Architecting Long-Context LLM Acceleration with Packing-Prefetch Scheduler and Ultra-Large Capacity On-Chip Memories

Faaiq Waqar; Hanchen Yang; Harsono Simka; Ming-Yen Lee; Muhammed Ahosan Ul Karim; Shimeng Yu

REVIEW 3 major objections 2 minor 13 references

Architecting Long-Context LLM Acceleration with Packing-Prefetch Scheduler and Ultra-Large Capacity On-Chip Memories

T0 review · 3 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that co-designing a packing-prefetch scheduler with ultra-large monolithic-3D BEOL-compatible on-chip memory removes the HBM KV-cache bottleneck, achieving 8.06x decode speedup and 1.83x overall latency reduction on Llama3.

desk verdict The headline result is arithmetically suspicious: 512MB of on-chip KV cache is only ~4K tokens for Llama3.1-8B, so an 8.06x 'long-context' decode speedup can't hold unless the context is short. read the letter →

arxiv 2508.08457 v1 pith:B7IGJXJF submitted 2025-08-11 cs.AR cs.ET

Ming-Yen Lee , Faaiq Waqar , Hanchen Yang , Muhammed Ahosan Ul Karim , Harsono Simka , Shimeng Yu This is my paper · ORCID

classification cs.ARcs.ET

keywords long-contextLLMinferenceKVcacheprefetchingrequestpackingmonolithic3DmemoryBEOLembeddedHBMbandwidthTPUaccelerator

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is trying to establish that long-context LLM inference can be accelerated by co-designing the request scheduler and the memory hierarchy: pack requests to reuse and coalesce KV-cache reads, prefetch KV data before attention needs it, and place an ultra-large on-chip memory close to the compute. The authors simulate TPUv6e-like hardware with an additional 512 MB BEOL-compatible memory and report 8.06x decode speedup and 1.83x overall latency reduction on Llama3.1-8B versus serial execution, and on multi-request workloads 1.7x-2.4x throughput improvement with 1.5x-2.4x less HBM traffic than packing-only methods. A sympathetic reader should care because HBM bandwidth for KV-cache transfer is currently a well-known bottleneck in long-context serving: if the memory and scheduler assumptions hold, the bottleneck shifts and long-context inference becomes substantially cheaper.

What carries the argument

The key machinery is the packing-prefetch scheduler together with ultra-large-capacity monolithic 3D (M3D) back-end-of-line (BEOL) compatible embedded memories. The scheduler makes two moves: packing, which batches or rearranges requests so that KV-cache reads share spatial and temporal locality, and prefetching, which brings KV data from HBM or the on-chip memory into the compute path before the attention layer needs it. The BEOL memory is a memory tier stacked on top of logic with very large on-chip capacity, and it does the work of keeping a large slice of the KV cache physically close to the attention units, so the system no longer depends on HBM bandwidth for every cache access.

What would settle it

Run the same packing-prefetch scheduler on a TPUv6e-like cycle-accurate simulator with two memory configurations: the assumed 512 MB BEOL memory, and an equal-capacity memory whose bandwidth and density match the highest values reported from real monolithic-3D BEOL test chips. If the decode speedup over serial falls from 8.06x toward the packing-only result, the central claim rests on the unproven memory assumption. A second check: on real hardware or a silicon-validated simulator, measure HBM traffic during long-context decode; if the HBM interface remains saturated with packing-prefetch enab

Watch

Extended reading notes

Core claim

The paper's central claim is that the growing bottleneck in long-context LLM decode is not attention arithmetic but KV-cache transfer over HBM, and that a co-designed packing-prefetch scheduler plus an ultra-large monolithic-3D BEOL-compatible on-chip memory removes that bottleneck. Packing groups requests so KV-cache accesses have spatial and temporal locality; prefetching issues those accesses ahead of the attention unit; and the 512 MB on-chip memory keeps enough of the KV working set near the compute so many reads never cross the HBM interface. On a TPUv6e-like architecture running Llama3.1-8B, the authors report 8.06x decode speedup and 1.83x overall latency reduction over serial execut

Load-bearing premise

The result rests on a 512 MB monolithic-3D BEOL-compatible on-chip memory being buildable and correctly modeled at the assumed bandwidth and density, and on the serial and packing-only baselines being fair; if either condition fails, the headline speedups collapse.

