# Kernel and benchmark dashboard This is the one-page technical overview of `quantem.gpu`: what the scientific kernels compute, where each runtime implements them, how parity is proved, and what the latest retained measurements actually mean. ```{admonition} Read the state before comparing the number :class: important First-process source load, prepared-source load, warm resident compute, and saved-result reopen are different experiments. A binned or cropped source is never presented as native resolution. Check the complete provenance ledger before using a number in a design or release decision. ``` **Dashboard review:** 2026-08-22. Every current measured timing below shows the device, test date, and exact source revision. The overview omits opaque evidence IDs and does not replace the [complete benchmark provenance ledger](performance/results.md). (coverage-and-next-runs)= ## Coverage and next runs The [filterable coverage registry](performance/coverage.md) now keeps every required configuration visible, including configurations that have never run. It separates complete measurements, partial evidence, pending work, refuted experiments, and fail-closed unsupported paths. Each open row names a stable runbook, its physical owner, and the exact artifact required for promotion. ```{include} _generated/benchmark_coverage.md :start-after: :end-before: ``` Agents and maintainers should begin with: ```bash python scripts/benchmark_registry.py next --limit 10 python scripts/benchmark_registry.py command GATE_ID ``` The overview tables below retain current measurements, qualified probes, and clearly marked diagnostics. Read the **State** cell before the time. The coverage registry is the canonical place to find missing combinations and their reproduction entry points. (speed-and-memory-at-a-glance)= ## Speed and memory at a glance The rows below are deliberately not a leaderboard: results with different fixtures, cache states, scientific plans, or wall-clock boundaries are not ranked against one another. Each row keeps those conditions together. Measurement tables are keyed first by **Platform**, then by the reproducible **Computer** class; local host nicknames are never public benchmark identifiers. ### Current qualified load measurements One row is one exact configuration on one reproducible computer. Platform and computer are the first two columns; scan geometry, detector geometry, bin, dtype, cache state, statistic, memory observations, device, date, and revision remain separate data. The table is generated from the benchmark registry so a superseded result cannot remain the dashboard headline. ```{include} _generated/benchmark_coverage.md :start-after: :end-before: ``` These measurements use a complete `512x512` scan and native `192x192 uint16` source detector, with scan bin 1 and no crop. The current Python MPS, Native Swift/Metal, and CPU rows use fixture `real-512x512x192x192-u16-bslz4-27shard-master-fixture-c`. The physical WebGPU smoke uses the separately fingerprinted `full-native-webgpu-512x512x192x192-u16` fixture; it is not a cross-fixture speed comparison. Every row retains its own fixture SHA-256 and source identity in the complete registry. Detector bin is explicit. Exact binned `uint16` output is source-identity-bound to its retained maximum-count audit; it is not a general license to narrow arbitrary input. For this immutable source, the complete maximum-count audit of 53 proves that an 8x8 exact detector sum is at most 3,392, so bins through 8 fit in `uint16`. Other sources must pass their own complete range audit or use a provably sufficient wider integer dtype. The package rows are prepared-index measurements. Python MPS source pages were uncontrolled after one same-process warmup. Native Swift/Metal uses prepared immutable indexes and controlled `F_NOCACHE` source descriptors. Neither is a cold arbitrary-source or application end-to-end result. Historical CUDA, WebGPU, earlier CPU, destination-reuse, and superseded measurements remain in [Verified benchmark results](performance/results.md) and the complete [coverage registry](performance/coverage.md); they are not silently discarded or mixed into this current table. Logical resident bytes, accelerator/driver peak, process RSS high-water, process physical-footprint peak, and swap delta are distinct observations and are not additive. A missing physical-footprint value is displayed as `n/a`, not inferred from RSS. Process RSS does not include every direct Metal allocation, so it cannot replace accelerator or physical-footprint telemetry. For example, the current full-native Python MPS row reports **18.442 GiB** of driver allocation for an 18.00 GiB logical resident tensor; that is a measured after-load boundary, not a universal peak estimate. On the 24 GB MacBook Pro, the full-native bin-1 smoke passed every exact volume and product hash but produced approximately 