Pith. sign in

REVIEW 4 major objections 5 minor 27 references

The paper claims that a purpose-built, memory-heavy decode accelerator can serve frontier Mixture-of-Experts models at a fraction of the GPU capital cost, with two fixed chip ratios explaining why mainstream GPUs waste compute during LLM de

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 07:23 UTC pith:LHLUCE23

load-bearing objection F/S and F/B framing is useful, but the economic headline rests on vendor numbers and a strawman GPU baseline; treat the design argument seriously, the cost claims as unverified. the 4 major comments →

arxiv 2607.13068 v1 pith:LHLUCE23 submitted 2026-07-10 cs.AR

The Economics of AI Decoding Chips: Rebalancing Compute, Capacity, and Bandwidth for Efficient LLM Inference

Michael J. Yuan , Ju Long This is my paper
classification cs.AR
keywords LLM inferencedecode phasememory-boundMixture-of-Expertsroofline modelF/B and F/S ratioscommodity DDR5 acceleratorDeepSeek-R1 671B
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

LLM decoding reads model weights from memory and does almost no arithmetic, yet every mainstream GPU bundles enormous compute with too little memory. This paper reduces the resulting waste to two fixed chip ratios — compute per memory bandwidth (F/B) and compute per memory capacity (F/S) — and plots every major accelerator on those axes. It then argues that the efficient "corner" is empty and that a purpose-built PCIe accelerator using commodity DDR5 fills it. The paper reports that a 4U server of four four-chip HTX-301 cards holds DeepSeek-R1 671B for about $28,000 and serves two users at a deterministic 20.3 tokens/s each, yielding about $12 per million output tokens against roughly $21 for an eight-GPU node at matching concurrency. If the throughput figure holds, the barrier to serving a frontier model drops from a $350,000 capital project to a workstation-priced purchase for low-concurrency deployments.

Core claim

The paper's central claim is that transformer decoding is memory-bound — each token reads model weights and performs only about two FLOPs per weight byte — so a chip's decode efficiency is set by two fixed design constants: F/B (peak 8-bit tensor compute divided by memory bandwidth, the roofline ridge point) and F/S (peak BF16 compute per GB of memory). Datacenter GPUs and TPUs sit at high values of both, so they idle below roughly a hundred concurrent users; SRAM accelerators sit at the opposite extreme; the low-low "efficient corner" is empty. The concrete discovery is that a purpose-built 28 nm PCIe accelerator using commodity DDR5, the HTX-301, lands in that corner and can hold DeepSeek-

What carries the argument

The argument rides on two fixed per-chip ratios and one cost identity. F/B — peak dense 8-bit tensor FLOPs over peak memory bandwidth — is the roofline ridge point; it sets how many concurrent users a batch needs before the arithmetic units saturate. F/S — peak dense BF16 TFLOPS per GB — sets how much compute a buyer is forced to purchase just to hold a model's weights; for a 671B MoE, the model size dictates a mandatory compute purchase. Plotting all accelerators on (F/S, F/B) reveals an empty "efficient corner" below about 1 TFLOP/GB and a ridge point below 64. The MoE tax (E/k = 32 for DeepSeek-R1) multiplies the concurrency requirement and makes GPUs untenable below hyperscale. Equation

Load-bearing premise

The headline numbers rest on vendor-supplied figures: the HTX-301's deterministic 20.3 tokens/s, the ~$28,000 integrated price, and an assumed power draw well below maximum rating; if sustained real throughput is lower or power is higher, the $12 per million tokens cost erodes, although the design-level F/B and F/S argument does not depend on those figures.

What would settle it

Measure a 4U HTX-301 system running the full 4-bit DeepSeek-R1 671B under two concurrent users for a sustained session, record tokens/s and wall power, and substitute the measured aggregate rate and power into Equation (2). If per-user throughput falls below about 12 tokens/s at that price, the per-token cost crosses the paper's $21 per million GPU-node figure; any material drop below 20.3 tokens/s shrinks the claimed saving. The design-level claim could also be tested by recomputing F/S and F/B from independent teardown or datasheet numbers using the same 8-bit basis for all devices and check

