LiquidGEMM: LLM Kernel and Detector Concepts

• LiquidGEMM is a hardware-efficient 4-bit weight GEMM kernel optimized for LLM inference on NVIDIA Hopper GPUs, reducing dequantization overhead through co-design.
• For cryogenic detectors, LiquidGEMM-like structures use misaligned GEMs in dual-phase liquid argon TPCs to provide optical decoupling and electron transmission without avalanche multiplication.
• The dual usage of LiquidGEMM underscores the importance of distinguishing low-bit GEMM innovations from GEM-inspired microstructures in detector research.

LiquidGEMM denotes two unrelated research concepts that share a near-identical name but belong to distinct technical domains. In high-performance computing for LLM serving, “LiquidGEMM” refers to a hardware-efficient $W4A8$ GEMM kernel built for NVIDIA Hopper-class GPUs, centered on the co-design of quantization, dequantization, data layout, and warp-group pipelining to mitigate the dequantization bottleneck in asymmetric low-bit matrix multiplication (Hu et al., 1 Sep 2025). In cryogenic detector R&D, the term is relevant by analogy rather than as the formal title of a detector: it describes GEM-like perforated structures used not as avalanche multipliers in liquid, but as liquid-filled electron-transmitting and optically decoupling barriers for dual-phase liquid-argon time projection chambers, especially in configurations involving photosensitive dopants (Ramos et al., 2024). A related line of work extends the GEM-like concept toward thick electroluminescence structures, or FAT-GEMs, for dense argon gas under cryogenic-detector conditions (Rodas-Rodríguez et al., 3 Feb 2026). The shared name is therefore potentially misleading: one literature concerns low-bit tensor kernels for LLM inference, the other GEM-like microstructures for noble-liquid and cryogenic detectors.

1. Nomenclature and scope

The computational LiquidGEMM is introduced in “LiquidGEMM: Hardware-Efficient W4A8 GEMM Kernel for High-Performance LLM Serving” (Hu et al., 1 Sep 2025). Its subject is quantized matrix multiplication for LLM decoding, with 4-bit weights and 8-bit activations. The detector-side work does not use “LiquidGEMM” as the paper title, but it is directly relevant to “LiquidGEMM-like ideas” for liquid-argon TPCs because it proposes “two stacked GEMs with offset patterns, operated fully in the liquid phase, beneath a gas pocket in traditional dual-phase LArTPCs, with applied voltage too small to induce electroluminescence” (Ramos et al., 2024).

This dual usage is not merely terminological. The computational work addresses throughput limits arising from online dequantization in asymmetric GEMM; the detector work addresses optical feedback and charge-transfer problems in doped dual-phase LArTPCs. The overlap is lexical, not methodological.

A plausible implication is that any encyclopedic treatment of “LiquidGEMM” must explicitly separate these threads to avoid conflating GEMM, meaning general matrix multiplication in numerical linear algebra, with GEM-like structures, meaning gaseous-electron-multiplier-inspired detector elements.

2. LiquidGEMM in LLM serving

In the LLM-serving literature, LiquidGEMM is a $W4A8$ GEMM kernel optimized for serving workloads in which dequantization overhead had previously erased the theoretical advantage of 4-bit weights (Hu et al., 1 Sep 2025). The motivation is a mismatch between roofline expectations and realized throughput: $W4A8$ should reduce weight bandwidth relative to $W8A8$ while preserving the advantages of 8-bit activations, yet prior kernels remained bottlenecked by CUDA-Core-side dequantization before Tensor Core MMA.

The paper argues that $W4A8$ is a favorable operating point because it “gives most of the accuracy benefits of 8-bit activations while cutting weight storage and bandwidth in half relative to W8A8, and improving arithmetic intensity relative to both W8A8 and W4A16” (Hu et al., 1 Sep 2025). It also explicitly contrasts $W4A8$ with $W4A4$, stating that heavily quantized activations can harm model quality. However, existing practice did not realize these benefits. On LLaMA2-7B at batch size 256, the paper reports that an existing $W4A8$ GEMM could be about $2\times$ slower than $W8A8$, which it attributes to hardware inefficiency in the dequantization path rather than to an inherent defect of the quantization format (Hu et al., 1 Sep 2025).

The kernel’s central claim is that efficient $W4A8$0 execution requires co-design across arithmetic representation and scheduling. LiquidGEMM therefore combines two techniques: LiquidQuant, abbreviated LQQ, and an implicit fine-grained pipeline, abbreviated ImFP (Hu et al., 1 Sep 2025). The former reduces the instruction cost of dequantization; the latter overlaps weight loading, dequantization, and MMA across warp groups without a separate dequant stage handing data to a separate MMA stage through shared memory.

