Performance
Determinism comes first. This page gives one scoped measurement, where PyTorch leads, and how the two compare.
On NVIDIA B200, a 4096 x 4096 matrix-multiply end-to-end round trip took 30.5 ms in Talario's deterministic mode against 43.5 ms for PyTorch with pinned memory (same box, same day; indicative; method).
On NVIDIA B300, a 4096 x 4096 matrix-multiply end-to-end round trip took 30.1 ms in Talario's deterministic mode against 42.3 ms for PyTorch with pinned memory (same box, same day; indicative; method).
Default configuration on B200 and B300, measured on rented hardware. Times vary by rental; the ratio held at about 1.4x. Indicative, not a guarantee. Results bit-identical to the CPU.
Precision is f32, with no Tensor Cores on the bit-identical path.
Where PyTorch leads
- The multiply alone: cuBLAS leads by about 3x to 15x depending on the GPU. (method)
- Small calls: at 1024 x 1024 PyTorch leads by about 2x to 4x. (method)
- GH200: PyTorch leads by about 6x at 4096. (method)
How Talario compares
| Topic | PyTorch | Talario |
|---|---|---|
| Same bits on CPU and GPU, across hardware By default subnormal and NaN bit patterns can differ on NVIDIA; an opt-in GPU-matching mode makes them match. | Not guaranteed across hardware or releases (PyTorch documentation) | Yes: 26 GPU models, NVIDIA and AMD Instinct, float32, attested set |
| One engine on NVIDIA and AMD | Separate CUDA and ROCm builds | Yes |
| Runs inside a PHP worker | No: needs a Python service | Yes |
| 4096 round trip on B200 and B300, deterministic mode (method) Default configuration, measured on rented hardware. Times vary by rental; the ratio held at about 1.4x. Indicative. | 43.5 and 42.3 ms (pinned memory) | 30.5 and 30.1 ms |
| Small calls (1024 round trip) (method) | Ahead by about 2x to 4x | Behind |
| The multiply alone (method) | Ahead by about 3x to 15x, varies by GPU (cuBLAS) | Behind |
| GH200 (method) | Ahead by about 6x at 4096 | Behind |
| Tensor Cores Talario's bit-identical path is float32 and uses no Tensor Cores. | Yes, mature | Experimental, opt-in, not bit-identical, not yet faster than the exact path |
| Multiple GPUs | Yes | Single device today |
| Operator breadth | Very broad | A focused set: transformers, GNNs, RL and ONNX import |
| Availability | Public, open source | Not yet available |
Talario offers determinism across hardware plus in-process PHP serving. It does not beat vendor libraries at raw compute.
Method
One float32 matrix multiply end to end (host data in, multiply, result back on the host), Talario's deterministic path, same GPU, same day.
Honest caveats
- Determinism is the design priority: f32 by default.
- An experimental tensor-core matmul (matmulFast) is opt-in per call, not bit-identical, and not yet faster than the exact path.
- Bit-identity is claimed only where attested: exact ops and gradients at 1024x1024 on seven NVIDIA generations (Pascal through Blackwell); the transcendental tier (softmax, GELU, layernorm, rmsnorm, attention), forward and gradients, on twenty-six GPU models: twenty-two NVIDIA across five generations (Turing, Ampere, Ada, Hopper, Blackwell) and four AMD Instinct (MI300X, MI325X, MI350X, MI355X); the 100-step training replay CPU==GPU on nineteen of the NVIDIA models, on x86 and Arm (aarch64) hosts.
- AMD Instinct is reached through the HIP backend (no Vulkan driver there); the training replay was not run on it.
- By default, values in the subnormal range and the bit encoding of NaN/infinity results can differ between CPU and GPU (NVIDIA GPUs flush subnormals); an opt-in GPU-matching mode makes them match.
- Each NVIDIA driver is attested: a new driver can change results until the kernels are checked against it.
- AMD Radeon and Intel GPUs are attested for exact ops at small tensor sizes only; full scale on them is not yet attested.
- Software Vulkan (llvmpipe) is never counted as evidence.
- Op catalogue is focused (not PyTorch's thousands); no higher-order/double-backward.
- Multi-GPU training is validated on one device only.