9b3)
  • Backend ROCm
  • Model Qwen3.8-27B UD-IQ4_XS
  • Configured context 131,072 tokens
  • Actual input 86,675 tokens
  • Full window tested no
  • Reported context Author also configured 262144; not the occupied depth in these rows.
  • Simultaneous clients 1
  • Workers 1
  • Topology dual-split (tensor split; PCIe 4.0 x16 CPU plus x4 chipset)
  • Speculative decoding MTP
  • CPU Ryzen 9 9950X
  • Memory configuration 96 GB reported; unit not verified
  • PCIe PCIe 4.0 x16 CPU plus x4 chipset
  • 69.25 tok/s — generation speed (per request) · API server · 1 client · input 86,675 tokens
    Compared with 29.28 tok/s (baseline single GPU/no MTP; reported dual tensor/MTP); changed between the two runs: GPU count, split mode, speculative decoding, context
    Same reported prompt; f16 KV, flash attention; baseline also changed context.
    This row: configured context: 131,072 tokens · speculative decoding: MTP · topology: dual-split · cards: 2
  • 361 tok/s — prompt speed (per request) · API server · 1 client · input 86,675 tokens
    Compared with 516 tok/s (single GPU without MTP vs dual tensor with MTP); changed between the two runs: GPU count, split mode, speculative decoding, context
    Same reported prompt; f16 KV, flash attention; baseline also changed context.
    This row: configured context: 131,072 tokens · speculative decoding: MTP · topology: dual-split · cards: 2

Not reported: system memory, artifact digest, sample count and variance, physical memory residency

  • UD-IQ4_XS differs from the linked Q8_0 recipe.
  • The comparison changes GPU count, split mode, MTP, and configured context; it does not isolate a speedup cause.
  • The larger configured window was not filled by the prompt.

Related reports from the same author or project — not independent replication

Q8_0 · 2 cards · Linux · custom llama.cpp HIP fork ROCm · ≈60–65 tok/s generation · ≈1400 tok/s prompt · by nasone32 on Reddit

Reported on Reddit by nasone32 Original report

Dual RX 7900 XTX tensor parallel on limited PCIe

Method (as reported): Author used rocprofv3 on a saturated chipset link and combined transfer compression with a larger patch set.

HEAD seen at retrieval: 15995a12d1d5 (not necessarily the commit the author measured)

  • Reported GPU setup RX 7900 XTX
  • GPU count 2
  • Operating system Linux (Ubuntu 24; ROCm 7.14)
  • Engine custom llama.cpp HIP fork
  • Backend ROCm
  • Model Qwen3.8-27B Q8_0
  • Actual input input length not reported
  • Reported context Prompt row has known input length; decode range may use another condition.
  • Workers 1
  • Topology dual-split (tensor split; one chipset-connected PCIe x4 card)
  • Speculative decoding adaptive MTP in decode range
  • PCIe one chipset-connected PCIe x4 card
  • ≈1400 tok/s — prompt speed (per request) · input 8,192 tokens
    Compared with 1100 tok/s (baseline patched AllReduce; reported Q8 transfer); changed between the two runs: engine build, other
    Approximate author comparison on the same asymmetric host.
    This row: topology: dual-split
  • ≈60–65 tok/s — generation speed (per request)
    Combined RDNA Boost and adaptive MTP patch set; not isolated transfer-only data.
    This row: speculative decoding: adaptive MTP · topology: dual-split

Not reported: configured context, simultaneous clients, CPU, system memory, exact tested fork commit, artifact digest, replications, quality delta, full launch arguments

  • Custom patch set differs from the linked recipe and is not a stock Windows path.
  • Same author and host as the topology report; not an independent replication.
  • Author had not evaluated quality impact of Q8 inter-card transfer.
  • Fork author says combined build will not be maintained.

Setup guide (external)

Related link (external)

Q8_0 · 2 cards · Linux · custom llama.cpp HIP ROCm · ≈50 tok/s generation · ≈1150 tok/s prompt · by nasone32 on Reddit

Reported on Reddit by nasone32 Original report

Notes on dual RX 7900 XTX tensor split topology

Method (as reported): Author compared layer and patched tensor split on an asymmetric dual-card host; exact test counts not supplied.