Editorial extensions

If this is right

On Llama3.1-8B with a TPUv6e-like architecture plus 512 MB of BEOL on-chip memory, decode latency would drop to roughly one-eighth of serial execution and end-to-end latency by 1.83x.
For multi-request serving, both Llama3.1-8B and Llama3.1-70B would see 1.7x-2.4x higher throughput and 1.5x-2.4x lower HBM bandwidth demand compared with packing-only scheduling.
KV-cache transfer would stop being the resource that saturates HBM in long-context decode, so serving capacity planning would shift from HBM bandwidth to on-chip memory capacity and scheduler locality.
Packing, prefetching, and BEOL memory would form a concrete co-design path for efficient long-context LLM inference without waiting for HBM bandwidth to outpace attention compute.
These gains are claimed for TPU-like architectures in simulation; the mechanism is the co-design itself rather than any single component.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An implication the paper leaves implicit: at a fixed 512 MB on-chip capacity, combining this scheduler with KV-cache compression or quantization should extend the same packing-prefetch gains to even longer contexts, since fewer bytes per cached token means the same memory covers a larger context window.
The quoted throughput gains likely depend on request arrival patterns; a natural test is to run the scheduler under bursty versus steady request streams and measure how much of the 1.7x-2.4x comes from scheduling locality versus raw on-chip capacity.
If the central claim holds, hardware designers of long-context accelerators would have a target design point: a 512 MB-class memory tier on the compute die plus scheduler lookahead, rather than simply adding more HBM channels.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's the short version. This is a co-design proposal for long-context LLM inference: a packing-prefetch scheduler paired with 512MB of M3D BEOL-compatible on-chip memory for KV-cache staging. The abstract reports 8.06x decode speedup, 1.83x latency reduction, and 1.7-2.4x throughput gains over baseline schedulers. Those are meaningful numbers if they hold.

What's new is the specific combination — packing and prefetching have been explored separately, and BEOL/M3D memory for LLM inference is an active line (some from this same group), but I don't know of a prior co-design that puts all three together with a scheduler. That's a legitimate contribution.

I can't check the methodology because the full text I have is corrupted mojibake, with even a stray astro-ph arXiv ID in it. So I'm reviewing the abstract only. Here's the load-bearing concern. For Llama3.1-8B, per-token KV is 32 layers × 8 KV heads × 128 dims × 2 (K and V) × 2 bytes = 128KiB. A 512MB memory holds exactly 4096 tokens. To get an 8.06x decode speedup by cutting HBM traffic, you'd need to eliminate roughly 88% of KV transfers. That means the working set has to fit in about 4.7K tokens. So either the experiments used contexts under ~5K tokens (making "long-context" misleading), or the speedup is dominated by something other than on-chip capacity, or the abstract's numbers are inconsistent. The abstract doesn't give context length. This isn't a fabricated flaw; it's basic arithmetic. If they use KV compression or quantized cache, that changes the arithmetic, but the abstract doesn't say so.

Another soft spot: the evaluation compares against serial execution and packing-only baselines, but the BEOL memory's capacity, bandwidth, and latency are themselves part of the proposed design. Unless those parameters come from a real tech roadmap, some of the gains are baked into the simulation inputs. The abstract doesn't show that derivation.

So my read: the idea is worth a look, but the public claims as stated don't yet hold water. This paper belongs in front of a reviewer who can demand the context length, the simulator configuration, and the memory technology assumptions. I'd also want code/data released. If the authors can show the speedup holds at 32K or 128K context, then it's a strong result. If not, the "long-context" angle collapses.

Who's this for? People working on KV-cache management and accelerator memory hierarchies. I'd mention it to a student, but I wouldn't cite it yet. Send it to peer review — a desk reject would be premature, but the review should be rigorous.

