When -sm row stopped working in my dual Tesla P40 setup, there was no dependable flag to swap in and preserve the old result. The useful fix was to test the exact build, model, and workload again—and then measure other ways to improve throughput. One important qualification: current upstream llama.cpp documentation retrieved around October 7, 2026 still lists row split, so this is an account of a configuration-specific failure, not proof that the mode has been universally removed.
What changed in my dual-P40 setup
In my earlier setup, row splitting had delivered roughly 12–14 tokens per second, compared with about 7 tokens per second for layer splitting. In an earlier 72B-model configuration, full GPU residency with row split reached approximately 10.3 generated tokens per second and 60 prompt tokens per second. These are my measurements in my environment, not independent benchmarks or promises of what another system will achieve. Source: my original account.
As an Amazon Associate I earn from qualifying purchases.
The title reflects how the change felt in that workflow: I had tuned around a flag and then encountered a model-and-backend condition where it no longer worked. But upstream documentation available around October 7, 2026 still lists row as a split mode. A July 12, 2026 issue reports a row-split failure on a particular CUDA build in a mixed CUDA/ROCm setup; that establishes a compatibility report, not universal removal. The behavior depends on the release, backend, device mix, model, and workload.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Why old split-mode results did not settle the question
In a March comparison, changing several factors at once obscured a substantial prompt-processing regression. A later comparison that changed one variable at a time reportedly found row split working on the original binary, layer split running at about half the speed, and graph split crashing on Pascal with an illegal-memory-access error. Those results describe that test sequence; they do not establish a general ranking of split modes.
#1 Best Overall
- System Compatibility Note: This 2-slot card measures 271 x 112 x 39 mm and requires a single 12V-2x6-pin power connector. Please verify chassis and PSU compatibility before purchase.
- Dedicated Support: Please contact us directly through Amazon for any product questions or assistance you may require.
- Professional Intel Arc Pro B70 GPU: Built on the Intel Xe2-HPG architecture, it features 32 Xe cores and 256 XMX engines, designed to accelerate AI, rendering, and complex visualization workloads.
- Massive 32GB GDDR6 VRAM: Equipped with 32GB of high-speed GDDR6 memory on a 256-bit bus, running at 19 Gbps, which allows for handling large AI models and complex datasets locally.
- High-Performance Engine Clock: Delivers an engine clock of 2540 MHz, providing the compute power needed for demanding professional applications and AI inference.
My account also describes a model-specific complication: Gemma 4’s shared KV layers, represented as tensor views, caused row split to fail in my multi-GPU CUDA setup, while my Qwen stacks continued to use row split. That is an observation about those configurations, not evidence that every Gemma 4 or multi-GPU setup behaves the same way.
What the split modes do—and what documentation can establish
The llama.cpp server README retrieved around October 7, 2026 lists -sm / --split-mode choices as none, layer, row, and tensor. It describes layer mode as the default, with layers and KV split across GPUs; row mode splits weights by rows; and tensor mode as experimental. The documentation describes behavior, not a universal performance ordering. llama.cpp server README.
Rank #2
- PLEASE NOTE: Exporting an NVIDIA RTX Pro 6000 GPU outside the US requires strict adherence to the U.S. Export Administration Regulations (EAR) and issuance of an export license from the Bureau of Industry and Security (BIS). Compliance and Know Your Customer (KYC) screening may be required as a condition of order acceptance. [NVIDIA Blackwell Streaming Multiprocessor] The new SM features increased processing throughput, and new neural shaders that integrate neural networks inside of programmable shaders | DLSS 4: Multi Frame Generation ensures ultra-smooth frame pacing for lifelike simulations.
- [Double-Flow-Through Design] The RTX PRO 6000 Blackwell features a double-flow-through cooling design, optimizing efficiency and airflow to sustain peak performance under 600W power loads. | [5th Gen Tensor Cores] Deliver up to 3X the performance of the previous generation and support for FP4 precision for faster AI model processing times with reduced memory usage, enabling local fine-tuning of LLMs and generative AI | [4th Gen Ray Tracing Cores] Double the ray-triangle intersection rate of the previous generation to create photoreal, physically accurate scenes and immersive 3D designs with RTX Mega Geometry, which enables up to 100X more ray-traced triangles.
- [PCIe Gen 5] Support for PCIe Gen 5 provides double the bandwidth of PCIe Gen 4, improving data-transfer speeds from CPU memory and unlocking faster performance for data-intensive tasks like AI, data science, and 3D modeling. | [GDDR7 Memory] With 96 GB of GPU memory and 1.8 TB ps bandwidth, it can tackle massive 3D and AI projects, fine-tune AI models locally, explore large-scale VR environments, and drive larger multi-app workflows.
- [DisplayPort 2.1] Achieve unparalleled visual clarity and performance, driving high resolution displays at up to 8K at 240 Hz and 16K at 60 Hz. Increased bandwidth enables seamless multi-monitor setups while HDR and higher color depth support ensures superior color accuracy for precision work, such as video editing, 3D design, and live broadcasting.
- [Universal MIG] Divide a single RTX PRO 6000 Blackwell into multiple isolated instances, each with dedicated resources, allowing for concurrent execution of multiple workloads, optimized GPU utilization, and secure isolation of different applications or users. [WARRANTY] 3 YR Manufacturer's Warranty. Bulk OEM Packaging. Retail Packaging is NOT included.
The CLI README retrieved around the same date also lists row split, parallel sequences, and speculative decoding modes including draft-mtp. These are mutable master pages, not guarantees for every release or backend. Check the documentation and behavior for the specific binary you run. llama.cpp CLI README.