723.31 MiB additional swap use and 758,448,128 B of swapouts. The safety gate therefore stopped after one trial; it is partial evidence, not a repeatable performance distribution. (dtype-support-and-peak-memory)= ### Dtype support and peak memory “Source,” “staging,” “accumulation,” and “resident” dtype describe different stages. A `uint8` row is scientifically exact only when the source is already `uint8` or a source-identity-bound complete audit proves `maximum <= 255` and `pixelsAbove255 == 0`. Otherwise a saturating `uint8` output, such as the WebGPU clip8 browse decode, clips values above 255 and is a browse representation, not raw-count evidence. The registry below separates those cases. Each row fixes one platform, computer class, scan, detector, bin, crop, source dtype, staging dtype, resident dtype, and scientific gate. **Browse-only** can be benchmarked, but it cannot satisfy an exact scientific gate. ```{include} _generated/benchmark_coverage.md :start-after: :end-before: ``` This capability matrix does not infer peak memory. The current load table is authoritative only for its matching shape, dtype, device, cache state, and wall boundary. In particular, browser RSS may be smaller than a WebGPU resident payload because it does not capture every device allocation; that is an incomplete peak, not evidence that the payload disappeared. The current C matrix measures native `uint16` input. The full-native logical payload is exactly 19,327,352,832 bytes (18.00 GiB). Python `io.load` on CUDA and MPS keeps that complete native detector and holds it ANS encoded, so its resident bytes depend on the acquisition, not on a detector bin. Detector bins are explicit options of the native Swift/Metal and WebGPU loaders and exact views of the encoded resident in the CUDA browse service. The native Swift/Metal primitive has admitted one full 18.00 GiB resident load on the 24 GB Apple M5, but repeated uncontended memory and paging evidence remains incomplete. A consumer may apply a conservative working-set limit; that application policy is separate from primitive capability and may not silently select detector bin 2. The current production WebGPU detector-bin-2/4/8 path accumulates into and stores `float32`. Those historical timings remain useful implementation history, but they do not satisfy the exact-integer resident contract. The platform/computer registry therefore marks exact WebGPU bins 2/4/8 **blocked** until integer accumulation and residency pass full-volume parity. Seven-run physical exact full-volume distribution evidence remains in the retained-measurement table; unavailable WebGPU device-allocation telemetry is still shown as `n/a`, never inferred from browser RSS. Public Python `io.load` keeps stored counts: omit `dtype` or pass `dtype="native"`. Its one conversion is an explicit `dtype="scaled_uint16"` for float32 sources, which stores calibrated integer codes, stays ANS encoded, and reports the measured conversion error. The CUDA, Python MPS and CPU reference rows above for uint8 outputs, uint32 widening and audited uint8 staging are therefore **Not supported**: those were load-time casts that `io.load` no longer offers. (minimum-device-memory-gates)= ### Minimum-device memory gates The public release floors are **6 GiB of dedicated VRAM for CUDA** and **8 GB of total laptop RAM for WebGPU**. The WebGPU number is the entire machine budget shared by the operating system, browser, JavaScript heap, staging buffers, and GPU—not memory available exclusively to one `GPUBuffer`. **✓** is awarded only after the complete load-and-product pipeline runs on a physical device at or below the stated floor with retained peak memory, pressure/swap, output parity, and responsiveness evidence. A calculated payload can prove **No**, but it can only establish a **Pending** candidate. An allocator cap on a larger GPU is a useful pre-check, not physical-device signoff. | Platform | Computer | Minimum device | Selected scan | Scan plan | Source detector | Detector bin | Output detector | Resident dtype | Resident payload | Gate | Reason | Device tested | Date tested | |---|---|---|---:|---|---:|---:|---:|---|---:|---|---|---|---| | [**CUDA**](platforms/cuda.md) | Physical 6 GiB CUDA computer pending | 6 GiB VRAM | `512x512` | Full | `192x192` | 1 | `192x192` | `uint16`, ANS encoded | **1.98 to 3.14 GiB** | **Pending** | Measured encoded payload of three real acquisitions; complete physical 6 GiB peak not retained | — | — | | [**Python MPS**](platforms/mps.md) | MacBook Air (M2, 8 GB) | 8 GB unified RAM | `512x512` | Full | `192x192` | 1 | `192x192` | `uint16`, ANS encoded | **Data-dependent** | **Pending** | Encoded payload and physical 8 GB pressure/parity run not retained | — | — | | [**Native Swift/Metal**](platforms/swift-metal.md) | MacBook Air (M2, 8 GB) | 8 GB unified RAM | `512x512` | Full | `192x192` | 1 | `192x192` | `uint16` | **18.00 