Support Matrix
Items verified in hart 0.0.6. Anything not listed here has not been verified.
Legend — ✅ Supported · 🟡 Limited · ❌ Not supported · 📅 Planned
The full stack combinations, including OS, kernel, driver, and firmware, are in the
SDK support matrix. This page covers what the runtime itself
supports.
| Item |
Status |
Notes |
| Linux, x86_64 |
✅ |
|
| Python 3.10 – 3.12 |
✅ |
|
| PyTorch 2.10.0 |
✅ |
Optional. Required only for framework integration |
| Windows, macOS |
❌ |
|
Device Versions
| Device version |
Status |
Notes |
BERTHA_CLOUD_EVT0 |
✅ |
DRAM word size 64 B |
BERTHA_CLOUD_EVT1 |
✅ |
DRAM word size 64 B |
BERTHA_EDGE_EVT0 |
✅ |
DRAM word size 32 B |
Memory Features
| Feature |
Status |
Notes |
| Contiguous allocation |
✅ |
|
| DRAM channel interleaved allocation |
✅ |
|
| 3D shape allocation |
✅ |
|
| Asynchronous allocation and free |
✅ |
|
| Host↔device copy |
✅ |
|
| D2D within the same device |
✅ |
|
| D2D across devices |
🟡 |
Depends on the device pair. Returns HaMemCpyNotSupported when unsupported |
Host↔host copy in HaMemCpyNd |
✅ |
Performed entirely on the host |
File→device load (HaMemLoadFileAsync) |
✅ |
|
Non-contiguous strides in HaBuildDeviceTensorLayout |
❌ |
Returns HaNotSupported. The parameter exists only for future compatibility |
| Returning freed memory |
✅ |
A segment emptied by a free is returned right away or kept in the pool depending on conditions. Kept segments are returned with device_mem_trim() (C++: HaDeviceMemTrim()) |
Tensor Layout Kinds
HaTensorKind |
Status |
Required rank |
DramWordInterleaved |
✅ |
>= 1 |
SinglePassInterleaved |
✅ |
>= 1 |
SingleChannel |
✅ |
>= 1 |
RowInterleaved |
✅ |
>= 2 |
MaxRowsInterleaved |
✅ |
>= 2 |
MPUTensor |
✅ |
>= 2 |
KvCacheKey / KvCacheValue |
✅ |
Exactly 5 |
Alignment constraints per kind are in the Memory Model.
Virtual Memory (VMM)
Support scope of the VMM API. The table below reflects the current hardware revision (engineering sample).
| Feature |
Status |
Notes |
KV pool allocation and free (HaVmmAllocatePool, HaVmmFreePool) |
✅ |
The pool is a single segment |
HW page initialization (HaVmmInitBlocks) |
✅ |
pool_base and pool_bytes required |
Per-request page mapping and return (HaVmmMapBlocksBatchFromPool, HaVmmMapBlocksBatchFromPools, HaVmmUnmapBlocksBatchFromPool) |
✅ |
Based on token count, block_size 1 |
Python bindings (hart.hart.ha_vmm_*) |
📅 |
Use the PyTorch vmm_* operators |
Execution and Synchronization
| Feature |
Status |
Notes |
| Synchronous kernel launch |
✅ |
|
| Asynchronous kernel launch |
✅ |
|
| Legato kernel launch |
✅ |
Both synchronous and asynchronous |
| Streams |
✅ |
10 per device |
| Events, including cross-stream waits |
✅ |
|
Observability
Profile runtime execution with the HyperAccel Profiler (hapti). For the collection scope, see the
profiler support matrix.
| Feature |
Status |
Notes |
| Device memory usage query |
✅ |
device_mem_get_stats() (C++: HaDeviceMemGetStats()) |
| Device utilization query |
❌ |
|
Limits
| Resource |
Limit |
| Devices per process |
8 |
| Streams per device |
10 (excluding the default stream, including streams created through the driver API) |
| Executor worker threads per device |
8, not configurable |
| Small/large allocation pool boundary |
1 MB |
Known Limitations
- Empty segments are kept in the pool even after every allocation is freed. While other allocations remain, an
empty segment may be returned right away, but when no allocation remains it is kept for later
allocations. Kept segments are returned only when
device_mem_trim() (C++: HaDeviceMemTrim()) is called or the
device is reset.
- A device reset invalidates every handle on that device. This covers allocations, streams,
and events. Stale stream and event handles and device pointers are rejected with an error, but if a new
allocation after the reset gets the same address, the stale pointer silently refers to the new allocation, so clean up
related resources before the reset.
- The runtime cannot report device utilization. Memory usage can be queried with
device_mem_get_stats() (C++: HaDeviceMemGetStats()).
- The executor thread count and the small/large allocation pool boundary (1 MB) cannot be tuned. The Huge tier
threshold and the Small pool split remainder can be set through environment variables.
See also