  • Reported GPU setup RX 7900 XTX
  • GPU count 2
  • Operating system Linux (ROCm Linux; Windows failure also described)
  • Engine custom llama.cpp HIP
  • Backend ROCm
  • Model Qwen3.8-27B Q8_0
  • Actual input input length not reported
  • Workers 1
  • Topology dual-split (PCIe 4.0 x16 CPU plus x4 chipset)
  • Speculative decoding MTP3 in generation comparison
  • PCIe PCIe 4.0 x16 CPU plus x4 chipset
  • ≈1150 tok/s — prompt speed (per request)
    Compared with 1500 tok/s (baseline layer split; reported patched tensor split); changed between the two runs: split mode, engine build
    Approximate author-reported comparison on asymmetric PCIe topology.
    This row: topology: dual-split
  • ≈50 tok/s — generation speed (per request)
    Compared with 30 tok/s (baseline layer split; reported patched tensor/MTP3); changed between the two runs: split mode, engine build, speculative decoding
    Approximate prose comparison; split and speculation both changed.
    This row: speculative decoding: MTP3 · topology: dual-split

Not reported: configured context, actual input length, simultaneous clients, CPU, system memory, exact build and patch digest, prompt and output counts, repetition and variance, Windows version and driver

  • Custom internal AllReduce runtime differs from the linked recipe.
  • The generation comparison changes both split and MTP settings.
  • Author-reported Windows P2P failure does not establish a native-Windows tensor recipe.
  • Predicted symmetric-slot gains are not measurements.

Related reports from the same author or project — not independent replication

UD-Q4_K_XL · 1 card · Windows · llama.cpp Vulkan · by pparuzel on GitHub

Reported on GitHub by pparuzel — posted 2026-08-18 Original report

RX 7900 XTX Windows/Vulkan Qwen3.8-27B — MTP N=2 acceptance rerun

Method (as reported): Separate N=2 rerun retained for draft-acceptance logs; warmups discarded across the same three prompt classes.

HEAD seen at retrieval: 1e514a89a102 (not necessarily the commit the author measured)

  • Reported GPU setup RX 7900 XTX 24GB
  • GPU count 1
  • Operating system Windows (Windows 11 native)
  • Environment Bare metal
  • Engine llama.cpp b9553 commit 9e3b928fd
  • Backend Vulkan
  • Model Qwen3.8-27B UD-Q4_K_XL (Unsloth)
  • Actual input input length not reported
  • Full window tested no
  • Reported context Author configured "131K"; exact window and occupied input are unknown. Request ceiling was 400.
  • Simultaneous clients 1
  • Server slots 1
  • Topology single (q4_0 K/V cache, flash attention on, thinking off; Vulkan0 selected for serving.)
  • Speculative decoding draft-mtp; n-max 2; p-min 0; acceptance rerun
  • Offload not reported — Vulkan0 selected; complete model and operation placement are not proved.
  • CPU AMD Ryzen 7 9800X3D
  • Host notes An AMD iGPU remained active under the same unreported driver.

Not reported: configured context, system memory, Exact GGUF filename and digest, AMD driver version and measurement date, Input count, exact configured window, occupied context and worker count, Authoritative generated-token counts, TTFT and full-request wall time, Output identity, task quality, independent sessions and sustained reliability

  • The pinned probe counts nonempty content or reasoning SSE deltas, not tokenizer or API usage tokens; no rate is normalized.
  • Draft-acceptance logs do not repair the output-rate numerator and are intentionally not normalized as a website metric.
  • This rerun differs from the original N=2 observation and is retained separately, not averaged or treated as replication.
  • Build b9553 differs from FastLLM b10698; one C1 author report cannot establish a general MTP advantage.

Setup guide (external)

Related link (external)

Related link (external)

Related link (external)

UD-Q4_K_XL · 1 card · Windows · llama.cpp Vulkan · by pparuzel on GitHub

Reported on GitHub by pparuzel — posted 2026-08-18 Original report

RX 7900 XTX Windows/Vulkan Qwen3.8-27B — original MTP N=2 arm

Method (as reported): Original N=2 arm: warmups discarded, then three recorded runs in each of three prompt classes using the pinned probe.