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Below hyperscale, frontier MoE models can be served from a $19K–$28K commodity-memory device instead of a $350K eight-GPU node.
  • At matched concurrency, the per-token cost is about $12 per million tokens versus $21, a roughly 40% saving that stays flat as the pool scales.
  • Concurrency becomes a linear purchase: replicating 4U servers adds users at the same per-token cost, avoiding the GPU path's step-function entry cost.
  • The design uses only commodity DDR5 and a mature 28 nm logic node, avoiding HBM, CoWoS, and EUV — all rationed inputs in the AI buildout.
  • The accelerator naturally fits disaggregated serving, with a GPU handling prefill and a pool of these cards handling the memory-bound decode phase.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The F/S–F/B map could be reused to size decode accelerators for future MoE models: as the E/k ratio changes, the saturation concurrency shifts, so the optimal ridge point moves with the model architecture.
  • If independent benchmarks confirm the 20.3 tokens/s figure, the same rebalancing logic likely extends to other memory-bound workloads — long-context KV-cache serving, retrieval, and embedding — where the two ratios predict similar GPU idling.
  • The paper's own concession that the per-token gap is modest implies the economic case rests more on capital entry and supply resilience than on operating cost; a change in power draw or amortization period would shift the crossover.
  • A smaller-process successor would raise single-user speed but re-enter expensive EUV and LPDDR5X cost tiers; the paper flags this trade but leaves the exact performance-versus-supply-curve unquantified.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper formalizes the inefficiency of GPU-based LLM decoding using two per-chip constants: F/B (the roofline ridge point) and F/S (compute per GB of memory). It argues that datacenter GPUs and even memory-rich consumer devices sit far outside the 'efficient corner' defined by these constants, and that a deliberately rebalanced accelerator—less compute, more commodity memory, lower bandwidth—would serve frontier MoE decoding below hyperscale at far lower capital and per-token cost. As a concrete instance, the paper analyzes the Skymizer HTX-301, a 28nm PCIe DDR5 accelerator, and claims that a 4U server of four 4-chip cards serves DeepSeek-R1 671B at a deterministic 20.3 tokens/s per user for about $28,000, with cost per million tokens about $12 versus $21 for an 8×H100 node. The design-level argument is transparent and largely independent of the vendor-specific numbers; the economic headline, however, rests on vendor-supplied throughput and prices and on a GPU baseline that may not be the true minimum.

Significance. The paper's two-constant framework is a useful and clearly presented way to reason about decode-oriented hardware, and the MoE-tax argument convincingly shows why large sparse models worsen the utilization problem at low concurrency. The supply-chain reasoning—avoiding HBM, CoWoS, and leading-edge logic—is a genuine contribution and is not tied to the specific throughput number. If the quantitative claims were independently validated, the paper would make a strong case for a class of memory-heavy, low-compute accelerators. As it stands, the central economic comparison is vulnerable because (i) the baseline GPU deployment may be unrepresentative (4×H200 is a natural cheaper alternative that the paper itself lists in Table 1), and (ii) the HTX-301's throughput and price are vendor-provided with no benchmark protocol or public artifact. The design argument survives these concerns; the economic headline needs substantial hardening.