GiB** | **Blocked** | Resident payload exceeds physical memory | — | — | | [**Native Swift/Metal**](platforms/swift-metal.md) | MacBook Air (M2, 8 GB) | 8 GB unified RAM | `512x512` | Full | `192x192` | 2 | `96x96` | `uint16` | **4.50 GiB** | **Pending** | Current clean-revision pressure/parity run not retained | — | — | | [**Native Swift/Metal**](platforms/swift-metal.md) | MacBook Air (M2, 8 GB) | 8 GB unified RAM | `512x512` | Full | `192x192` | 4 | `48x48` | `uint16` | **1.125 GiB** | **Historical** | Earlier physical run passed; current clean-revision repeat remains pending | Apple M2 MacBook Air (`Mac14,2`, 8 GB) | 2026-08-18 | | [**Native Swift/Metal**](platforms/swift-metal.md) | MacBook Air (M2, 8 GB) | 8 GB unified RAM | `512x512` | Full | `192x192` | 8 | `24x24` | — | — | **Not supported** | Current native load-plan contract supports bins 1, 2, and 4 | — | — | | [**WebGPU**](platforms/webgpu.md) | MacBook Air (M2, 8 GB) | 8 GB total RAM | `512x512` | Full | `192x192` | 1 | `192x192` | `uint16` | **18.00 GiB** | **Blocked** | Resident payload exceeds total machine RAM | — | — | | [**WebGPU**](platforms/webgpu.md) | MacBook Air (M2, 8 GB) | 8 GB total RAM | `512x512` | Full | `192x192` | 2 | `96x96` | `uint16` | **4.50 GiB** | **Blocked** | Production detector binning stores `float32`; exact-integer accumulation and residency must be integrated before the physical pressure/parity gate | — | — | | [**WebGPU**](platforms/webgpu.md) | MacBook Air (M2, 8 GB) | 8 GB total RAM | `512x512` | Full | `192x192` | 4 | `48x48` | `uint16` | **1.125 GiB** | **Blocked** | Production detector binning stores `float32`; exact-integer accumulation and residency must be integrated before the physical pressure/parity gate | — | — | | [**WebGPU**](platforms/webgpu.md) | MacBook Air (M2, 8 GB) | 8 GB total RAM | `512x512` | Full | `192x192` | 8 | `24x24` | `uint16` | **0.28125 GiB** | **Blocked** | Production detector binning stores `float32`; exact-integer accumulation and residency must be integrated before the physical pressure/parity gate | — | — | Python `io.load` on CUDA and MPS has no detector bin: it keeps the complete native detector ANS encoded, so its payload depends on the counts. Three real `512x512x192x192` `uint16` acquisitions encode to 2,127,983,353, 2,279,376,159 and 3,374,307,987 bytes (0.11 to 0.18 of the 18.00 GiB of counts) on the CUDA loader (NVIDIA RTX PRO 6000 Blackwell, 2026-10-05). The exact integer WebGPU bin-2/4/8 payloads can fit within the nominal 8 GB total-RAM floor, but payload arithmetic alone is not acceptance. Production WebGPU still needs integer accumulation/residency plus a physical browser run that captures browser, adapter, staging, operating-system, pressure, swap, and scientific parity. Likewise, the current Blackwell timings do not prove a physical 6 GiB CUDA floor. ### What a 4 or 6 GiB budget can hold This capacity chart fixes the full scan at `512x512` and the native detector at `192x192`. “Payload” excludes decoder scratch, staging buffers, allocator reserve, and other GPU users unless the row reports a measured process peak. | Detector bin | Resident dtype | Resident payload | 4 GiB fit | 6 GiB fit | Evidence state | |---:|---|---:|---|---|---| | 1 | `uint16` | **18.00 GiB** | No | No | Calculated dense payload (native Swift/Metal and WebGPU bin 1) | | 1 | `uint16`, ANS encoded | **1.98 to 3.14 GiB** | Candidate | Candidate | Measured CUDA `io.load` payload of three real acquisitions; physical 4/6 GiB signoff Pending | | 1 | `uint8` | **9.00 GiB** | No | No | Complete-audit lossless path only (WebGPU low8) | | 4 | `uint16` | **1.125 GiB** | Candidate | Candidate | Physical 8 GB M2 Air evidence (native Swift/Metal); 4/6 GiB signoff Pending | The encoded CUDA/MPS load needs no memory plan of its own: `io.load` reads and encodes the acquisition in bounded scan blocks, and `read(...)` and the detector kernels decode bounded blocks, so the working set is the encoded payload plus bounded decode scratch. Each `512x512 float32` product map is only **1 MiB**, and one `192x192` float32 mean diffraction pattern is **144 KiB**. The source working set—not the final BF/ADF/DF/DPC image—is the capacity problem. For a full scan with output detector bin $b$ and $w$ resident bytes per value, the payload alone is $$ B_{\mathrm{payload}} =N_{R_r}N_{R_c} \left\lceil\frac{N_{k_r}}{b}\right\rceil \left\lceil\frac{N_{k_c}}{b}\right\rceil w. $$ Peak memory is larger: the benchmark must also report live compressed bytes, decode and reduction scratch, staging/upload buffers, allocator reserve, products, concurrent GPU users, and—on unified memory—process pressure and swap. A calculated payload is never relabeled as a measured peak. ```{admonition} Small-GPU support today :class: important The CUDA/MPS `screening.prepare` path loads the complete acquisition once into encoded storage and emits mean