HEAD seen at retrieval: 1e514a89a102 (not necessarily the commit the author measured)

  • Reported GPU setup RX 7900 XTX 24GB
  • GPU count 1
  • Operating system Windows (Windows 11 native)
  • Environment Bare metal
  • Engine llama.cpp b9553 commit 9e3b928fd
  • Backend Vulkan
  • Model Qwen3.8-27B UD-Q4_K_XL (Unsloth)
  • Actual input input length not reported
  • Full window tested no
  • Reported context Author configured "131K"; exact window and occupied input are unknown. Request ceiling was 400.
  • Simultaneous clients 1
  • Server slots 1
  • Topology single (q4_0 K/V cache, flash attention on, thinking off; Vulkan0 selected for serving.)
  • Speculative decoding draft-mtp; n-max 2; p-min 0
  • Offload not reported — Vulkan0 selected; complete model and operation placement are not proved.
  • CPU AMD Ryzen 7 9800X3D
  • Host notes An AMD iGPU remained active under the same unreported driver.

Not reported: configured context, system memory, Exact GGUF filename and digest, AMD driver version and measurement date, Input count, exact configured window, occupied context and worker count, Authoritative output count, original-arm draft acceptance and request wall time, Output identity, task quality, independent sessions and sustained reliability

  • The pinned probe counts nonempty content or reasoning SSE deltas, not tokenizer or API usage tokens; no rate is normalized.
  • The original N=2 row and later acceptance rerun are distinct observations and are not merged as repetitions.
  • Build b9553 is different from FastLLM b10698; this record does not transfer compatibility or correctness between them.
  • One author, host, C1 client and server slot do not establish concurrency or a general MTP advantage.

Setup guide (external)

Related link (external)

Related link (external)

Related link (external)

UD-Q4_K_XL · 1 card · Windows · llama.cpp Vulkan · by pparuzel on GitHub

Reported on GitHub by pparuzel — posted 2026-08-18 Original report

RX 7900 XTX Windows/Vulkan Qwen3.8-27B — MTP N=3 arm

Method (as reported): N=3 arm with draft-acceptance logs; warmups discarded, then three recorded runs in each of three prompt classes.

HEAD seen at retrieval: 1e514a89a102 (not necessarily the commit the author measured)

  • Reported GPU setup RX 7900 XTX 24GB
  • GPU count 1
  • Operating system Windows (Windows 11 native)
  • Environment Bare metal
  • Engine llama.cpp b9553 commit 9e3b928fd
  • Backend Vulkan
  • Model Qwen3.8-27B UD-Q4_K_XL (Unsloth)
  • Actual input input length not reported
  • Full window tested no
  • Reported context Author configured "131K"; exact window and occupied input are unknown. Request ceiling was 400.
  • Simultaneous clients 1
  • Server slots 1
  • Topology single (q4_0 K/V cache, flash attention on, thinking off; Vulkan0 selected for serving.)
  • Speculative decoding draft-mtp; n-max 3; p-min 0
  • Offload not reported — Vulkan0 selected; complete model and operation placement are not proved.
  • CPU AMD Ryzen 7 9800X3D
  • Host notes An AMD iGPU remained active under the same unreported driver.

Not reported: configured context, system memory, Exact GGUF filename and digest, AMD driver version and measurement date, Input count, exact configured window, occupied context and worker count, Authoritative generated-token counts, TTFT and full-request wall time, Output identity, task quality, independent sessions and sustained reliability

  • The pinned probe counts nonempty content or reasoning SSE deltas, not tokenizer or API usage tokens; no rate is normalized.
  • Draft-acceptance logs do not repair the output-rate numerator and are intentionally not normalized as a website metric.
  • Build b9553 is different from FastLLM b10698; this record does not transfer compatibility or correctness between them.
  • One author, host, C1 client and server slot do not establish concurrency or a general MTP advantage.

Setup guide (external)

Related link (external)

Related link (external)

Related link (external)

UD-Q4_K_XL · 1 card · Windows · llama.cpp Vulkan · by pparuzel on GitHub

Reported on GitHub by pparuzel — posted 2026-08-18 Original report

RX 7900 XTX Windows/Vulkan Qwen3.8-27B — MTP N=4 arm

Method (as reported): N=4 arm with draft-acceptance logs; warmups discarded, then three recorded runs in each of three prompt classes.

