R1–R9

Evidence library

Every key figure on this site is followed by an ID that points to a report listed here. This page publishes what each report tested, on which platform, and what it concluded; the original documents are available on request.

Updated November 12, 2026

9 test reports

R1FX100 Comprehensive LLM Inference and Training Benchmark (8 × AMD MI308X)One report, five questions: how much the LMCache parallel-read patch helps, whether KV tiering scales across GPUs, how much faster 8 GPUs load models concurrently, how effective utilization changes under frequent model switching, and how much faster concurrent training checkpoint writes become. Test models are Qwen2.5-32B and 7B.2026-07-03
8 × AMD Instinct MI308X · ROCm 7.2
R2FX100 KV Cache Benchmark (480B, TP8, Long Context)Qwen3-Coder-480B-FP8 deployed as one 8-GPU instance (TP=8), the standard production topology, under a long-context cold-recovery load. A three-way comparison of FX100, a single local NVMe drive and full recompute without external storage, stepped across concurrency 8 / 16 / 32. Most throughput and TTFT figures quoted on this site come from this report.2026-07-05
8 × AMD Instinct MI308X · ROCm 7.2
R3FX100 KV Cache Benchmark Summary (480B, TP4×2, All Metrics)The same 8-GPU server redeployed as two 4-GPU instances (TP4×2), with every metric recorded across nine load levels: TTFT p50 / p90 / p99, TPOT, throughput and disk bandwidth. The report also verifies two things: the fs:// shared pool lets both instances hot-share one cache, and two independent runs at the same level deviate by no more than 5%.2026-07-06
8 × AMD Instinct MI308X · ROCm 7.2
R4FX100 KV Cache Benchmark (480B, Multi-Instance, No. -006)The official compilation of R2 and R3, issued by a third-party testing organization as report No. -006. If your procurement or compliance process needs a document from an independent body, this is the one to cite.2026-07-06
8 × AMD Instinct MI308X · ROCm 7.2
R5FX100 KV Cache Benchmark (14B, HBM Efficiency)This report answers whether an external KV tier can stand in for GPU memory. A seven-way comparison on a 14B model (GDS direct read, recompute at two HBM sizes, direct read with low HBM) builds a three-step argument for HBM equivalence. The platform is a single MetaX N260 GPU; the method carries across platforms, but the numbers are never mixed with MI308X results.2026-07-03
Single MetaX N260 GPU
R6ComfyUI + LTX-Video 2.3 Full-Model Deployment, Adaptation and Usage Report (V2)Proof of adaptation for video generation: all 7 delivered models (3 VAEs, the fp8 text encoder, the text projection and 2 LoRAs) run end to end on MI308X / ROCm 7.2 with no dtype, operator or kernel errors.2026-07-07
AMD Instinct MI308X · ROCm 7.2
R7ComfyUI + LTX-Video 2.3 Model Adaptation Report (AMD MI308X)The per-model detail behind R6: individual run records for the 3 VAEs and the Singularity LoRA, plus the adaptation points that matter on MI308X.2026-07-07
AMD Instinct MI308X
R8FX100 KV Cache Test Code Export and Raw Benchmark DataEverything a third party needs to rerun the tests: the LMCache parallel-read patch (git patch plus full before/after sources), the load client, orchestration scripts, forensic probe scripts and the raw benchmark data.2026-07
AMD Instinct MI308X · ROCm 7.2
R9FX100 on Huawei Ascend 910B: Inference and Training Benchmark (vs. NFS Baseline)Three comparison groups against an NFS baseline on an Ascend platform with 8 × Huawei Atlas 910B: model loading for inference (DeepSeek-32B / 70B), training weight and checkpoint I/O (Qwen-7B), and training-data acceleration (YOLOv8 / COCO). The platform is Ascend 910B, not MI308X, and is labeled as such wherever this site quotes it.2026-05-30
8 × Huawei Atlas 910B (Kunpeng-920)

Key figures and their sources

+29–40%

Higher inference throughput

vs. a single local NVMe drive

Qwen3-Coder-480B-FP8 in TP=8 production topology under a long-context cold-recovery load: +29% at concurrency 8 and +40% at concurrency 16, the operating point with the largest gain; +35–36% for the whole machine in the dual-instance TP4×2 layout.

MeasuredSourceR2 R3

−26–32%

Lower time to first token (TTFT)

vs. a single local NVMe drive

Across three concurrency levels of the same 480B TP8 load, TTFT p50 falls from 10.17–35.73 s to 7.53–26.35 s.

MeasuredSourceR2

8.6–20×

Faster than recomputing without external storage

vs. full recompute with no external storage

At concurrency 16, TTFT p50 is 149.5 s when recomputing and 11.85 s with FX100; throughput is 4.1 tok/s versus 74.9 tok/s.

MeasuredSourceR2

4.1×

TTFT gain from the LMCache parallel-read patch

vs. before the patch

One GPU, concurrency 16, cold read from disk (Qwen2.5-32B): TTFT 37.97 s → 9.30 s; read bandwidth 0.98 → 5.23 GB/s, about 5.3×.

MeasuredSourceR1

6.2–9.3×

Faster model loading

vs. NFS, on Huawei Ascend 910B

On Huawei Atlas 910B: DeepSeek-32B serving load 691 s → 112 s (6.2×); DeepSeek-70B 1399 s → 150 s (9.3×).

MeasuredSourceR9

1.9×

Faster training checkpoint saves

vs. the baseline in report R1

8-GPU 32B LoRA with 65.6 GB full-model snapshots: 178 s → 94 s; sustained write bandwidth 3.26 → 6.40 GB/s (+96%).