DP, total, BF, ABF, ADF, DF, CoM, rotation, and iDPC without cropping the scan. Physical 4 and 6 GiB product-pipeline signoff is **Pending**. See the {ref}`screening API ` before choosing a plan. ``` ## Platform-first module dashboard Every table keeps the scientific module as its section and puts the execution platform in the first column. Empty cells are forbidden: - **✓** — verified with retained real-data or physical-device parity evidence. - **Test** — deterministic source/test coverage without an equivalent retained physical or native-data run. - **Pending** — implementation exists, but that exact size, bin, or timing evidence has not been retained yet. - **Ref** — CPU correctness adjudication, never a production fallback. - **—** — unsupported or not a target. ### I/O and first usable product — `quantem.gpu.io` #### Exact configuration gaps The measured table above owns retained full-scan timings. The [filterable coverage registry](performance/coverage.md) owns every missing combination as one **Platform → Computer → detector-bin → cache-state** row; this page does not maintain a second bundled list. Real-space crop remains an explicit selector and is never an implicit performance policy. #### Selective scan rectangles The physical WebGPU rows below are prepared **whole-shard-selective** rectangle loads. A retained frame-span manifest lets the loader omit nonintersecting shards and decode/upload only selected scan rows. It does not issue byte-range reads inside an intersecting shard. | Platform | Computer | Selected scan | Rectangle `(row_start,row_stop,column_start,column_stop)` | Source detector | Source dtype | Shards read | Storage bytes read | Samples | Loader p50 | Loader p95 | Loader maximum | Logical resident | Browser-tree RSS peak | Observed swap delta | Parity | Device tested | Date tested | |---|---|---:|---|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|---|---| | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | `64x64` | `(0,64,0,64)` | `192x192` | `uint16` | 4 of 27 | 488,224,242 B | 5 | **0.147 s** | **0.1544 s** | **0.156 s** | 301,989,888 B | 1,724,317,696 B | 0 B | ✓ | Chrome 151, Apple M5 Max Metal | 2026-08-22 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | `256x256` | `(128,384,128,384)` | `192x192` | `uint16` | 14 of 27 | 1,705,556,941 B | 5 | **0.381 s** | **0.3924 s** | **0.394 s** | 4,831,838,208 B | 3,002,875,904 B | 0 B | ✓ | Chrome 151, Apple M5 Max Metal | 2026-08-22 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | `384x384` | `(64,448,64,448)` | `192x192` | `uint16` | 20 of 27 | 2,432,636,897 B | 5 | **0.574 s** | **0.582 s** | **0.584 s** | 10,871,635,968 B | 3,896,934,400 B | 0 B | ✓ | Chrome 151, Apple M5 Max Metal | 2026-08-22 | All rows use prepared block indexes and a prepared frame-span manifest. The operating-system source-page state was uncontrolled/unspecified and no eviction was performed. Frame-manifest read/encoding, DevTools injection, checksum harness, products, and application E2E are excluded. A negative control without the frame-span manifest read all 27 shards (3.17 GB), so it is explicitly **not** selective evidence. CUDA and Python MPS load the complete acquisition and return a rectangle with `read(scan_region=..., detector_region=...)` from the encoded resident; native Swift/Metal bulk selection and physical cross-backend timing remain pending. The parity check mark means all five runs for that rectangle passed independent CPU `uint16` checksum probes at the first, middle, and last retained raw frames, plus output shape, dtype, row-major order, and selection metadata. Full-tensor readback parity was not performed. The WebGPU fixture view records source identity `1be810b9...`; fixture C records `c9c0d968...`, and the CUDA master record is `4802ec16...`. These rows are not a cross-lane fixture-controlled comparison. ### Screening and prepared-product caches — `quantem.gpu.screening` | Platform | Computer | Support | Operation | Source plan | Statistic | Time | Device tested | Date tested | |---|---|---|---|---|---|---:|---|---| | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | ✓ | Exact streamed screening | Full `512x512x192x192` `uint16`; no crop/bin; warm source pages unspecified; empty result cache | p50 of 6 | **1.205 s** | NVIDIA RTX PRO 6000 Blackwell Workstation Edition, GPU 0 | 2026-08-22 | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | ✓ | Exact screening build | Full `512x512x192x192` `uint16`; no crop/bin; exact fallback pass | Single run | **6.711 s** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | ✓ | Validated screening-v3 reopen | Prepared derived products | p50 | **20.803 ms** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **Native Swift/Metal** | MacBook Air (M2, 8 GB) | ✓ | Validated exact-summary reopen | Prepared derived products | p50 | **0.029 s** | Apple M2 MacBook Air (`Mac14,2`, 8 GB) | 2026-08-19 | | **WebGPU** | N/A (not implemented) | — | Prepared screening workflow | — | — | — | — | — | | **CPU reference** | Portable CI runner | Ref | Independent adjudication | Reference fixtures | — | **Pending** | — | — | The promoted MPS build derives its detector mean and masks from the full scan. When the provisional and final masks differ, BF/DF are recomputed exactly. The first-chunk-mask candidate failed parity and is retained only as a rejected experiment; its timing is not repeated on a current page. Saved-product reopen is likewise never presented as source load. The current CUDA pinned-slot candidate reduced the like-for-like package p50 from `1.356516 s` to `1.204713 s`; candidate p95/max were `1.325760/1.329731 s`. All six public arrays were byte exact in every trial. This is warm/source-pages-unspecified exact screening, not cold HDF5 and not a prepared-result reopen. Detailed pinned-registration and memory measurements remain in the provenance ledger. ### Virtual images — `quantem.gpu.detector` #### Complete platform, computer, and detector-bin matrix The registry keeps one row per exact product configuration instead of hiding an unrun combination. Use the **Platform**, **Computer**, **State**, and **Detector bin** filters to isolate the implementation or physical computer you care about. A pending, blocked, or unsupported row is intentionally visible; only a measured row publishes a timing distribution. The 8 GB MacBook Air detector-bin-2 row is the locked downstream acceptance configuration. Detector bins 4 and 8 remain separate library-capability cells, not silent application fallbacks. Detector bin 1 is explicitly blocked where its 18 GiB exact resident input cannot be admitted. ```{include} _generated/benchmark_coverage.md :start-after: :end-before: ``` #### Retained operation-level timings | Platform | Computer | Operation | Scan grid | Detector | Detector bin | Input state | Statistic | Time | Device tested | Date tested | |---|---|---|---:|---:|---:|---|---|---:|---|---| | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | Mean diffraction | `512x512` | `192x192` | 1 | Warm resident | p50 | **18.392 ms** | NVIDIA RTX PRO 6000 Blackwell Max-Q, GPU 1 | 2026-08-19 | | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | BF exact sum | `512x512` | `192x192` | 1 | Warm resident | p50 | **3.768 ms** | NVIDIA RTX PRO 6000 Blackwell Max-Q, GPU 1 | 2026-08-19 | | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | ADF exact sum | `512x512` | `192x192` | 1 | Warm resident | p50 | **5.586 ms** | NVIDIA RTX PRO 6000 Blackwell Max-Q, GPU 1 | 2026-08-19 | | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | DF exact sum | `512x512` | `192x192` | 1 | Warm resident | p50 | **3.747 ms** | NVIDIA RTX PRO 6000 Blackwell Max-Q, GPU 1 | 2026-08-19 | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | Mean diffraction | `512x512` | `48x48` | 4 | Warm resident | p50 | **74.805 ms** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | BF exact sum | `512x512` | `48x48` | 4 | Warm resident | p50 | **2.502 ms** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | ADF exact sum | `512x512` | `48x48` | 4 | Warm resident | p50 | **4.404 ms** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | DF exact sum | `512x512` | `48x48` | 4 | Warm resident | p50 | **2.642 ms** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **Native Swift/Metal** | MacBook Pro (M5 Max, 128 GB) | Fused BF, ABF, ADF, total, and row/column moments | `512x512` | `192x192` | 1 | Controlled source load | p50 | **119.040 ms** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-22 | | **Native Swift/Metal** | MacBook Air (M2, 8 GB) | BF, ABF, ADF, total, and row/column moments | `512x512` | `48x48` | 4 | Prepared resident-cache fallback | Single run | **103.0 ms** | Apple M2 MacBook Air (`Mac14,2`, 8 GB) | 2026-08-19 | | **Native Swift/Metal** | MacBook Air (M2, 8 GB) | Exact moment widening for resident summary | `512x512` | `48x48` | 4 | Same validated fused source pass; no resident traversal | Single run | **0.569 ms** | Apple M2 MacBook Air (`Mac14,2`, 8 GB) | 2026-08-20 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | Mean diffraction | `512x512` | `192x192` | 1 | Warm resident | p50 | **50.9 ms** | Chrome 151, Apple M5 Max Metal-3 | 2026-08-19 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | BF exact sum | `512x512` | `192x192` | 1 | Warm resident | p50 | **5.5 ms** | Chrome 151, Apple M5 Max Metal-3 | 2026-08-19 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | ADF exact sum | `512x512` | `192x192` | 1 | Warm resident | p50 | **15.0 ms** | Chrome 151, Apple M5 Max Metal-3 | 2026-08-19 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | DF exact sum | `512x512` | `192x192` | 1 | Warm resident | p50 | **43.4 ms** | Chrome 151, Apple M5 Max Metal-3 | 2026-08-19 | | **CPU reference** | MacBook Pro (M5 Max, 128 GB) | Virtual-image adjudication | `512x512` | `192x192` | 1 | One independent traversal | Single run | **31.08 s** | Apple M5 Max CPU | 2026-08-19 | The current integer and mean-DP rows pass their independent CPU reference. CUDA uses native detector resolution on fixture D; MPS uses explicit detector bin 4 on fixture C. WebGPU uses native detector resolution on D. The current native bin-1 source pass produces three virtual images, detector sums, and exact `uint32` total and detector moments in one fused stage. That stage overlaps source IO and is not additive with the package-wall measurement. Revision `70bc366` proves when bin-4 accumulators fit, then widens the three moment maps to `uint64` in one small dispatch. The 0.569 ms row is only that incremental widening; it is not the full source pass. The cache-only fallback row remains because it has a different input state. Neither historical bin-4 row is a compressed-source load time. ### Detector moments and phase contrast — `quantem.gpu.dpc` | Platform | Computer | Operation | Scan grid | Detector | Detector bin | Input state | Statistic | Time | Device tested | Date tested | |---|---|---|---:|---:|---:|---|---|---:|---|---| | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | CoM row and column | `512x512` | `192x192` | 1 | Warm resident | p50 | **13.002 ms** | NVIDIA RTX PRO 6000 Blackwell Max-Q, GPU 1 | 2026-08-19 | | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | Fixed-orientation iDPC | `512x512` | `192x192` | 1 | CPU small-field integration | p50 | **21.272 ms** | NVIDIA RTX PRO 6000 Blackwell Max-Q, GPU 1 | 2026-08-19 | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | CoM row and column | `512x512` | `48x48` | 4 | Warm resident | p50 | **4.637 ms** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | Fixed-orientation iDPC | `512x512` | `48x48` | 4 | CPU small-field integration | p50 | **12.678 ms** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **Native Swift/Metal** | MacBook Air (M2, 8 GB) | CoM, DPC, iDPC, and display statistics | `512x512` | `48x48` | 4 | Prepared exact `uint64` moments | Single run | **11.389 ms** | Apple M2 MacBook Air (`Mac14,2`, 8 GB) | 2026-08-19 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | DPC row | `512x512` | `192x192` | 1 | Warm cached CoM | p50 | **0.9 ms** | Chrome 151, Apple M5 Max Metal-3 | 2026-08-19 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | DPC column | `512x512` | `192x192` | 1 | Warm cached CoM | p50 | **0.7 ms** | Chrome 151, Apple M5 Max Metal-3 | 2026-08-19 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | iDPC | `512x512` | `192x192` | 1 | Explicit 0-degree rotation | p50 | **1.4 ms** | Chrome 151, Apple M5 Max Metal-3 | 2026-08-19 | | **CPU reference** | MacBook Pro (M5 Max, 128 GB) | Rotation and iDPC adjudication | `512x512` | `192x192` | 1 | CPU reference | Single run | **177.6 ms** | Apple M5 Max CPU | 2026-08-19 | The current CUDA and MPS CoM rows pass their respective independent references. The native row is the complete derived-product stage after exact moments exist; it includes center/mean, alignment, Metal iDPC, and float-surface construction. It does not include resident-cache traversal or source loading. The CUDA/MPS cross-fixture timing rows are not compared numerically. The prior same-fixture detector-bin-4 comparison remains a historical block for iDPC at `2.84e-5` maximum error. On the retained M5 Max hardware run, WebGPU DPC row and column were byte exact. Optimized and zero-rotation iDPC had zero frozen tolerance violations; optimized rotation had maximum absolute error `1.52587890625e-5` and maximum tolerance ratio `0.8993483035`. ### Single-sideband ptychography — `quantem.gpu.SSB` These are square scan-grid sizes, not detector dimensions. Current `512x512` operation timing is separated from the size-support matrix. Source loading, `G(\mathbf k,\boldsymbol{\nu})` preparation, and UI paint are excluded. | Platform | Computer | State | Operation | Detector plan | BF policy | Boundary | Statistic | Time | Device tested | Date tested | |---|---|---|---|---|---|---|---|---:|---|---| | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | Partial | Complex object | Native `192x192` | 8,928 active | Warm resident GPU | p50 | **13.883 ms** | NVIDIA RTX PRO 6000 Blackwell Max-Q, GPU 1 | 2026-08-19 | | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | Partial | Exact phase | Native `192x192` | 8,928 active | Warm resident GPU | p50 | **32.035 ms** | NVIDIA RTX PRO 6000 Blackwell Max-Q, GPU 1 | 2026-08-19 | | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | Partial | Exact phase and loss | Native `192x192` | 8,928 active | Warm resident GPU | p50 | **32.335 ms** | NVIDIA RTX PRO 6000 Blackwell Max-Q, GPU 1 | 2026-08-19 | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | Partial | Exact phase and loss | Explicit detector bin 2 to `96x96` | 2,275 calibrated | Single synchronized reconstruction | Single run | **497.187 ms** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **Native Swift/Metal** | MacBook Pro (M5 Max, 128 GB) | Partial | Complex object | Native-detector exact BF columns | 9,074 logical / 2,459 executed | Warm complete Hermitian cache | p50 | **8.911 ms** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **Native Swift/Metal** | MacBook Pro (M5 Max, 128 GB) | Partial | Exact phase-variance loss | Native-detector exact BF columns | 9,074 logical / 2,459 executed | Warm complete Hermitian cache | p50 | **25.120 ms** | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | Refuted diagnostic | Complex object | Native `192x192` companion | 3,418 active | Readback-complete compute wall | p50 | **32.5 ms** | Chrome 151, Apple M5 Max Metal-3 | 2026-08-19 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | Refuted diagnostic | Exact phase | Native `192x192` companion | 3,418 active | Readback-complete compute wall | p50 | **102.1 ms** | Chrome 151, Apple M5 Max Metal-3 | 2026-08-19 | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | Refuted diagnostic | Exact phase and loss | Native `192x192` companion | 3,418 active | Readback-complete compute wall | p50 | **189.4 ms** | Chrome 151, Apple M5 Max Metal-3 | 2026-08-19 | | **CPU reference** | Portable CI runner | Reference | SSB | Frozen adjudication only | Frozen fixture | Reference | — | **Pending** | — | — | The WebGPU SSB values are retained diagnostic timings, not accepted scientific performance. All 262,144 phase values differed from the frozen reference and the wrapped maximum error was `0.0597773 rad` against a `0.0002 rad` gate. That implementation remains refuted until phase parity is restored. The raw detector-bin-2 MPS phase agrees with an independent CUDA reference to `1.2815e-6` wrapped radians maximum; loss differs by `7.45e-9`. A prepared BF-column companion from the same campaign is rejected because its stored columns do not match the declared detector-bin coordinate grid. Its faster timings are not published as scientific results. Calibration is a separate operation: | Platform | Computer | Search | Refinement | Repetitions | Statistic | Time | Result | Device tested | Date tested | |---|---|---|---|---:|---|---:|---|---|---| | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | Seeded Optuna TPE, 200 trials | Nelder–Mead | 3 | p50 | **11.168 s** | Byte-deterministic parameters, phase, object, and loss | NVIDIA RTX PRO 6000 Blackwell Max-Q, GPU 1 | 2026-08-19 | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | Optuna TPE, 200 trials | Nelder–Mead | — | — | **Pending** | Current compatible source not profiled | — | — | | **Native Swift/Metal** | MacBook Pro (M5 Max, 128 GB) | Seeded TPE, 200 trials | Nelder–Mead | 3 | p50 | **6.061 s** | Deterministic parameters and loss | Apple M5 Max (`Mac17,6`, 40-core GPU) | 2026-08-19 | | **WebGPU** | N/A (not implemented) | — | — | — | — | — | Unsupported | — | — | | **CPU reference** | Portable CI runner | — | — | — | — | — | Reference only | — | — | Levenberg–Marquardt is not implemented in any current SSB backend. An earlier CUDA atomic-objective calibration split into two different fitted minima under an identical seed; it is retained as a rejected experiment, not a benchmark row. | Platform | Computer | Scan grid | Source kind | BF policy | State | |---|---|---:|---|---|---| | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | `128x128` | Fixed-size parity | Frozen fixture | Test | | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | `256x256` | Fixed-size parity | Frozen fixture | Test | | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | `512x512` | Native real acquisition | Full active BF | Partial physical evidence | | **CUDA** | Linux CUDA workstation (dual 96 GB Blackwell GPUs) | `1024x1024` | Fixed-size parity | Frozen fixture | Test | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | `128x128` | Resized/synthetic | Fixed-size fixture | Test | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | `256x256` | Resized/synthetic | Fixed-size fixture | Test | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | `512x512` | Native real acquisition | Full active BF | Partial physical evidence | | **Python MPS** | MacBook Pro (M5 Max, 128 GB) | `1024x1024` | Synthetic | Fixed-size fixture | Test | | **Native Swift/Metal** | MacBook Pro (M5 Max, 128 GB) | `128x128` | — | — | Not supported | | **Native Swift/Metal** | MacBook Pro (M5 Max, 128 GB) | `256x256` | — | — | Not supported | | **Native Swift/Metal** | MacBook Pro (M5 Max, 128 GB) | `512x512` | Native real acquisition | Full active BF | Partial physical evidence | | **Native Swift/Metal** | MacBook Pro (M5 Max, 128 GB) | `1024x1024` | — | — | Not supported | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | `128x128` | Real BF30 parity | Radius 30 px | Partial | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | `256x256` | Deterministic