HEAD seen at retrieval: 1e514a89a102 (not necessarily the commit the author measured)

  • Reported GPU setup RX 7900 XTX 24GB
  • GPU count 1
  • Operating system Windows (Windows 11 native)
  • Environment Bare metal
  • Engine llama.cpp b9553 commit 9e3b928fd
  • Backend Vulkan
  • Model Qwen3.8-27B UD-Q4_K_XL (Unsloth)
  • Actual input input length not reported
  • Full window tested no
  • Reported context Author configured "131K"; exact window and occupied input are unknown. Request ceiling was 400.
  • Simultaneous clients 1
  • Server slots 1
  • Topology single (q4_0 K/V cache, flash attention on, thinking off; Vulkan0 selected for serving.)
  • Speculative decoding draft-mtp; n-max 4; p-min 0
  • Offload not reported — Vulkan0 selected; complete model and operation placement are not proved.
  • CPU AMD Ryzen 7 9800X3D
  • Host notes An AMD iGPU remained active under the same unreported driver.

Not reported: configured context, system memory, Exact GGUF filename and digest, AMD driver version and measurement date, Input count, exact configured window, occupied context and worker count, Authoritative generated-token counts, TTFT and full-request wall time, Output identity, task quality, independent sessions and sustained reliability

  • The pinned probe counts nonempty content or reasoning SSE deltas, not tokenizer or API usage tokens; no rate is normalized.
  • Draft-acceptance logs do not repair the output-rate numerator and are intentionally not normalized as a website metric.
  • Build b9553 is different from FastLLM b10698; this record does not transfer compatibility or correctness between them.
  • One author, host, C1 client and server slot do not establish concurrency or a general MTP advantage.

Setup guide (external)

Related link (external)

Related link (external)

Related link (external)

UD-Q4_K_XL · 1 card · Windows · llama.cpp Vulkan · by pparuzel on GitHub

Reported on GitHub by pparuzel — posted 2026-08-18 Original report

RX 7900 XTX Windows/Vulkan Qwen3.8-27B — speculation-off arm

Method (as reported): Same-probe baseline backfilled on the same serve with speculation flags removed; warmups discarded across three prompt classes.

HEAD seen at retrieval: 1e514a89a102 (not necessarily the commit the author measured)

  • Reported GPU setup RX 7900 XTX 24GB
  • GPU count 1
  • Operating system Windows (Windows 11 native)
  • Environment Bare metal
  • Engine llama.cpp b9553 commit 9e3b928fd
  • Backend Vulkan
  • Model Qwen3.8-27B UD-Q4_K_XL (Unsloth)
  • Actual input input length not reported
  • Full window tested no
  • Reported context Author configured "131K"; exact window and occupied input are unknown. Request ceiling was 400.
  • Simultaneous clients 1
  • Server slots 1
  • Topology single (q4_0 K/V cache, flash attention on, thinking off; Vulkan0 selected for serving.)
  • Speculative decoding none
  • Offload not reported — Vulkan0 selected; complete model and operation placement are not proved.
  • CPU AMD Ryzen 7 9800X3D
  • Host notes An AMD iGPU remained active under the same unreported driver.

Not reported: configured context, system memory, Exact GGUF filename and digest, AMD driver version and measurement date, Input count, exact configured window, occupied context and worker count, Authoritative generated-token counts, TTFT and full-request wall time, Output identity, task quality, independent sessions and sustained reliability

  • The pinned probe counts nonempty content or reasoning SSE deltas, not tokenizer or API usage tokens; no rate is normalized.
  • The same-serve backfill is useful qualitative control evidence, not an independent session or statistical baseline.
  • Build b9553 is different from FastLLM b10698; this record does not transfer compatibility or correctness between them.
  • Windows GPU counters and Vulkan0 selection do not prove complete operation placement or physical residency.

Setup guide (external)

Related link (external)

Related link (external)

Related link (external)

DeepSeek-V4-Flash-0731

Bitworks

Bitworks: no recipe for this card yet.