The system is specialized to NVIDIA Hopper-class GPUs, especially H100 and H800, because the implementation depends on Hopper features such as WGMMA, warp groups, and TMA (Hu et al., 1 Sep 2025). The paper does not claim portability to older architectures.

3. Quantization design and dequantization path

LiquidQuant is a two-level quantization framework. First, weights are quantized from FP16 to INT8 using affine quantization, with group-wise quantization and a protected first-level INT8 range $W4A8$1 (Hu et al., 1 Sep 2025). Second, rather than mapping signed INT8 directly to UINT4, LiquidQuant shifts the signed INT8 range into an unsigned UINT8 domain and then quantizes UINT8 to UINT4. This construction is used to make dequantization overflow-safe while preserving a form expressible in efficient native integer instructions.

The paper’s crucial runtime identity is that dequantization can be rewritten so that the final INT8 result is obtained by a packed integer multiply-add followed by a top-bit flip. Operationally, the runtime formula is

$W4A8$2

where $W4A8$3 (Hu et al., 1 Sep 2025). The construction depends on the facts that $W4A8$4, that $W4A8$5, and that the intermediate sum remains within UINT8 range under the protected INT8 bounds.

The paper states that four elements can be dequantized with just two hardware arithmetic instructions, specifically IMAD and XOR; including unpacking from packed UINT4 to bytewise form, eight elements are dequantized with seven instructions total (Hu et al., 1 Sep 2025). This instruction count is the main quantitative reason the scheme can keep pace with Tensor Core consumption. The authors further argue that, on H100/H800-class hardware, hiding dequantization requires the per-element dequantization instruction count $W4A8$6 to be roughly at or below 5. Prior dequantizers exceeded that threshold, especially once overflow handling was included (Hu et al., 1 Sep 2025).

The critique of prior work is concrete. QServe, the primary baseline, also uses two-level $W4A8$7 quantization, but its dequantization involves a vadd-style packed subtraction path that is not a native hardware instruction and is lowered to around a dozen low-level operations. Nsight profiling reported in the paper attributes 21% of warp stalls in QServe on the FFN layer of LLaMA2-7B to subtraction involving vadd (Hu et al., 1 Sep 2025).

4. Kernel organization and performance

LiquidGEMM reformulates the GEMM as

$W4A8$8

rather than $W4A8$9, in order to better match Hopper’s INT8 WGMMA instruction shapes, especially at small batch sizes (Hu et al., 1 Sep 2025). Within each thread block, the output is tiled over $W4A8$0, and the main loop advances through $W4A8$1 in chunks of $W4A8$2. The thread block uses warp groups of four warps, or 128 threads, organized as one Load WG and two Compute WGs (Hu et al., 1 Sep 2025).

A major hardware-facing component is the Dual-MMA packed layout. Because one MMA requires only 16 UINT4 elements per thread but an LDS.128 fetches 128 bits, or 32 four-bit elements, the weights are reordered offline so that a thread’s operands for two consecutive MMAs are contiguous in one 128-bit chunk (Hu et al., 1 Sep 2025). The same layout is used in global memory and shared memory, enabling LDG.128 and LDS.128 while reducing memory instruction count and address-generation overhead.

The implicit fine-grained pipeline distinguishes LiquidGEMM from a more obvious explicit staged design. In an explicit coarse-grained pipeline, one would have a Load WG, a Dequant WG, and an MMA WG, with synchronization and shared-memory handoff between stages. The paper argues that this incurs two penalties: register-to-shared-memory-to-register round trips for dequantized weights, and software synchronization bubbles (Hu et al., 1 Sep 2025). In ImFP, each Compute WG both dequantizes and immediately consumes the dequantized values in WGMMA, while another Compute WG may already be at a different phase. The Load WG acts as a producer and the Compute WGs as consumers in a single-producer, multiple-consumer model (Hu et al., 1 Sep 2025).

The reported gains are substantial. At batch size 256, kernel-level speedups over QServe are 2.75× on LLaMA2-7B, 2.87× on LLaMA2-13B, and 2.90× on LLaMA2-70B (Hu et al., 1 Sep 2025). The abstract states “up to 2.90x speedup over state-of-the-art W4A8 kernels and up to 4.94x end-to-end system-level speedup” (Hu et al., 1 Sep 2025). Compared to TensorRT-LLM quantized kernels, the paper reports 1.12–1.63× performance gains and up to 1.63× system-level speedup (Hu et al., 1 Sep 2025). Representative end-to-end decoding results include 6,721 tok/s for LLaMA2-7B at batch 194, 4,105 tok/s for LLaMA2-13B, 3,695 tok/s for LLaMA2-70B at batch 184, 6,999 tok/s for Yi-34B, and 10,745 tok/s for Mixtral-8×7B at batch 256 (Hu et al., 1 Sep 2025).