Referee Report

3 major / 2 minor

Summary. The paper proposes a co-design of a packing-prefetch scheduler with ultra-large on-chip monolithic-3D/BEOL-compatible memories for long-context LLM inference. The abstract reports 8.06x decode speedup and 1.83x overall latency reduction for Llama3.1-8B on TPUv6e-like hardware with an additional 512 MB BEOL memory, plus 1.7x-2.4x throughput improvement and 1.5x-2.4x HBM-bandwidth reduction versus packing-only methods on Llama3.1-8B and Llama3.1-70B. The body text, however, is severely corrupted and unreadable, so the methodology, simulator details, baseline definitions, and the numerical evidence behind these claims cannot be assessed.

Significance. If the quantitative claims were substantiated, the paper would address a real bottleneck in long-context LLM serving and would make a specific case for co-designing scheduling policy with emerging BEOL memory technology. The problem statement is interesting and the proposed directions are plausible. However, the submission provides no readable methodology, no reproducible artifact, no simulator calibration, and no error analysis. As submitted, the work is not assessable; the significance of the claimed results can only be evaluated after a readable manuscript is provided.

major comments (3)

[Full text (entire body)] The body of the paper is functionally unreadable: the text is severely corrupted (mojibake), with no accessible equations, tables, algorithms, or evaluation descriptions. The abstract alone cannot support verification of the central claims. No simulator details, baseline configurations, context lengths, batch sizes, or measurement error bars are available. This is a load-bearing problem: every quantitative claim in the abstract is unsupported by accessible evidence. The authors need to resubmit in a readable form before the paper can be reviewed.
[Abstract, first paragraph] The claimed 8.06x decode speedup appears arithmetically inconsistent with the stated 512 MB on-chip capacity for Llama3.1-8B. For that model, KV cache per token is 2 (K,V) × 32 layers × 8 KV heads × 128 head_dim × 2 bytes = 128 KiB, so 512 MB can hold at most 4096 tokens. For a long context of 32K tokens, the maximum fraction of the KV working set that can be cached is about 12.8%, giving an upper-bound speedup of about 1/(1−0.128) ≈ 1.15x if HBM transfer is the bottleneck. Packing and prefetching hide latency but do not reduce HBM bytes, so they cannot close this gap. The abstract must specify the context length and batch size at which 8.06x was measured; otherwise the claim is internally inconsistent with the stated memory capacity.
[Evaluation and simulator assumptions] The BEOL memory's capacity, bandwidth, and latency are part of the proposed architecture, but the abstract presents the resulting gains as an evaluation outcome. Without sensitivity analysis or calibrated device parameters, it is unclear whether the reported 1.5x-2.4x HBM-bandwidth reduction is a result or an assumption built into the simulator. The packing-only baseline also needs a precise definition; otherwise the throughput gains cannot be separated from the effect of the memory model itself. This issue is central to the paper's contribution and cannot be checked in the current submission.

minor comments (2)

[Full text, header] The PDF header contains an arXiv identifier 'arXiv:2508.08463v1 [astro-ph.EP]', which is inconsistent with the stated paper ID and subject class. Please re-check the submission file.
[Full text, presentation] All figures, tables, equations, and cross-references are unreadable. A clean PDF with selectable text is required for any further review.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the reported speedups are simulator outputs from proposed hardware inputs, not reductions to the paper's own definitions.

full rationale

From the readable abstract, the claimed speedups are simulation results of a proposed co-design (packing-prefetch scheduler plus 512MB M3D/BEOL memories) compared against serial execution and packing-only baselines. The BEOL memory capacity, bandwidth, and placement are design inputs to the evaluation, not parameters fitted to reproduce the reported 8.06x decode speedup; the speedup is an output of the simulator under those assumptions. No equation or definition in the available text equates the predicted speedup to the input assumptions by construction, and no load-bearing self-citation is visible. The skeptic's arithmetic (512MB holds roughly 4K tokens of Llama3.1-8B KV cache under standard GQA assumptions) raises a feasibility and consistency concern about whether the claimed long-context speedup is physically achievable; that is a correctness/modeling critique, not circularity, because the derivation chain does not define the result in terms of its own inputs. The full text is corrupted and undecodable, so no further equations, references, or limitation passages could be examined; absent specific evidence of a reduction, the appropriate finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No new natural entities are proposed; the packing-prefetch scheduler and the BEOL memory system are engineered designs rather than postulated physical entities, so the graviton problem does not apply. The main unverified inputs are the assumed memory parameters and the simulator baselines.

free parameters (2)