Community

UD-Q8_K_XL main GGUF plus Q2_K-Q8_0 DSpark drafter · 2 cards · Linux · llama-server ROCm ROCm · 10.5 tok/s generation · 52 tok/s prompt · by Inevitable-Big-1131 on Reddit

Reported on Reddit by Inevitable-Big-1131 Original report

DeepSeek V4 Flash host offload on dual RX 7900 XTX

Method (as reported): Author supplies launch command with two GPUs, layer split, CPU expert tensors, Q8 KV, DSpark draft and swap-disabled cgroup.

  • Reported GPU setup RX 7900 XTX
  • GPU count 2
  • Operating system Linux (systemd cgroup)
  • Engine llama-server ROCm
  • Backend ROCm
  • Model DeepSeek-V4-Flash-0731 UD-Q8_K_XL main GGUF plus Q2_K-Q8_0 DSpark drafter
  • Configured context 131,072 tokens
  • Actual input input length not reported
  • Reported context Configured context only; actual prompt occupancy unreported.
  • Simultaneous clients 1
  • Server slots 1
  • Workers 1
  • Topology dual-split (explicit devices and layer split)
  • Speculative decoding DSpark draft
  • Offload other (host system-memory offload with systemd cgroup limits)
  • Memory configuration 128 GB reported; unit not verified
  • 52 tok/s — prompt speed (per request) · API server · 1 client
    Single-client exploratory request; Q8 KV, CPU experts, draft and swap-disabled cgroup.
    This row: configured context: 131,072 tokens
  • 10.5 tok/s — generation speed (per request) · API server · 1 client
    Same author-reported setup; context occupancy unknown.
    This row: configured context: 131,072 tokens

Not reported: engine build, model file checksum, CPU, model and drafter digests, actual input depth, sample count, PCIe topology, peak committed RAM

  • This is DeepSeek-V4-Flash, not Qwen3.8-Flash-Next.
  • Single-client observation does not establish concurrent service.
  • Host RAM and CPU placement dominate transferability.

Qwen3.8-Flash-Next

Bitworks

Bitworks: no recipe for this card yet.

Community

IQ3_XXS · 1 card · Strata AMD HIP HIP · ≈45–70 tok/s generation · ≈800 tok/s prompt · by EmPips on Reddit

Reported on Reddit by EmPips Original report

Strata Flash-Next offload on single RX 7900 XTX

Method (as reported): User anecdote on DDR4 host with offloaded experts; exact OS/build and repetition not supplied.

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

  • Reported GPU setup RX 7900 XTX
  • GPU count 1
  • Operating system OS not reported — GPU-only evidence
  • Engine Strata AMD HIP
  • Backend HIP
  • Model Qwen3.8-Flash-Next IQ3_XXS
  • Actual input input length not reported
  • Reported context Author says usual prompt size around 60K; no exact occupied window.
  • Workers 1
  • Topology single
  • Speculative decoding MTP-dependent
  • Offload experts-to-host (host-memory offload; exact partition unknown)
  • Host notes Slow DDR4 host reported; exact capacity/speed unknown
  • ≈45–70 tok/s — generation speed (per request)
    Compared with 22.5 tok/s (baseline tuned llama.cpp; reported Strata on same host); changed between the two runs: engine, speculative decoding
    Anecdotal coding use; Strata range depends on MTP.
    This row: engine: Strata AMD HIP · speculative decoding: MTP-dependent
  • ≈800 tok/s — prompt speed (per request)
    Approximate report around usual long prompts; exact count unknown.
    This row: engine: Strata AMD HIP

Not reported: operating system, configured context, simultaneous clients, CPU, OS and Strata build, host RAM capacity and bandwidth, pack digest, expert cache hit rate, MTP acceptance, quality grader

  • OS, exact runtime, memory partition and test repetitions are unknown.
  • Another user reports a slower setup; no numeric cross-user comparison is included.
  • A separate commenter preferred 27B Q8 answers; cause was not isolated.
  • Flash-Next has separate Qwen Community 1.0 terms, not the catalog Apache-2.0 terms.

Setup guide (external)

Related link (external)

References

Discussion

Share setups, questions and your own results. Posts are reviewed before they appear. Replies from Bitworks staff are marked.

Log in Create an account to join the discussion.

Bitworks listings for this card