Pith Number

pith:XR4JGHSU

pith:2026:XR4JGHSU2QQVVMICQ52A3RJJOM

not attested not anchored not stored refs resolved

Exceeding the Numerical and Performance Characteristics of IEEE-754 SGEMM with BFloat16 Tensor Cores on GPUs for Scientific Computing

Addison Richards, Cherin Joseph, Cole Brower, Dmitry Lyakh, Greg Henry, Haicheng Wu, Harun Bayraktar, Jack Kosaian, John Gunnels, Lukas Mosimann, Paul Springer, Victor Podlozhnyuk

Using BFloat16 tensor cores with FP32 accumulation on GPUs exceeds the speed and numerical accuracy of native IEEE-754 FP32 SGEMM for scientific workloads.

arxiv:2605.16617 v1 · 2026-05-15 · cs.DC

Open paper page JSON Open Graph Bundle Merged state Verified badge What is a Pith Number?

Add to your LaTeX paper

\usepackage{pith}
\pithnumber{XR4JGHSU2QQVVMICQ52A3RJJOM}

Prints a linked badge after your title and injects PDF metadata. Compiles on arXiv. Learn more · Embed verified badge

Record completeness

1 Bitcoin timestamp

2 Internet Archive

3 Author claim open · sign in to claim

4 Citations open

5 Replications open

✓ Portable graph bundle live · download bundle · merged state

The bundle contains the canonical record plus signed events. A mirror can host it anywhere and recompute the same current state with the deterministic merge algorithm.

Claims

C1strongest claim

This paper examines the performance, efficiency, power, and numerical characteristics of FP32 matrix multiplication via BF16-based emulation and demonstrates how it exceeds numerical and performance characteristics of native FP32 for scientific applications.

C2weakest assumption

The assumption that BF16 tensor core operations accumulated into FP32 accumulators, combined with Blackwell-specific scaling hardware, produce results that are both faster and numerically superior to native IEEE-754 FP32 SGEMM across relevant scientific workloads without hidden accuracy losses from rounding or denormal handling.

C3one line summary

BF16 tensor cores on GPUs emulate FP32 SGEMM with superior performance, power efficiency, and numerical accuracy compared to native FP32, including a library implementation that handles denormals.

References

27 extracted · 27 resolved · 3 Pith anchors

[1] Machine behaviour 2019 · doi:10.1038/s41586-

[2] Baboulin, M., Buttari, A., Dongarra, J., Kurzak, J., Langou, J., Langou, J., Luszczek, P., and Tomov, S.Accelerating scientific computations with mixed precision algorithms.Computer Physics Communicat 2009

[3] Bayraktar, H., Charara, A., Clark, D., Cohen, S., Costa, T., Fang, Y.- L. L., Gao, Y., Guan, J., Gunnels, J., Haidar, A., Hehn, A., Hohnerbach, M., Jones, M., Lubowe, T., Lyakh, D., Morino, S., Spring 2023

[4] E., and Sherrill, C 2013 · doi:10.1021/ct400250u

[5] Dongarra, J. J., Croz, J. D., Hammarling, S., and Hanson, R. J.Algo- rithm 656: An extended set of Fortran basic linear algebra subprograms: Model implementation and test programs. 18–32

Formal links

2 machine-checked theorem links

Receipt and verification

First computed	2026-05-20T00:02:32.740281Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

bc78931e54d4215ab10287740dc529733a66a40194bdffc0efe3d6398b634e7c

Aliases

arxiv: 2605.16617 · arxiv_version: 2605.16617v1 · doi: 10.48550/arxiv.2605.16617 · pith_short_12: XR4JGHSU2QQV · pith_short_16: XR4JGHSU2QQVVMIC · pith_short_8: XR4JGHSU

Agent API

Resolver JSON Graph JSON Events JSON Schema Signing key

Verify this Pith Number yourself

curl -sH 'Accept: application/ld+json' https://pith.science/pith/XR4JGHSU2QQVVMICQ52A3RJJOM \
  | jq -c '.canonical_record' \
  | python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: bc78931e54d4215ab10287740dc529733a66a40194bdffc0efe3d6398b634e7c

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "7202c08d7290dacfc13424c7f0b728bb8c095152f517a436d62c3b8c1ceae93d",
    "cross_cats_sorted": [],
    "license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
    "primary_cat": "cs.DC",
    "submitted_at": "2026-05-15T20:37:49Z",
    "title_canon_sha256": "55ff7533e0d1ea6e01754bff217154a5b4c9af3e35c4186304d5500b3b07d9bd"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2605.16617",
    "kind": "arxiv",
    "version": 1
  }
}