Draft 04 | Primary source: the Solidigm × Giga Computing solution brief (PDF, all 4 pages downloaded and parsed). Verification date: 2026-11-15. Note: the original solidigmtechnology.com URL redirects to the same path on solidigm.com; the web-fetch tool does not support PDFs, so the file was downloaded and its full text extracted with pypdf. PDF metadata shows creation on 2026-09-03 and revision on 2026-09-04.

The Most Expensive Component in the Cluster Is Idling

Official Solidigm D5-P5336 product photo (E1.L, E3.S and U.2 form factors)
Official Solidigm D5-P5336 product photo (E1.L, E3.S and U.2 form factors)

The most hidden waste in a training cluster is its most expensive component sitting idle. H100s bill by the hour, and data has to clear several gates on the way from the data lake to GPU memory — storage reads, network transfer, preprocessing, copies. If any gate lags, the GPU waits. The brief's "customer challenge" section is refreshingly restrained: storage must keep pace with the GPUs, but the network, the memory hierarchy, and the preprocessing pipeline all affect efficiency too, and the system has to be viewed as a whole. That posture is far more credible than "swap the drives and everything is fixed" marketing. Storage is a common bottleneck — not the only one.

Solidigm (the former Intel NAND business, an SK hynix subsidiary since 2021-12) teamed with Giga Computing (spun out of Gigabyte) to validate a D7-PS1010 + D5-P5336 combination on the MLPerf Storage v1.0 benchmark. What follows unpacks the brief as written.

MLPerf Storage v1.0: A Storage Exam That Needs No Real GPUs

Solidigm official graphic: D5-P5336 on the path to 122TB capacity
Solidigm official graphic: D5-P5336 on the path to 122TB capacity

MLPerf Storage, maintained by MLCommons, doesn't test the drive in isolation — it tests whether a storage system can feed training data to accelerators fast enough. The clever part is that accelerators are simulated in software: real data is read at the intensity of real training, then the simulator "sleeps" for the measured compute time of each batch (what the brief calls think time). The data path stays fully real; only the arithmetic is skipped. That means storage pressure can be reproduced without a single physical GPU, and integrators can run low-cost proofs of concept.

The headline metric is accelerator utilization, AU, defined in the brief's appendix as AU = total compute time ÷ total benchmark runtime × 100%. Version 1.0 runs three workloads: Unet3D (medical 3D image segmentation, dominated by large-file sequential reads), ResNet50 (image classification on a 50-layer residual network, small files at a high sample rate), and CosmoFlow (a cosmology 3D convolutional network, bandwidth-hungry on large samples). A supplementary source — MLCommons' own explanation via Blocks & Files — gives v1.0 average sample sizes: roughly 114,660 bytes for ResNet50, 146.6 MB for Unet3D, and 2.83 MB for CosmoFlow. That spread explains why the three stress storage in completely different shapes.

The brief reports AU results of >90% for Unet3D, >90% for ResNet50, and >70% for CosmoFlow — consistent with MLPerf's pass/fail rule based on AU thresholds (MLCommons' official page states the current round's Unet3D threshold is 90%; the v1.0 thresholds for all three workloads appear in the supplementary source). One anti-gaming rule is worth memorizing: the dataset must be at least 5× the total memory of the client nodes, so the benchmark measures storage, not the memory cache.

D7-PS1010 + D5-P5336: Who Does What, and the Measured Gains

Official Solidigm D7-PS1010 close-up render
Official Solidigm D7-PS1010 close-up render

The brief spells out the division of labor: the D7-PS1010 is the PCIe 5.0 performance tier, built for low-latency direct-attached storage; the D5-P5336 is the PCIe 4.0 QLC capacity tier, aimed at network-attached use (per the official Chinese product brief: up to 61.44 TB, with read performance on par with TLC). One handles "fast," the other handles "big."

Measured figures (brief, ResNet50, simulated NVIDIA H100):

MetricSolidigm D7-PS1010Micron 9550 (as cited in the brief)
AU92.68%90.33%
GPUs supported7058
Throughput115,805 samples/s; 12,663 MB/sBrief states Solidigm is nearly 24% higher

On the D5-P5336 side: ResNet50 throughput 22% higher than the Micron 6500 ION (brief, verbatim claim).

A few calculated values, labeled as such: 70 versus 58 GPUs is about 20.7% more accelerators supported; the AU gap is 2.35 percentage points; and 12,663 ÷ 70 ≈ 181 MB/s of sustained read demand per GPU. On that basis, an 8-GPU node running ResNet50-class training alone needs roughly 1.45 GB/s of stable read bandwidth (181 × 8 ≈ 1,448 MB/s). That's the starting point for sizing storage.

Test platform (brief appendix): a Gigabyte R163-Z35 (1U, single AMD 9555P, 64 cores — AMD's site confirms EPYC 9005 series, 64 cores/128 threads, 12-channel DDR5, PCIe 5.0), Samsung 64 GB DDR5 ×12 (768 GB total, calculated), with Intel 1.6 TB U.2 ×2 as the system drives.

The sentence that must be highlighted: the brief's appendix explicitly states that the Micron results came from a different server system and tuning configuration — this was not a same-condition controlled experiment. Quote the "24%/22% lead" without that premise and you've turned a vendor brief into an objective review. It's the easiest sentence in the document to skip, and the one that most affects how much the conclusions deserve trust.