On-chip BEOL memory capacity = 512 MB per accelerator
Design choice that determines how much KV cache can be placed on-chip; the claimed 8.06x decode speedup depends on this capacity being available, and it is an assumed hardware target, not a derived quantity (abstract).
Simulator model constants for BEOL memory and TPUv6e-like fabric = not stated in abstract
The bandwidth, latency, and energy parameters for the 512 MB BEOL memory and the TPUv6e-like baseline are not visible in the abstract and cannot be audited because the full text is corrupted; the quantitative results rest on these unstated constants.

assumptions (3)

domain assumption KV-cache transfer over HBM is the dominant bottleneck in long-context decode.
Motivating premise stated in the abstract; if false, the co-design's value proposition weakens, but the claim itself is framed in terms of this bottleneck.
domain assumption TPUv6e-like hardware plus 512 MB BEOL memory is a valid and representative evaluation platform.
The abstract says the evaluation uses this configuration; transferring the reported speedups to real hardware depends on this modeling being accurate and on the memory technology being feasible.
domain assumption Packing and prefetching are scheduling-only changes that preserve model outputs and latency semantics.
No accuracy change or approximation is mentioned; if packing alters results or prefetching misses degrade effective latency, the reported improvements need re-baselining.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Architecting Long-Context LLM Acceleration with Packing-Prefetch Scheduler and Ultra-Large Capacity On-Chip Memories." pith.science (2026). https://pith.science/paper/B7IGJXJF

@misc{pith2026250808457,
  author       = {Pith},
  title        = {Pith review of: Architecting Long-Context LLM Acceleration with Packing-Prefetch Scheduler and Ultra-Large Capacity On-Chip Memories},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B7IGJXJF}},
  note         = {Machine review of arXiv:2508.08457}
}

read the original abstract

Long-context Large Language Model (LLM) inference faces increasing compute bottlenecks as attention calculations scale with context length, primarily due to the growing KV-cache transfer overhead that saturates High Bandwidth Memory (HBM). While prefetching techniques mitigate cache misses by fetching KV data in advance, their spatial and temporal benefits present new opportunities to exploit. This work proposes a packing-prefetch scheduling architecture with monolithic 3D (M3D) back-end-of-line (BEOL) compatible embedded memories with ultra-large on-chip capacity to accelerate long-context LLM inference. Our optimizations demonstrate 8.06x decode speedup and 1.83x overall latency reduction on Llama3.1-8B using TPUv6e-like hardware with additional 512MB BEOL memories over the serial execution. Evaluations of multi-request workloads on TPU-like architectures show 1.7x-2.4x throughput improvement and 1.5x-2.4x HBM bandwidth reduction compared to packing-only methods on Llama3.1-8B and Llama3.1-70B models. With the co-design of packing, prefetching, and BEOL memories, our approach alleviates HBM constraints and enables efficient long-context LLM inference.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 3 canonical work pages

[1]

FLAT: An optimized dataflow for mitigating attention bottlenecks,

S.-C. Kao, S. Subramanian, G. Agrawal, A. Yazdanbakhsh, and T. Krishna, “FLAT: An optimized dataflow for mitigating attention bottlenecks,” in Proc. 28th ACM Int. Conf. Architectural Support for Programming Languages and Operating Systems (ASPLOS ’23), Vol. 2, Vancouver, BC, Canada, Mar. 25–29, 2023, pp. 295–310, doi:10.1145/3575693.3575747 https://dl.acm...

arXiv 2023
[2]

Sarathi: Efficient LLM inference by piggybacking decodes with chunked prefills,

A. Agrawal, A. Panwar, J. Mohan, N. Kwatra, B. S. Gulavani, and R. Ramjee, “Sarathi: Efficient LLM inference by piggybacking decodes with chunked prefills,” arXiv preprint arXiv:2308.16369, Aug. 2023, doi:10.48550/arXiv.2308.16369 https://doi.org/10.48550/arXiv.2308.16369

arXiv doi:10.48550/arxiv.2308.16369 2023
[3]