fixture | Test BF | Test | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | `512x512` | Real interaction | Frozen phase reference | Refuted | | **WebGPU** | MacBook Pro (M5 Max, 128 GB) | `1024x1024` | Real interaction | Incomplete frozen reference | Partial | | **CPU reference** | Portable CI runner | `128x128` | — | — | Not retained | | **CPU reference** | Portable CI runner | `256x256` | — | — | Not retained | | **CPU reference** | Portable CI runner | `512x512` | Independent adjudication | Frozen fixture | Ref | | **CPU reference** | Portable CI runner | `1024x1024` | — | — | Not retained | Native Swift/Metal has a package-owned 512×512 implementation. Other native Swift scan sizes remain unsupported rather than inferred from CUDA/MPS. Untimed CUDA and MPS sizes retain fixed-size parity coverage. The WebGPU 512×512 phase result is explicitly refuted; its timing cannot be promoted until parity is restored. See the [SSB performance history](maintainer/ssb-performance.md) for size-specific historical experiments. Current numerical rows remain in this dashboard and the verified-results ledger. ### Cross-module platform map A one-row-per-platform map hides the computer, configuration, and evidence state, and previously made a refuted WebGPU SSB result look accepted. It is no longer maintained. Use the [filterable atomic coverage matrix](performance/coverage.md), where every row begins with **Platform** and **Computer** and keeps module, bin, dtype, cache state, and next gate separate. A warm resident kernel, prepared source, first-process application load, and saved-result reopen answer different questions. See [Benchmark methodology](performance/methodology.md) and [Verified benchmark results](performance/results.md) before comparing them. The public scientific array is always $$ I[R_r,R_c,k_r,k_c], $$ where $\mathbf R=(R_r,R_c)$ is the probe/scan coordinate and $\mathbf k=(k_r,k_c)$ is the detector coordinate. Runtime-specific layout, tiling, fusion, and dispatch are private optimizations; shape, sampling, precision, calibration, and provenance are shared contracts. ## Where an implementer starts | Goal | Read first | Then inspect | Acceptance evidence | |---|---|---|---| | Change scientific meaning or add an operation | [Scientific contract](concepts/scientific-contract.md) | [Scientific kernels](kernels/index.md) | Operation-specific equation, provenance schema, and independent reference | | Optimize CUDA | [CUDA implementation](platforms/cuda.md) | The domain's `cuda/` package, for example `resident/cuda` | Real NVIDIA profile plus exact/frozen parity | | Optimize Python on Apple Silicon | [Python MPS](platforms/mps.md) | The domain's `mps/` package, for example `resident/mps` | Physical Apple device profile plus exact/frozen parity | | Build a native Apple client/library | [Native Swift and Metal](platforms/swift-metal.md) | `Package.swift` and `native/swift/{Sources,Tests}` | `swift test`, physical Metal timing, and cross-language fixtures | | Optimize a browser client | [WebGPU](platforms/webgpu.md) | Domain WebGPU TypeScript/WGSL resources | Real adapter, headed browser gate, and matching scientific output | | Deploy CUDA behind a process boundary | [QuantEM.GPU Remote](remote/index.md) | Deployment, protocol, and admission pages | Same array/provenance contract plus transport and capacity checks | | Add or review a benchmark | [Benchmark methodology](performance/methodology.md) | [Continuous profiling](performance/continuous-profiling.md), [parity](performance/parity.md), and the [optimization ledger](maintainer/backend-optimization-matrix.md) | Date, revision, device, source plan, cache state, memory, wall boundary, and parity artifact | ## Dashboard maintenance rule Update a dashboard row only after its detailed evidence row is complete. The detail remains authoritative and must record measurement date, exact source revision, physical device/runtime, source shape and dtype, cache state, crop/bin/load plan, benchmark definition, peak memory or swap where available, and numerical or hash parity. Keep an older result when the newer experiment changes any of those conditions; label both instead of silently replacing one. Keep every row atomic: a different bin, dtype path, fixture, cache state, or statistic is another row. Documentation tests enforce that the landing page stays timing-free and that current overview values remain owned by this dashboard. Accepted and rejected experiments remain in the [optimization ledger](maintainer/backend-optimization-matrix.md), and the machine-readable evidence fingerprints are in [`performance/evidence_manifest.json`](performance/evidence_manifest.json). The platform/module cadence, runner ownership, and known harness gaps are machine-checked from `benchmarks/profile_matrix.json`; see [Continuous profiling](performance/continuous-profiling.md).