Hardware / NVIDIA Grace Blackwell · GB10 (product GX10)

DGX Spark · NVIDIA GB10

A Blackwell GPU sharing 128 GB of memory with a 20-core Arm CPU. Coppelius trains here through the compiler's sm_121 back end.

AArch64 host · Blackwell compute capability 12.1

What has run here

Qualified

Coppelius trains, resumes and predicts here, including runs of 10,000 steps (25–26 Sep 2026). Its sm_86 kernels are lowered to sm_121 by the compiler; nothing is tuned for this GPU yet. Vathek's short training runs also pass here.

Driver facts

Typed profile

Read at build time from Platform.DGXSpark.Profile at 2877a0c1. Systems bind to these values by name; none branches on a card name.

FactValueAlpha definition
Profile identitydgx-spark-bringup-v1dgxSparkProfileIdentity
Performance cores10 × Cortex-X925dgxSparkPerformanceCoreCount
Efficiency cores10 × Cortex-A725dgxSparkEfficiencyCoreCount
Unified memory130,596,048,896 B (121.6 GiB)dgxSparkTotalMemory
Page size4096 BdgxSparkPageSize
Level 2 cache26,214,400 B (25 MiB)dgxSparkLevel2Cache
Level 3 cache25,165,824 B (24 MiB)dgxSparkLevel3Cache
Accelerator compute capability12.1dgxSparkDeviceComputeMajor / Minor

Profile coverage

What has been measured, and what has not

A fully profiled device has three tiers. A tier that is not measured yet lists the fields it will fill, so the gap is explicit.

  1. 1

    Driver facts

    Live

    What the host reports: core parts and counts, unified memory, page size, caches, and the GPU's compute capability.

    • ARM implementer and core parts
    • core counts
    • unified memory
    • page size
    • L2 / L3 cache
    • GB10 compute capability
  2. 2

    Micro-benchmarks

    First reading

    Measured instruction and memory behaviour. Copy bandwidth by memory kind has been measured here (below); instruction-level costs have not.

    • issue rate and latency per typed instruction (FFMA, IMAD, MUFU, LDG, STG, LDSM, HMMA…)
    • shared-memory bandwidth and bank behaviour
    • global-memory bandwidth by access pattern and width
    • tensor-core throughput at 16×8×16
    • barrier and warp-shuffle cost
    • submission, doorbell and semaphore round-trip cost
    • sustained clocks under the training workload
  3. 3

    Model-shaped calibration

    First reading

    The cost of the exact executables Alpha emits for each system on this device. End-to-end throughput and per-launch device times from the GPU's own timestamps exist (below); occupancy and register pressure have not been measured.

    • per-kernel time for every launch in each system's schedule
    • occupancy and register / shared-memory pressure per kernel
    • end-to-end step time and its split between host protocol, launches and waits
    • memory footprint against the profile's video memory

Measured on this card · 2026-09-26

Measured readings

From the run record. These measure Alpha's current executables on this card, not the card's limits.

ReadingValue
One Coppelius training step (866 launches, run 21)45.35 ms on the device's own clock
Copy bandwidth, L2-resident, GPU-cached memory737–774 GB/s
Copy bandwidth, DRAM-sized225–235 GB/s
10,000-step runsPer-step losses and final checkpoints byte-identical across runs 19–21

Sources

Pairings and evidence