PRESERVE: Prefetching model weights and KV‑cache in distributed LLM serving,

A. C. Yüzügüler, J. Zhuang, and L. Cavigelli, “PRESERVE: Prefetching model weights and KV‑cache in distributed LLM serving,” arXiv preprint arXiv:2501.08192, Jan. 2025, doi:10.48550/arXiv.2501.08192 https://doi.org/10.48550/arXiv.2501.08192

arXiv doi:10.48550/arxiv.2501.08192 2025
[4]

CMOS+X: Stacking Persistent Embedded Memories based on Oxide Transistors upon GPGPU Platforms

F. Waqar, M.-Y. Lee, S. Yoon, S. Lim, and S. Yu, "CMOS+ X: Stacking Persistent Embedded Memories based on Oxide Transistors upon GPGPU Platforms," arXiv preprint arXiv:2506.23405, Jun. 2025, doi:10.48550/arXiv.2506.23405 https://doi.org/10.48550/arXiv.2506.23405

work page Pith review arXiv doi:10.48550/arxiv.2506.23405 2025
[5]

Timeloop: A systematic approach to DNN accelerator evaluation,

A. Parashar et al., “Timeloop: A systematic approach to DNN accelerator evaluation,” in Proc. IEEE Int. Symp. Performance Anal. Syst. Softw. (ISPASS), Mar. 2019, pp. 304–315, doi:10.1109/ISPASS.2019.00042 https://doi.org/10.1109/ISPASS.2019.00042

arXiv 2019
[6]

A-IGZO FETs with High Current and Remarkable Stability for Vertical Channel Transistor(VCT) / 3D DRAM Applications,

J.-E. Yang et al., "A-IGZO FETs with High Current and Remarkable Stability for Vertical Channel Transistor(VCT) / 3D DRAM Applications," 2024 IEEE Symposium on VLSI Technology and Circuits, Honolulu, HI, USA, 2024, pp. 1-2, doi:10.1109/VLSITechnologyandCir46783.2024.10631550 https://doi.org/10.1109/VLSITechnologyandCir46783.2024.10631550

arXiv 2024
[7]

Integration of 0.75V VDD Oxide-Semiconductor 1T1C Memory with Advanced Logic for An Ultra-Low-Power Low-Latency Cache Solution,

K.H. Chiang et al., "Integration of 0.75V VDD Oxide-Semiconductor 1T1C Memory with Advanced Logic for An Ultra-Low-Power Low-Latency Cache Solution," 2025 IEEE Symposium on VLSI Technology and Circuits, Kyoto, Japan, 2025

work page 2025
[8]

AMD Next-Generation “Zen 4

R. Bhargava and K. Troester, "AMD Next-Generation “Zen 4” Core and 4th Gen AMD EPYC Server CPUs," in IEEE Micro, vol. 44, no. 3, pp. 8-17, May-June 2024, doi:10.1109/MM.2024.3375070 https://doi.org/10.1109/MM.2024.3375070

work page arXiv 2024

Show all 13 references

[9]

Taming throughput-latency tradeoff in LLM inference with Sarathi‑Serve,

A. Agrawal et al., “Taming throughput-latency tradeoff in LLM inference with Sarathi‑Serve,” in *Proc. 18th USENIX Symp. Operating Syst. Design Implementation (OSDI)*, Santa Clara, CA, USA, July 10–12, 2024, pp. 117–134, doi:10.5555/3691938.3691945 https://dl.acm.org/doi/10.55...

2024
[10]

FlashAttention‑2: Faster attention with better parallelism and work partitioning,

T. Dao, “FlashAttention‑2: Faster attention with better parallelism and work partitioning,” arXiv preprint arXiv:2307.08691, Jul. 2023, doi:10.48550/arXiv.2307.08691 https://doi.org/10.48550/arXiv.2307.08691

2023 doi arXiv
[11]

The LLaMA 3 herd of models,

A. Dubey et al., “The LLaMA 3 herd of models,” arXiv e‑prints, Jul. 2024, arXiv:2407.21783, doi:10.48550/arXiv.2407.21783 https://doi.org/10.48550/arXiv.2407.21783

2024 doi arXiv
[12]

GPT‑4 technical report,

J. Achiam et al., “GPT‑4 technical report,” arXiv preprint arXiv:2303.08774, Mar. 2023, doi:10.48550/arXiv.2303.08774 https://doi.org/10.48550/arXiv.2303.08774