major comments (4)
  1. [§6.1, Table 2] The 'minimum GPU deployment' for DeepSeek-R1 is taken to be an 8×H100 node at $350K, but the paper's own Table 1 lists the H200 with 141 GB. Four H200s provide 564 GB, enough for the ~400 GB model plus KV cache, and aggregate bandwidth ~19 TB/s is far above the ~10.5 GB/token sparse read at batch 16. At 2026 street prices of $25–40K per GPU plus chassis, a 4×H200 server is roughly $120–180K; plugging into Eq. (2) gives approximately $8–14 per million tokens at 16 users, at or below the claimed $12/M. If true, the order-of-magnitude capital advantage and the $12-vs-$21 comparison are artifacts of an unrepresentative baseline. The authors must add the 4×H200 configuration as a baseline or explicitly justify its exclusion (availability, software, power, or other constraints).
  2. [§6.1, Eq. (2), Table 2] The paper's central quantitative claims—20.3 tokens/s deterministic throughput, $28K integrated price, and the resulting $12/M—are vendor-supplied inputs with no benchmark protocol, software stack, measured power draw, or public artifact. The $12/M figure is obtained by substituting these inputs into Eq. (2), not by an independent model. If sustained throughput is lower due to PCIe overhead, quantization kernels, prefill integration, or expert-routing contention, the per-token cost advantage shrinks or disappears. Please provide a sensitivity analysis over throughput and price, or an independent benchmark, before the economic claims can be evaluated.
  3. [§6.1, Table 2, Eq. (2)] The ownership model in Eq. (2) includes a power term W·ce, but the paper never states the sustained power draw W for the HTX-301 4U system or the 8-card system. The reported $12/M depends on an unstated W. The four-chip card is rated at 164 W, and a 4U system with four such cards would draw at least ~0.66 kW plus host overhead, which is material to the cost calculation. Please report the assumed W for each row of Table 2 and check the sensitivity of $/M to W.
  4. [§6, 'Note' on F/B basis] The paper notes that HTX-301 and Mac Studio ridge points use BF16 rates because no 8-bit tensor rate is published, while GPU ridge points use 8-bit. The note says the ordering is unchanged, but the quantitative claim that the HTX-301 is 'an order of magnitude below every GPU' on ridge point depends on this basis choice. Since 8-bit rates are typically 2× BF16, a consistent 8-bit basis would move the HTX-301 and Mac Studio by a factor of 2. The qualitative conclusion likely survives, but the exact 'order of magnitude' statement should be re-verified on a common basis.
minor comments (5)
  1. [Eq. (2)] Variables P_acq, W, and c_e are introduced after the equation; define them before or in the equation for readability.
  2. [Figure 1/4 captions] The 'efficient corner' is shaded but not defined in the caption; the text gives F/B < 64 and F/S < 1 later, but the reader must hunt for it. State the thresholds in the captions.
  3. [§6, Mac Studio] The Mac Studio's F/S is computed using 96 GB, the post-shortage maximum. If the pre-shortage 512 GB option were used, F/S would be ~0.13 and the device would sit inside or at the edge of the corner. The paper should state why 96 GB is the relevant configuration.
  4. [References] Reference [21] has two URLs and a somewhat confusing parenthetical; split or clarify. Reference [23] gives 5nm wafer pricing, but the text claims 28nm costs 'roughly 6× less per wafer'; a direct 28nm price source would be more appropriate.
  5. [§7] The prefill phase is only mentioned as a disaggregation option; for a paper about decoding chips this is acceptable, but a sentence quantifying prefill cost or stating it is out of scope would help set expectations.

Circularity Check

0 steps flagged

No circularity: the economic figures are transparent arithmetic from stated inputs; the F/B–F/S design argument is independent.

full rationale

The paper's central derivation chain is not circular. The F/B and F/S constants, the roofline/concurrency argument, the MoE tax (E/k = 32 for DeepSeek-R1), and the identification of an empty efficient corner are computed from vendor datasheets and model architecture, independent of the economic headline. The quantitative comparison in §6.1 is an explicit cost-accounting exercise: Equation (2) defines $/M = C_hour/R, and the HTX-301 figures are obtained by substituting the vendor-supplied $28K system price, 2 × 20.3 tokens/s throughput, and power estimate. The paper does not claim these are predictions derived from F/B or F/S; it labels the rate as 'rated' and the prices as integrated. The resulting $12/M is therefore dependent on unvalidated vendor inputs, and the $350K eight-H100 baseline may be unrepresentative (e.g., a four-H200 node could hold the model), but dependence on inputs is a data-quality concern, not circularity. There is no self-citation chain: reference [22] is the vendor's announcement, not prior work by the authors, and the acknowledgments disclose the vendor relationship. The F/B–F/S design thesis would survive even if the specific $12/M figure were wrong, so the score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central quantitative comparison rests on vendor-supplied throughput and BOM prices plus an ownership model with chosen amortization and electricity assumptions. The efficient-corner boundaries are also chosen to match the paper's target concurrency, so the headline cost numbers are arithmetic consequences of inputs rather than independently derived predictions.