The gains are not uniform across all regimes. The paper states that at very small batch sizes, especially for MoE or GEMV-like regimes, TensorRT-LLM’s specialized kernels can outperform LiquidGEMM, particularly TRT-W4A16 and TRT-FP8 below batch 32 on Mixtral-8×7B (Hu et al., 1 Sep 2025). This suggests that LiquidGEMM is optimized for mainstream decode and serving workloads in which GEMM dominates and sufficient work exists to exploit the pipeline.

5. LiquidGEMM-like detector concepts in dual-phase liquid argon

In detector physics, the term “LiquidGEMM” is relevant to a specific use of GEM-like structures in liquid argon rather than to a formal detector name. The key problem is introduced in “Production of GEM-like structures using laser-cutting techniques” (Ramos et al., 2024). The paper considers dual-phase LArTPCs with photosensitive dopants whose ionization energies lie below the scintillation threshold of argon. Such dopants can increase ionization yield, which is advantageous for low-energy sensitivity, but gas-phase electroluminescence photons can propagate back into the liquid, photoionize the dopant, and trigger a positive feedback loop (Ramos et al., 2024).

The proposed solution is an optically blocking but electron-transmitting barrier between the liquid and gas, realized with two stacked GEM-like structures having misaligned holes (Ramos et al., 2024). Their operating mode is explicitly not liquid-phase avalanche multiplication. Instead, the holes are intended to remain filled with liquid argon, and the bias is kept low enough that drifting electrons are collected into and transmitted through the holes without amplification. The paper states that “incident drifting electrons are drawn into the holes but not amplified. Instead, amplification will occur in the gas phase above the structures, as typical for dual-phase TPCs” (Ramos et al., 2024).

This architecture repurposes GEM geometry from a multiplication structure into what may be described, in editor’s terms, as a “transmission-shield barrier.” The two misaligned perforated sheets provide optical decoupling because photons generated in the gas region lack a direct line of sight back into the liquid (Ramos et al., 2024). A plausible implication is that the principal innovation lies in interface management—charge transfer combined with optical isolation—rather than in microscopic gain.

The paper does not provide quantitative detector-performance validation. It explicitly omits field maps, extraction efficiencies, transparency values, charge gains, optical attenuation factors, and detector-physics formulas (Ramos et al., 2024). Consequently, the concept is clear at the architectural level but remains experimentally incomplete at the level of liquid-argon transport and optical suppression.

6. Materials, fabrication, and cryogenic electroluminescence structures

The thin GEM-like structures for optical decoupling are based on polyethylene naphthalate, or PEN, with 125 $W4A8$3m thickness emphasized for the self-vetoing application, though laser tests were performed on 25, 50, and 125 $W4A8$4m PEN sheets (Ramos et al., 2024). The first batch used a hexagonal hole pattern with 200 $W4A8$5m diameter and 1000 $W4A8$6m pitch. The electrodes are transparent conductive polymer coatings of PEDOT:PSS on both sides, with target thickness about 50 nm and room-temperature sheet resistance about $W4A8$7 (Ramos et al., 2024). PEN is chosen partly because it scintillates in the visible spectrum, leading the authors to argue that “the risk of increased radioactivity due to the larger mass compared to traditional wire grids is negated by the potential to veto its own radioactivity” (Ramos et al., 2024).

Fabrication proceeds by ultrasonic cleaning in Alconox and ultrapure water, rinsing, drying, perimeter masking, spray deposition of diluted PEDOT:PSS with a Paasche Talon TG-3F airbrush, curing at 80 °C for 30 min, and only then laser drilling of the hole patterns (Ramos et al., 2024). Two laser systems were studied. A 445 nm, 6 W blue diode laser could drill only 25 $W4A8$8m PEN with minimal visible burn marks; no clean-hole configuration was found for 50 and 125 $W4A8$9m PEN (Ramos et al., 2024). A 55 W CO$W8A8$0 laser, xTool P2, produced successful results on 50 and 125 $W8A8$1m PEN, with no visible burning and a minimum consistent hole diameter of about 200 $W8A8$2m. For 125 $W8A8$3m PEN, the paper reports a minimum hole diameter of 0.2 mm with $W8A8$4 mm (Ramos et al., 2024).

