Summary
- The MI325X data sheet prints five AI precision rows with two figures each under a heading reading
AI PEAK THEORETICAL PERFORMANCE with sparsity, and a separate HPC block with one figure per row, so one page carries 81.7 and 5,229.8 as the same kind of noun. - The naming of the two columns appears on AMD's product pages, the doubling rule appears in prose in the CDNA 3 white paper, and a third AMD table prints a single unqualified column; the independent ranking that measures these parts orders machines by a dense, full-64-bit achieved rate and uses theoretical peak only to break ties.
One page, eight numbers
The rate table is the first thing on the MI325X accelerator data sheet that looks like arithmetic, and it is worth reading slowly. Under the heading AI PEAK THEORETICAL PERFORMANCE with sparsity it prints five rows at five precisions, each with two figures: TF32 at 653.7 and 1,307.4 TFLOPs; FP16 at 1,307.4 and 2,614.9; BFLOAT16 at the same pair; INT8 at 2,614.9 and 5,229.8 TOPS; FP8 at 2,614.9 and 5,229.8 TFLOPS. Even the units shift inside the block, because the INT8 row counts integer operations per second while its neighbours count floating-point ones. Further down the same sheet a separate block headed HPC PEAK THEORETICAL PERFORMANCE (TFLOPS) gives one number per row: FP64 vector 81.7, FP32 vector 163.4, FP64 matrix 163.4, FP32 matrix 163.4 (MI325X data sheet).
So the smallest and largest numbers on that page, 81.7 and 5,229.8, describe the same physical part performing the same class of operation, and the ratio between them is roughly sixty-four to one. A procurement document that carries away one figure and not the other has carried away one of several printed quantities, not a summary of the part.
What the heading covers, and what it leaves outside
The visible qualification for the two AI columns is the heading itself, with sparsity, plus one sentence: Footnote explanations are available at: followed by the address of AMD's claims page. That is the whole of it on the sheet. Nothing in the retrieved text labels a cell as dense or sparse, and I state that as a limit of what I obtained rather than as a property of the document: both AI column labels failed to survive extraction, and footnote superscripts are among the first casualties of a text render.
On the MI300X data sheet the same table carries one extra note, *The MI200 Series does not support TF32, FP8, or sparsity*, which tells a reader something about the previous generation and nothing about which column is which (MI300X data sheet). The MI300X platform data sheet repeats the heading and adds AMD's own framing of what the feature buys: With the introduction of sparsity, AI models lacking dense data structures can be accelerated with even greater efficiency (platform data sheet). The word sparsity is doing an enormous amount of work in that sentence, and the sheet still does not tell you which of its two columns assumes it.
Where the pair gets names
The labelling is on the product pages, where each precision row is printed twice with names attached. For the eight-bit format: Peak Eight-bit Precision (FP8) Performance (E5M2, E4M3) at 2.61 PFLOPs beside Peak Eight-bit Precision (FP8) Performance with Structured Sparsity (E5M2, E4M3) at 5.22 PFLOPs. FP16 appears at 1.3 against 2.61; TF32 at 653.7 TFLOPs against 1.3 PFLOPs with structured sparsity; INT8 at 2.6 against 5.22. The FP64 and FP32 rows carry no sparsity-qualified variant on either page (MI325X product page, MI300X product page).
Two details matter here. First, the pairing is exactly the pairing the sheet leaves unnamed, so the mapping 2,614.9 → dense and 5,229.8 → sparse is a fact of a second document, not of the table. Second, the product pages round where the sheet is exact: 2.61 PFLOPs stands in for 2,614.9 TFLOPS. Both printings are true; only one is recomputable from the other without a rounding step.
The doubling, stated in prose
The rule behind the second column is written down, just not where the numbers are. The CDNA 3 architecture white paper states that The Matrix Core can take advantage of situations where at least two values within a group of four input values are zero (50%+ sparsity), and that this enables doubling the computational throughput (CDNA 3 white paper). The same document carries AMD's deferred endnote MI325-019 as its condition set: Based on internal measurements and calculations by AMD Performance Labs as of February 2025, on an AMD Instinct MI325X GPU (1000W), 256GB HBM3E, at 2,100 MHz peak boost against an MI250X at 1,700 MHz, with per-clock multipliers of 256 for FP64 matrix, 128 for FP64, 256 for FP32, 1,024 for TF32, 2,048 for FP16 and BFLOAT16, 4,096 for FP8, closing on results may vary.
Read those two passages together and the architecture of the whole sheet becomes visible. A theoretical peak is a clock rate multiplied by a per-clock, per-compute-unit count. Sparsity is a separate multiplier of two. The document states both, in two different places, in prose. The table that prints the result prints the multiplied number without saying that it has been multiplied.