From the official launch release (2024-08-06): the PS1010 spans 1.92–15.36 TB, rated 1 DWPD over 5 years, with 28 PBW lifetime writes on the 15.36 TB version, in E3.S/U.2 form factors. Solidigm claims up to 70% better efficiency than comparable drives; the footnote basis is the 7.68 TB version against Samsung's PM1743, at 3,874 versus 2,272 IOPS/W.

Checkpoints and Data-Lake Tiers: Every Drive Doing the Right Job

MLCommons' official description of checkpointing deserves to be quoted in full: while checkpoints are written, training stalls synchronously; on recovery, the entire cluster waits; fsync forces data to media, and no write cache can hide slow drives. The current official round uses the Llama-3 family as its example — a 70B-class model writes roughly 911 GiB per checkpoint. Do the estimate (calculated): 911 GiB ≈ 978 GB; at 10 GB/s of sustained write bandwidth a checkpoint takes about 98 seconds; at 3 GB/s it takes about 5.4 minutes. Every failure recovery and every save cycle converts that difference into real, billable training time.

So the tiering logic writes itself: the checkpoint/cache tier needs sustained writes and low latency — give it the PS1010. For heavier write duty, the same family's PS1030 is an option (official release: write performance more than doubled versus the previous generation, 70 PBW on the 12.8 TB version, explicitly recommended for cache and checkpoint use). Hand the dataset lake to big-capacity QLC like the P5336 — it's positioned as read-optimized, so don't make it a checkpoint target; QLC's sustained-write performance and endurance are its weak spots (editorial judgment). Below that sits the object-storage/HDD cold tier. The brief notes that BeeGFS, Mangoboost, and NCHC have expressed interest in the solution — indirect confirmation that "parallel filesystem plus tiered SSDs" is the mainstream deployment shape.

Sizing AI Cluster Storage: A Checklist for Guangzhou and PRD Integrators

Calculate bandwidth before buying drives: per-GPU demand × GPU count × a redundancy factor. Mind the workload spread — the ~181 MB/s per GPU (calculated) for ResNet50-class jobs is a small-file scenario only. MLCommons' current-round data shows Unet3D-class large-file workloads demanding up to ~5.9 GiB/s per GPU (on the v3.0/B200 basis, not this brief's v1.0/H100). That's an order of magnitude apart; pick the wrong reference and you'll spec the wrong drives.

Don't chase Gen5 blindly. Gen4 QLC is enough for the capacity tier; save Gen5 for the performance tier. Network-attached designs need the network math too: 100GbE is theoretically ~12.5 GB/s (calculated), and the fastest drives are pointless behind a 25GbE link. Most installed PRD customer networks are still 10/25GbE, so upgrade budgets must include switches and NICs. Check endurance in PBW (28 PBW for the 15.36 TB PS1010, official), and for clusters that checkpoint frequently, compute annual write volume first.

Run MLPerf Storage as the pre-delivery acceptance test: it's open source, needs no real GPUs, and lets you demand AU targets instead of just peak bandwidth. One localization note: the brief's test platform uses the EPYC 9555P, which AMD's regional availability listing marks "China: Not available" — domestic BOMs can't be copied verbatim; pick an equivalent SKU that's actually available in China. Risk flags: a vendor benchmark is not your workload; AU is sensitive to dataset and cache configuration (the 5× memory rule exists precisely for that); "70 GPUs supported" is a simulation result, not a production promise; and the brief's competitive comparisons are not same-condition — cite them with that caveat attached.

Sources

  1. Solidigm × Giga Computing, *Optimizing AI Training with Solidigm™ SSDs and Giga Computing Servers* (solution brief, PDF) — https://www.solidigmtechnology.com/content/dam/solidigm/en/site/products/technology/gigabyte-solidigm-ai-training-solution-brief/documents/Solidigm-Giga-Optimizing-AI-Training-Solutions-Brief.pdf (access redirects to the same path on solidigm.com) — PDF metadata: created 2026-09-03, revised 2026-09-04; downloaded and parsed in full 2026-11-15.
  2. MLCommons, *MLPerf Storage* official benchmark page — https://mlcommons.org/benchmarks/storage/ — accessed 2026-11-15 (the page currently presents v3.0 round rules; v1.0 workload composition follows the brief).
  3. Solidigm Newsroom, *Solidigm Extends AI SSD Portfolio Leadership with Lightning-Fast Additions to their D7 Product Line* (D7-PS1010/PS1030 launch release) — https://news.solidigm.com/en-WW/240188-solidigm-extends-ai-ssd-portfolio-leadership-with-lightning-fast-additions-to-their-d7-product-line/ — 2024-08-06.
  4. Solidigm, *D5-P5336 Product Brief* (in Chinese) — https://www.solidigmtechnology.cn/products/technology/d5-p5336-product-brief.html — 2023-06-27.
  5. AMD, *EPYC 9555P* product page — https://www.amd.com/zh-tw/products/processors/server/epyc/9005-series/amd-epyc-9555p.html — accessed 2026-11-15 (product launch date 2024-10-10).
  6. Gigabyte, *R163-Z35-AAH1 (rev. 3.x)* product page — https://www.gigabyte.cn/Enterprise/Rack-Server/R163-Z35-AAH1-rev-3x — accessed 2026-11-15.
  7. DOIT (compiled from Blocks & Files, including MLCommons' official explanation), *MLPerf Storage Benchmark: The Official Interpretation from MLCommons* (in Chinese) — https://www.doit.com.cn/p/521156.html — 2023-07-13 (supplementary source: v1.0 sample sizes and AU thresholds).