2023 doi arXiv
[13]

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEco...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

FLAT: An optimized dataflow for mitigating attention bottlenecks,

S.-C. Kao, S. Subramanian, G. Agrawal, A. Yazdanbakhsh, and T. Krishna, “FLAT: An optimized dataflow for mitigating attention bottlenecks,” in Proc. 28th ACM Int. Conf. Architectural Support for Programming Languages and Operating Systems (ASPLOS ’23), Vol. 2, Vancouver, BC, Canada, Mar. 25–29, 2023, pp. 295–310, doi:10.1145/3575693.3575747 https://dl.acm...

arXiv 2023

[2] [2]

Sarathi: Efficient LLM inference by piggybacking decodes with chunked prefills,

A. Agrawal, A. Panwar, J. Mohan, N. Kwatra, B. S. Gulavani, and R. Ramjee, “Sarathi: Efficient LLM inference by piggybacking decodes with chunked prefills,” arXiv preprint arXiv:2308.16369, Aug. 2023, doi:10.48550/arXiv.2308.16369 https://doi.org/10.48550/arXiv.2308.16369

arXiv doi:10.48550/arxiv.2308.16369 2023

[3] [3]

PRESERVE: Prefetching model weights and KV‑cache in distributed LLM serving,

A. C. Yüzügüler, J. Zhuang, and L. Cavigelli, “PRESERVE: Prefetching model weights and KV‑cache in distributed LLM serving,” arXiv preprint arXiv:2501.08192, Jan. 2025, doi:10.48550/arXiv.2501.08192 https://doi.org/10.48550/arXiv.2501.08192

arXiv doi:10.48550/arxiv.2501.08192 2025

[4] [4]

CMOS+X: Stacking Persistent Embedded Memories based on Oxide Transistors upon GPGPU Platforms

F. Waqar, M.-Y. Lee, S. Yoon, S. Lim, and S. Yu, "CMOS+ X: Stacking Persistent Embedded Memories based on Oxide Transistors upon GPGPU Platforms," arXiv preprint arXiv:2506.23405, Jun. 2025, doi:10.48550/arXiv.2506.23405 https://doi.org/10.48550/arXiv.2506.23405

work page Pith review arXiv doi:10.48550/arxiv.2506.23405 2025

[5] [5]

Timeloop: A systematic approach to DNN accelerator evaluation,

A. Parashar et al., “Timeloop: A systematic approach to DNN accelerator evaluation,” in Proc. IEEE Int. Symp. Performance Anal. Syst. Softw. (ISPASS), Mar. 2019, pp. 304–315, doi:10.1109/ISPASS.2019.00042 https://doi.org/10.1109/ISPASS.2019.00042

arXiv 2019

[6] [6]

A-IGZO FETs with High Current and Remarkable Stability for Vertical Channel Transistor(VCT) / 3D DRAM Applications,

J.-E. Yang et al., "A-IGZO FETs with High Current and Remarkable Stability for Vertical Channel Transistor(VCT) / 3D DRAM Applications," 2024 IEEE Symposium on VLSI Technology and Circuits, Honolulu, HI, USA, 2024, pp. 1-2, doi:10.1109/VLSITechnologyandCir46783.2024.10631550 https://doi.org/10.1109/VLSITechnologyandCir46783.2024.10631550

arXiv 2024

[7] [7]

Integration of 0.75V VDD Oxide-Semiconductor 1T1C Memory with Advanced Logic for An Ultra-Low-Power Low-Latency Cache Solution,

K.H. Chiang et al., "Integration of 0.75V VDD Oxide-Semiconductor 1T1C Memory with Advanced Logic for An Ultra-Low-Power Low-Latency Cache Solution," 2025 IEEE Symposium on VLSI Technology and Circuits, Kyoto, Japan, 2025

work page 2025

[8] [8]

AMD Next-Generation “Zen 4

R. Bhargava and K. Troester, "AMD Next-Generation “Zen 4” Core and 4th Gen AMD EPYC Server CPUs," in IEEE Micro, vol. 44, no. 3, pp. 8-17, May-June 2024, doi:10.1109/MM.2024.3375070 https://doi.org/10.1109/MM.2024.3375070

work page arXiv 2024