A third table, one column
AMD's ROCm documentation prints the same class of figure a third time, and there the sparsity dimension is absent altogether. Its table lists Matrix FP64 163.4, Vector FP64 81.7, Matrix FP32 163.4, Vector FP32 163.4, Vector TF32 653.7, Matrix FP16 1,307.4, Matrix BF16 1,307.4, Matrix FP8 2,614.9 and Matrix INT8 2,614.9 TFLOPS, described as the theoretical peak performance of the OAM (Instinct MI300 architecture documentation). Every value matches the first column of the data sheet. It also labels the TF32 row Vector TF32, a path distinction the data sheet leaves unstated for the same number.
So a reader comparing the sheet with the documentation has three facts to reconcile and only two of them appear in either table: the AI rows on the sheet and the rows in the documentation agree at every precision, which tells you the documentation is describing the dense column, and the TF32 row carries a path label in one document and not in the other.
The ranking that measures instead
Independent of all of this, TOP500 maintains two columns and an explicit ordering rule. Rmax - Maximal LINPACK performance achieved, and Rpeak - Theoretical peak performance: In the TOP500 List table, the computers are ordered first by their Rmax value. In the case of equal performances (Rmax value) for different computers, we have chosen to order by Rpeak. (TOP500 description)
The measurement behind Rmax is narrow on purpose. The benchmark solves a dense system of linear equations; the operation count must be 2/3 n^3 + O(n^2) double-precision floating-point operations, and that excludes the use of fast matrix multiplication algorithms (such as Strassen's method), algorithms that compute a solution in a precision lower than full 64-bit floating-point arithmetic and then refine it iteratively, and methods that emulate floating-point arithmetic in software using integer arithmetic (The Linpack Benchmark). The FAQ adds the sentence that fixes the convention: In order to have an entry included in the Linpack Benchmark report the results must be computed using full precision. By full precision we generally mean 64 bit floating point arithmetic or higher. It describes the theoretical peak as a paper computation, an upper bound, and the number manufacturers often cite (TOP500 FAQ).
That is a measured, dense, full-64-bit number, ordered on first, with theoretical peak demoted to a tiebreak. It is the opposite arrangement from a spec sheet whose headline table is theoretical, printed twice per row, and headed by a sparsity condition. Two conventions, each internally coherent, neither one wrong.
The inconsistency I recorded and did not bridge
One thing in the reviewed set does not close. The white paper's Table 1 lists MI325X FP16, BFLOAT16 and INT8 at 1,024 FLOPS per clock per compute unit against 2,048 for MI300X, while both parts print the same 1,307.4 and 2,614.9 peak figures at the same 2,100 MHz. The retrieved excerpt does not explain the difference, and I am not going to invent the bridge. It sits in the record as an unexplained internal inconsistency in a document that elsewhere states its conditions carefully.
My arithmetic, labelled as mine
The following divisions and doublings are mine, applied to the printed values, not statements the documents make.
653.7 × 2 = 1,307.4, exact. 1,307.4 × 2 = 2,614.8, and the sheet prints 2,614.9, so the FP16 pair is one-tenth of a teraflop above exact doubling and the ratio is 2.0001. 2,614.9 × 2 = 5,229.8, exact. In the HPC block, 81.7 × 2 = 163.4, which is exact for both the FP64 matrix row and the FP32 vector-to-matrix relationship. And 5,229.8 ÷ 81.7 ≈ 64.0, which is the spread between the largest AI figure and the smallest HPC figure on one sheet.
Whether those relationships are the intended reading of the two columns is a different question, and one the sheet does not answer in its own text. The arithmetic shows the columns are related two-to-one; the white paper's sparsity sentence explains why a two-to-one relationship would exist; the product page says which side is which.
What a reader has to hold before the number means anything
Three conditions travel with every peak in that table and none of them is in the cell. The precision format is in the row label. The sparsity assumption is in the heading. The theoretical-versus-measured status is in the words peak theoretical, unqualified, twice. A fourth condition, the matrix-or-vector path, appears for some rows in the ROCm documentation and not in the data sheet.
There is also a scoped absence worth stating plainly, because it is a fact about ten documents rather than about the world: among the seven AMD documents and three TOP500 pages I read, no AMD document cross-references TOP500, and no TOP500 page names these accelerators. Whatever comparison a buyer makes between the two conventions, neither publisher has made it for them. AMD's own company record sits in our directory at Advanced Micro Devices, Inc..
Reading limits. Every AMD and TOP500 passage above came from retrieval of the exact URLs, not from a native PDF or browser render, so the two AI column labels and any footnote superscripts may have been lost with the layout. The claims page the data sheet points to was not retrieved, and nothing here reports what it says. No TOP500-listed system built on these accelerators was examined. No price, order, shipment, revenue or customer claim is involved.
The documents read. MI325X data sheet · MI300X data sheet · MI325X product page · MI300X product page · CDNA 3 white paper · MI300X platform data sheet · Instinct MI300 architecture documentation · The Linpack Benchmark · TOP500 description · TOP500 FAQ.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