Electrical stability tests are preliminary but nontrivial. All produced PEN samples were tested successfully up to 800 V “without observing a significant increase in the drained current,” while one structure intentionally driven toward breakdown reached 7 kV bias and then experienced significant discharges that damaged the structure (Ramos et al., 2024). These results show insulating behavior after laser processing, but they do not constitute detector operation in liquid argon.

A related but distinct structure is the FAT-GEM, discussed in “Production of GEM-like structures for cryogenic applications, using laser-cutting techniques” (Rodas-Rodríguez et al., 3 Feb 2026). FAT-GEM denotes a Field Assisted Transparent Gaseous Electroluminescence Multiplier: a 5 mm-thick PMMA structure designed for optical electroluminescence readout in dense argon gas, below charge multiplication (Rodas-Rodríguez et al., 3 Feb 2026). It uses magnetron-sputtered ITO on one face, with 400 $W8A8$5 surface resistivity and 90% transparency at 420 nm, and evaporated aluminum of approximately 400 nm thickness on the opposite face (Rodas-Rodríguez et al., 3 Feb 2026). The structure further incorporates TPB inside the holes as a wavelength shifter, with emission around 420 nm (Rodas-Rodríguez et al., 3 Feb 2026).

The 2026 paper reports operation in pure argon at 4 bar, described as “at a gas density close to the one expected in LAr conditions” (Rodas-Rodríguez et al., 3 Feb 2026). The best reported energy resolution is $W8A8$6 FWHM at 5.9 keV, obtained at $W8A8$7 using a 5×5 cm$W8A8$8, 5 mm-thick conical-hole structure with 2 mm diameter, 4 mm pitch, TPB in holes, and ITO + Al electrodes (Rodas-Rodríguez et al., 3 Feb 2026). The paper also reports an average 21% improvement in light yield over previous productions, stable operation up to $W8A8$9 at 4 bar, and no evidence of charging-up effects (Rodas-Rodríguez et al., 3 Feb 2026). However, it does not demonstrate actual liquid-argon operation, saturated-vapor operation above liquid, or dual-phase extraction through the holes (Rodas-Rodríguez et al., 3 Feb 2026).

7. Significance, limitations, and potential misconceptions

The principal misconception associated with LiquidGEMM is that there is a single unified research program under that name. In fact, the 2025 LiquidGEMM paper concerns LLM-serving kernels, whereas the detector-related papers concern GEM-like structures for cryogenic detectors and are only “LiquidGEMM-like” in conceptual relevance (Hu et al., 1 Sep 2025, Ramos et al., 2024). The shared name should therefore be treated as homonymous.

A second misconception concerns the detector-side use of GEM-like structures. The 2024 paper does not propose conventional GEM amplification in liquid argon. It proposes low-bias liquid-filled transmission structures whose purpose is electron guidance and optical decoupling, with amplification or electroluminescence only in the gas above the barrier (Ramos et al., 2024). Likewise, the 2026 FAT-GEM paper demonstrates dense-gas electroluminescence, not in-liquid electroluminescence (Rodas-Rodríguez et al., 3 Feb 2026).

A third misconception concerns maturity. In HPC, LiquidGEMM is presented as already deployed in the authors’ production LLM serving infrastructure and evaluated in a full serving stack including SmoothQuant, PagedAttention, FlashAttention-2, and INT8 KV-cache quantization (Hu et al., 1 Sep 2025). By contrast, the detector-side work remains at a fabrication and enabling-technology stage. The 2024 paper explicitly states that “First tests indicate good electrical stability, while the performance assessment is still ongoing” (Ramos et al., 2024). The crucial detector metrics that remain unproven include electron transparency through one or two liquid-filled perforated layers, optical blocking effectiveness of the misaligned double layer, extraction into gas, long-term cryogenic HV reliability, PEDOT:PSS behavior in LAr over time, and compatibility with photosensitive dopants (Ramos et al., 2024).

Taken together, the literature shows that “LiquidGEMM” currently names one mature software-hardware co-design for $W4A8$0 LLM inference and one broader family of detector ideas in which GEM-like microstructures mediate electron transport, optical decoupling, or electroluminescent readout in cryogenic noble-detector systems. The former is a specific kernel architecture with quantified throughput improvements on Hopper GPUs (Hu et al., 1 Sep 2025). The latter is an emerging detector-design space supported by fabrication feasibility, preliminary electrical stability, and dense-gas EL demonstrations, but still lacking direct validation in the liquid-phase regimes that define its ultimate scientific use case (Ramos et al., 2024, Rodas-Rodríguez et al., 3 Feb 2026).