MeasuredSourceR1

46.7% → 62.8%

Effective compute utilization under model switching

at 20 model switches per hour

Modeled from measured load times: at 20 model switches per hour, effective compute utilization rises from 46.7% to 62.8%.

MeasuredSourceR1

16% / 30%

Concurrent model loading on 8 GPUs

vs. the baseline in report R1

Eight GPUs cold-reading models at once: 16% faster wall-clock overall and 30% faster per model.

MeasuredSourceR1

≤5%

Deviation between two independent runs

same load level, measured twice

Two independent runs at the same load level deviate by no more than 5%; cross-instance hot sharing through the fs:// shared pool was verified in the same campaign.

MeasuredSourceR3

128 GB

HBM residency an external KV tier can stand in for

MetaX N260 single GPU, 14B model, seven-way comparison

Conclusion of a seven-way comparison on a 14B model: an external KV tier can stand in for 128 GB-class HBM residency. That campaign ran on a single MetaX N260 GPU; the method carries across platforms, the numbers are not mixed with MI308X results.

MeasuredSourceR5

7 / 7

ComfyUI + LTX-Video 2.3 models running end to end

AMD MI308X, ROCm 7.2

All 7 delivered models (3 VAEs, the fp8 text encoder, the text projection and 2 LoRAs) run end to end on MI308X / ROCm 7.2 with no dtype, operator or kernel errors.

MeasuredSourceR6 R7

1 : 8

Arrays to inference nodes

production sizing basis

One FX100 serves 8 inference nodes; 100 GbE sustaining KV read-back for 8 nodes is a measured anchor. On the joint-test basis, storage spend stays within roughly 10% of the GPU budget.

EstimatedSourceR1 R2 R3

480B · TP8 raw readings

FX100 all-flash arraySingle local NVMe drive

TTFT p50 (s)

↓ lower is better

Concurrency8

Local NVMe drive · Concurrency 8: 10.17 s
FX100 · Concurrency 8: 7.53 s

−26%

Concurrency16

Local NVMe drive · Concurrency 16: 17.31 s
FX100 · Concurrency 16: 11.85 s

−32%

Concurrency32

Local NVMe drive · Concurrency 32: 35.73 s
FX100 · Concurrency 32: 26.35 s

−26%

Throughput (tok/s)

↑ higher is better

Concurrency8

Local NVMe drive · Concurrency 8: 43.9 tok/s
FX100 · Concurrency 8: 56.6 tok/s

+29%

Concurrency16

Local NVMe drive · Concurrency 16: 53.6 tok/s
FX100 · Concurrency 16: 74.9 tok/s

+40%

Concurrency32

Local NVMe drive · Concurrency 32: 53.9 tok/s
FX100 · Concurrency 32: 71.6 tok/s

+33%

Fig. 2 · Qwen3-Coder-480B-FP8 · TP=8 · 8 × MI308X · long-context cold-recovery load. TTFT is p50 in seconds; throughput in tok/s; full recompute was measured at concurrency 16 only.R2

Table 1 · Qwen3-Coder-480B-FP8 · TP=8 · 8 × MI308X · long-context cold-recovery load. TTFT is p50 in seconds; throughput in tok/s; full recompute was measured at concurrency 16 only. Source: R2
Concurrency	FX100 TTFT	Local NVMe drive TTFT	Full recompute TTFT	FX100 tok/s	Local NVMe drive tok/s	Full recompute tok/s
8	7.53 s	10.17 s	—	56.6	43.9	—
16	11.85 s	17.31 s	149.48 s	74.9	53.6	4.1
32	26.35 s	35.73 s	—	71.6	53.9	—

Test platform

Applies to R1, R2, R3, R4. Other reports state their platform in their own entries.
Accelerators	8 × AMD Instinct MI308X (192 GB HBM each, gfx942)
Host	2 × AMD EPYC 9654 (384 threads), about 1.5 TB RAM
Software stack	ROCm 7.2 · vLLM 0.20.1+rocm721 · LMCache (built from upstream mainline, 2026-06-29)
Device under test	FX100 all-flash NVMe-oF array: 4-drive RAID0 (14 TB, XFS), RoCEv2, one 100 GbE port
Baselines	A single local NVMe drive (PCIe Gen4, 2 TB); no external storage, full recompute
480B model	Qwen3-Coder-480B-FP8 (MoE, about 450 GB of weights)

Four kinds of data

Measured: From a signed or official test report, with a report ID.
Vendor: Nominal product specifications and list reference prices.
Public: From publicly available third-party material, with a link to the source.
Estimated: Derived from measured anchors, with a costing model reproducible in Python.

Request the original reports

Tell us which report IDs you need and what for, and we will reply with how to get them. The R8 code export and raw data are provided on request in the same way.

Request the original report

A note on names: in earlier test-report filenames FX100 appears as AISSD5000, and before that as WS5000 / GP5000; some reports are signed “Mingxin”. They all refer to the same product. This site uses the FX naming throughout (FX100 / FX200 / FX300 / FX400), while report entries keep their original titles so you can cross-check.

Test jointly first, decide afterwards

Every engagement starts with a gated joint test of about 10 weeks. Four gates are passed in sequence; if one is missed, the project stops there, so you never pay for a benefit that was not demonstrated.

Book a joint test See the evidence

Contact: Wang Limin
Phone: +86 131 6181 8898
Email: mingxin@agentmail.to
The inbox is monitored and triaged by AI around the clock