free parameters (5)
  • HTX-301 integrated-system throughput = 20.3 tokens/s per user at 2 users
    Vendor-rated; used in Eq. (2) to compute $12/M; no measurement methodology in the paper.
  • HTX-301 system prices and BOM = $7K (4S-128GB), $19K (8S-512GB), $28K (4U), $224K (8×4U)
    Vendor-supplied integrated prices; drive all cost comparisons and the main capital-cost claims.
  • Efficient-corner thresholds = F/S < ~1 TFLOP/GB; F/B < ~64 FLOP/byte
    Chosen from the paper's 16-user target and ~4 FLOP/byte INT4 decode intensity; not derived from first principles.
  • Ownership-model parameters = 2-year amortization; $0.20/kWh; W=5.0 kW for 8×H100
    Inputs to Eq. (2); varying them changes $/M proportionally, and power draws are estimates rather than measurements.
  • HTX-301 system power draw W = not stated; implied ~0.66 kW from 4×164 W cards
    Required for Eq. (2) but not reported; affects the $/M calculation.
axioms (6)
  • standard math Roofline model: a workload saturates compute only when arithmetic intensity exceeds F/B
    Invoked in §2 via [1] to define concurrency-to-saturate.
  • domain assumption Decode arithmetic intensity is ~2 FLOP per weight byte (≈4 FLOP/byte at INT4)
    Used throughout §2–3 to convert F/B into users-to-saturate; assumes transformer linear-layer counts without overhead.
  • domain assumption DeepSeek-R1 has 256 experts with top-8 routing, 37B active parameters per token, ~400 GB at 4-bit
    Taken from DeepSeek technical report and third-party benchmarks [12, 14].
  • domain assumption An 8-GPU node's KV cache holds ~2,200 concurrent sequences at 32K context
    Used as the practical batch ceiling in Figure 2; source is a third-party deployment benchmark [14] not reproduced here.
  • domain assumption HBM and CoWoS are supply-constrained into 2026–27, and 28 nm wafer cost is ~6× cheaper than leading edge
    Basis of the supply-chain advantage; from Epoch/SemiAnalysis and Tom's Hardware, not independently quantified in the paper.
  • domain assumption Mac Studio M3 Ultra delivers ~65 BF16 TFLOPS and 819 GB/s
    Apple publishes no TFLOPS figure; the paper derives it from core count, clock, and FP32 rate [20].

pith-pipeline@v1.3.0-alltime-deepseek · 3609 in / 3703 out tokens · 171647 ms · 2026-08-02T07:23:12.148693+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of The Economics of AI Decoding Chips: Rebalancing Compute, Capacity, and Bandwidth for Efficient LLM Inference." pith.science (2026). https://pith.science/paper/LHLUCE23

@misc{pith2026260713068,
  author       = {Pith},
  title        = {Pith review of: The Economics of AI Decoding Chips: Rebalancing Compute, Capacity, and Bandwidth for Efficient LLM Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LHLUCE23}},
  note         = {Machine review of arXiv:2607.13068}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Every mainstream GPU is built compute-heavy and capacity-light: it pairs enormous arithmetic throughput with too little memory to hold a modern model. In contrast, large language model decoding requires little compute and a large amount of memory: a GPU's floating-point units run at single-digit-percent utilization during decoding, and the memory the workload does need is sold only bundled with yet more compute. The compute is recovered only at hyperscale, where Mixture-of-Experts (MoE) models are spread across 96--320-GPU expert-parallel clusters serving thousands of concurrent users, a scale available to a handful of operators. We formalize the inefficiency with two fixed per-chip constants. F/B, the roofline ridge point, determines whether the compute can be utilized; F/S, the compute bundled with each GB of memory, determines how much compute must be bought. We then argue for a rebalanced decode accelerator: less compute, far more commodity memory, and a deliberately lower and cheaper bandwidth. The Skymizer HTX-301, a purpose-built 28nm PCIe accelerator using commodity DDR5, occupies that design point. Its entry cost is low. A single eight-chip card holds DeepSeek-R1 671B for about \$19,000, and a 4U server of four four-chip cards serves two users at a deterministic 20.3 tokens per second each for about \$28,000. Either costs less than a single H100, while the minimum GPU deployment for the model is an eight-GPU node near \$350,000. Concurrency then scales out by adding hardware: eight 4U servers carry sixteen users for about \$224,000, two-thirds of the node's price, with the cost per token unchanged at about \$12 per million against the node's \$21. The HTX-301's decisive advantage is a supply chain free of every rationed input: it uses no high-bandwidth memory, no CoWoS, and no leading-edge logic.