The July 12, 2026 issue is useful as a reminder to preserve configuration details: its reported failure concerned a particular CUDA build and mixed CUDA/ROCm setup, and it also describes other split-mode failures in that environment. Issue report.
Rank #3
- System Compatibility Note: 2-slot card, 271x112x39mm, single 8-pin power, 200W TDP. Verify chassis clearance and PSU capacity before purchase.
- Dedicated Support: Please contact us directly through Amazon for any product questions or assistance you may require.
- 24GB GDDR6 on 192-Bit Bus: Massive 24GB memory with 456 GB/s bandwidth – ideal for LLMs, AI inference, 3D rendering, and generative design.
- Intel Xe2-HPG Architecture: Built on Intel's next-gen architecture with 20 Xe cores and 160 XMX engines for AI acceleration (197 INT8 TOPS).
- PCIe 5.0 Support: PCI Express 5.0 x16 interface for maximum bandwidth with the latest workstation platforms.
How to re-measure after a flag or model change
- Record the setup. Note the llama.cpp release or commit, build and backend, GPU models and device mix, model and quantization, split mode, and relevant launch options. A mutable documentation page is not a substitute for identifying the binary that produced a result.
- Choose a repeatable workload. Keep the model, prompt, generation length, and concurrency consistent when comparing runs. Track prompt processing and generation separately; a change can affect one without making the other obvious.
- Change one factor at a time. Test the modes that the exact build exposes, and record whether each run starts, completes correctly, or fails. Treat a crash or incorrect output as a compatibility result, not merely a low score.
- Measure both latency and throughput. A single request answers a different question from several concurrent sequences. Record per-stream speed as well as aggregate tokens per second when concurrency matters.
- Repeat the comparison when the architecture or software changes. A model with different KV behavior, a new backend, or a different binary can invalidate a previously useful tuning rule.
What helped recover throughput in my later stack
Parallel slots increased aggregate throughput
With layer split in a later stack, I measured 8.46 tokens per second for one stream. Running four parallel slots raised my reported aggregate throughput to 15.0 tokens per second; two slots measured 12.8. That is a throughput result from my workload, not evidence that four slots improve single-request latency or will help every model and GPU. The current server documentation names --parallel / -np as the number of parallel sequences to decode; confirm the option and its behavior against your build.
MTP improved single-stream speed in my test
In the same later stack, I measured single-stream speed increasing from 8.46 to about 13.3 tokens per second with MTP speculative decoding, a reported 57% gain. I recorded acceptance rates from 0.38 to 0.63 and checked output correctness. These are my results, not independently reproduced measurements. The CLI documentation retrieved around October 7, 2026 lists draft-mtp among speculative decoding modes; actual availability and setup depend on the build and model.
Rank #4
- 【High-Performance APU】The MS-S1 MAX features an AMD Ryzen AI Max+ 395 APU, integrating a Zen 5 architecture CPU (up to 5.1GHz, 16C/32T, 64M L3 Cache), an RDNA 3.5 GPU, and an NPU (50 TOPS). The total system output is 126 TOPS. It provides powerful parallel computing capabilities for demanding AI workflows. It is ideal for running local LLMs, multimodal models, and computationally intensive tasks
- 【128GB UMA Memory】Equipped with up to 128GB of LPDDR5x-8000MT/s unified memory, it enables the CPU and GPU to access a shared, high-bandwidth memory pool with extremely low latency. Ideal for large-scale AI inference, 3D workloads, and complex timelines in video editing. It eliminates traditional VRAM bottlenecks, ensuring smoother data transfer during high-intensity computations. The UMA design maximizes performance stability under high loads
- 【Flexible Expansion】The MS-S1 MAX features USB4 V2 (up to 80Gbps), dual 10GbE LAN, HDMI 2.1 (up to 8K60), a full-length PCIe x16 expansion slot, and dual M.2 slots supporting up to 16TB RAID 0/1. Wi-Fi 7 provides stronger signal coverage and a more stable wireless experience. The slide-out design facilitates upgrades and maintenance. It easily adapts to personal, studio, or rack-mount enterprise environments
- 【High-Efficiency Cooling System】Utilizing an aerospace-grade aluminum alloy chassis, copper base plate, six heat pipes, dual turbine fans, and advanced PCM thermal conductive material, it maintains stable cooling performance even under continuous load. This system supports 130W continuous power and 160W peak power operation, with a built-in 320W power supply. It boasts multiple global certifications including CCC, FCC, UL, CE, and UKCA, ensuring stable and reliable operation in various environments
- 【Cluster Design】Two MS-S1 MAX units can be configured as a dual-unit cluster to run a large 235B Q4 model locally, achieving an output speed of 10.87 tok/s. Supporting 2U rack deployment, multiple MS-S1 MAX units can be cascaded into a distributed cluster to create a high-efficiency AI computing center. A cluster of four MS-S1 MAX units successfully ran a DeepSeek-R1 671B Q4 large model. A reserved cluster power-on interface allows for unified start-up and shutdown
How to interpret the numbers
A tokens-per-second figure is only useful alongside the conditions that produced it. For a comparison you want to reuse, preserve at least the build identity, backend, model, workload, hardware, and whether the number measures one stream or aggregate concurrent throughput. Prompt processing and generated-token speed should not be treated as interchangeable. My results show why re-measuring matters; they do not establish that row, layer, graph, or tensor mode is best in general.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