Figures

Figures reproduced from arXiv: 2607.13068 by Ju Long, Michael J. Yuan.

Figure 1
Figure 1. Figure 1: The figure plots each accelerator by F/S (horizontal, TFLOPS/GB) and F/B (vertical, [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The figure plots, for each GPU, the concurrency needed to saturate its compute on a dense [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The figure plots cost per million output tokens for DeepSeek-R1 against GPU cluster size [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The figure repeats the F/S versus F/B map of Figure 1 with both HTX-301 configurations [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 4 linked inside Pith

  1. [1]

    Roofline: an insightful visual performance model for multicore architectures.Communications of the ACM52(4), 65–76, 2009

    Williams, S., Waterman, A., & Patterson, D. Roofline: an insightful visual performance model for multicore architectures.Communications of the ACM52(4), 65–76, 2009

    2009

  2. [2]

    NVIDIA. RTX Spark: slim laptops and small desktops (announced at Computex, June 2026; up to a 6,144-core Blackwell RTX GPU, a 20-core Grace CPU, 128 GB unified memory, and 1 petaflop of FP4 AI performance). https://www.nvidia.com/en-us/products/rtx-spark/

    2026

  3. [3]

    Nvidia unveils RTX Spark Superchip for laptops and desktop PCs at Computex

    Tom’s Hardware. Nvidia unveils RTX Spark Superchip for laptops and desktop PCs at Computex

  4. [4]

    G., et al

    Recasens, P. G., et al. Mind the memory gap: unveiling GPU bottlenecks in large-batch LLM inference. arXiv:2503.08311, 2025. https://arxiv.org/abs/2503.08311

    Pith/arXiv arXiv 2025

  5. [5]

    Ironwood: the first Google TPU for the age of inference

    Google Cloud. Ironwood: the first Google TPU for the age of inference. April 2025. https: //blog.google/products/google-cloud/ironwood-tpu-age-of-inference/

    2025

  6. [6]

    GroqCard accelerator product spec sheet (v1.5)

    Groq, Inc. GroqCard accelerator product spec sheet (v1.5). https://cdn.sanity.io/files/chol0sk5/pr oduction/ec9c0e1b3f1e07f7389df879af71529abd66d662.pdf

  7. [7]

    Cerebras announces third-generation Wafer-Scale Engine

    Cerebras Systems. Cerebras announces third-generation Wafer-Scale Engine. March 2024. https: //www.cerebras.ai/press-release/cerebras-announces-third-generation-wafer-scale-engine

    2024

  8. [8]

    Introducing Cerebras inference: AI at instant speed

    Cerebras Systems. Introducing Cerebras inference: AI at instant speed. August 2024. https: //www.cerebras.ai/blog/introducing-cerebras-inference-ai-at-instant-speed

    2024

  9. [9]

    HGX B200 datasheet

    NVIDIA. HGX B200 datasheet. https://www.primeline-solutions.com/media/categories/server/nac h-gpu/nvidia-hgx-h200/nvidia-blackwell-b200-datasheet.pdf 12

  10. [10]

    Instinct MI300X data sheet

    AMD. Instinct MI300X data sheet. https://www.amd.com/content/dam/amd/en/documents/instinc t-tech-docs/data-sheets/amd-instinct-mi300x-data-sheet.pdf

  11. [11]

    DGX Spark

    NVIDIA. DGX Spark. https://www.nvidia.com/en-us/products/workstations/dgx-spark/

  12. [12]

    DeepSeek-V3 technical report

    DeepSeek-AI. DeepSeek-V3 technical report. arXiv:2412.19437, 2024. https://arxiv.org/abs/2412 .19437

    Pith/arXiv arXiv 2024

  13. [13]

    DeepGEMM: grouped GEMM kernels (BLOCK M = 128)

    DeepSeek-AI. DeepGEMM: grouped GEMM kernels (BLOCK M = 128). https://github.com/dee pseek-ai/DeepGEMM

  14. [14]

    DeepSeek V3/R1 deployment benchmarks, 8×H100 (AWQ 4-bit, vLLM)

    dzhsurf. DeepSeek V3/R1 deployment benchmarks, 8×H100 (AWQ 4-bit, vLLM). https://github.c om/dzhsurf/deepseek-v3-r1-deploy-and-benchmarks

  15. [15]

    MoE inference economics from first principles

    Tensor Economics. MoE inference economics from first principles. 2025. https://www.tensorecon omics.com/p/moe-inference-economics-from-first

    2025

  16. [16]

    DeepSeek-V3/R1 inference system overview (Open Infra Index, Day 6)

    DeepSeek-AI. DeepSeek-V3/R1 inference system overview (Open Infra Index, Day 6). https: //github.com/deepseek-ai/open-infra-index/blob/main/202502OpenSourceWeek/day 6 one mor e thing deepseekV3R1 inference system overview.md

  17. [17]

    Deploying DeepSeek with PD disaggregation and large-scale expert parallelism on 96 H100 GPUs

    LMSYS. Deploying DeepSeek with PD disaggregation and large-scale expert parallelism on 96 H100 GPUs. 2025-05-05. https://www.lmsys.org/blog/2025-05-05-large-scale-ep/

    2025

  18. [18]

    AI chip supply-chain constraints (packaging and HBM bottleneck)

    Epoch AI. AI chip supply-chain constraints (packaging and HBM bottleneck). https://epoch.ai/dat a-insights/ai-chip-supply-chain-constraints

  19. [19]

    AI capacity constraints: CoWoS and HBM supply chain

    SemiAnalysis. AI capacity constraints: CoWoS and HBM supply chain. https://newsletter.semiana lysis.com/p/ai-capacity-constraints-cowos-and

  20. [20]

    Mac Studio (M3 Ultra) technical specifications, 80-core GPU and 819 GB/s unified memory, https://support.apple.com/en-us/122211

    Apple. Mac Studio (M3 Ultra) technical specifications, 80-core GPU and 819 GB/s unified memory, https://support.apple.com/en-us/122211. Apple publishes no TFLOPS figure; the compute is derived from the architecture, 80 cores × 128 FP32 ALUs × 2 FLOP × ∼1.6 GHz ≈ 33 TFLOPS FP32, and about 65 TFLOPS at FP16/BF16 (twice the FP32 rate)

  21. [21]

    Mac Studio (2025) technical specifications, https://support.apple.com/en-us/122211 (M3 Ultra announced with up to 512 GB, Newsroom, March 2025)

    Apple. Mac Studio (2025) technical specifications, https://support.apple.com/en-us/122211 (M3 Ultra announced with up to 512 GB, Newsroom, March 2025). During the 2026 memory shortage the 128, 256, and 512 GB options were withdrawn and 96 GB became the maximum: Tom’s Hardware, March 2026, and Macworld, June 2026, https://www.tomshardware.com/tech-i ndustr...

    2025

  22. [22]

    HTX-301 announcement (April 2026) and company profile v1.0 (March 2026)

    Skymizer Taiwan Inc. HTX-301 announcement (April 2026) and company profile v1.0 (March 2026)

    2026

  23. [23]

    TSMC wafer prices by process node

    Tom’s Hardware. TSMC wafer prices by process node. https://www.tomshardware.com/news/tsmc s-wafer-prices-revealed-300mm-wafer-at-5nm-is-nearly-dollar17000

  24. [24]

    DistServe: disaggregating prefill and decoding for goodput-optimized large language model serving

    Zhong, Y ., et al. DistServe: disaggregating prefill and decoding for goodput-optimized large language model serving. OSDI 2024. arXiv:2401.09670. https://arxiv.org/abs/2401.09670

    Pith/arXiv arXiv 2024

  25. [25]

    (Microsoft Azure Research)

    Patel, P., Choukse, E., et al. (Microsoft Azure Research). Splitwise: efficient generative LLM inference using phase splitting. arXiv:2311.18677, 2023. https://arxiv.org/abs/2311.18677 13

    Pith/arXiv arXiv 2023

  26. [26]

    DeepSeek-R1 provider pricing

    Artificial Analysis. DeepSeek-R1 provider pricing. https://artificialanalysis.ai/models/deepseek-r1/ providers 14

  27. [2026]

    2026-06-01. https://www.tomshardware.com/laptops/nvidia-unveils-rtx-spark-superchip-a t-computex-2026-new-platform-promises-to-turn-windows-into-an-agentic-ai-os-with-arm-cpu -blackwell-gpu-and-128gb-unified-memory

    2026