GPU Specifications
Hardware specs for all supported GPU types -- VRAM, memory type, bandwidth, interconnect, and compute performance
Quick reference for all GPU types supported by urun.
Effective VRAM available to your model is less than total VRAM — CUDA kernels, framework overhead, and memory fragmentation typically consume 1-3 GB depending on workload.
Hardware specifications
| GPU | Spec string | VRAM | Memory Type | Memory Bandwidth | Interconnect |
|---|---|---|---|---|---|
| B200 SXM | "b200:N" | 192 GB | HBM3e | 8.0 TB/s | NVLink 5 (1.8 TB/s) |
| H200 SXM | "h200:N" | 141 GB | HBM3e | 4.8 TB/s | NVLink 4 (900 GB/s) |
| RTX PRO 6000 Blackwell SE | "rtx6000:N" | 96 GB | GDDR7 | 1.6 TB/s | PCIe 5.0 (128 GB/s) |
| H100 SXM | "h100:N" | 80 GB | HBM3 | 3.35 TB/s | NVLink 4 (900 GB/s) |
| A100 SXM | "a100:N" | 80 GB | HBM2e | 2.0 TB/s | NVLink 3 (600 GB/s) |
| L40S | "l40s:N" | 48 GB | GDDR6 | 864 GB/s | PCIe 4.0 (64 GB/s) |
| A10G | "a10g:N" | 24 GB | GDDR6 | 600 GB/s | PCIe 4.0 (64 GB/s) |
| L4 | "l4:N" | 24 GB | GDDR6 | 300 GB/s | PCIe 4.0 (64 GB/s) |
The RTX PRO 6000 Blackwell Server Edition's 96 GB of VRAM fits ~80GB-class models on a single card that would otherwise need multi-GPU sharding on 48 GB parts.
Compute performance
Dense (non-sparse) tensor-core throughput. The B200's exact dense FP16/FP8/FP4 figures depend on the final shipping clocks — confirm against NVIDIA's current Blackwell datasheet before sizing on them.
| GPU | FP16 TFLOPS | FP8 TFLOPS | FP4 TFLOPS | TDP |
|---|---|---|---|---|
| B200 SXM | see datasheet | see datasheet | see datasheet (Blackwell-only) | ~1000W |
| RTX PRO 6000 Blackwell SE | see datasheet | see datasheet | see datasheet (Blackwell) | 600W |
| H100 SXM | 989 | 1,979 | N/A | 700W |
| H200 SXM | 989 | 1,979 | N/A | 700W |
| A100 SXM | 312 | N/A | N/A | 400W |
| L40S | 362 | 733 | N/A | 350W |
H200 shares the H100's GH100 compute die, so its dense FP16/FP8 throughput matches the H100; the H200's advantage is memory (141 GB HBM3e at 4.8 TB/s vs 80 GB at 3.35 TB/s), not raw FLOPS. FP4 is a Blackwell-class (B200) capability and is not available on Hopper/Ampere/Ada parts.
Specifying a GPU
The canonical way to request hardware is the gpus="type:count" string on @app.function:
@app.function(gpus="h100:8") # 8x H100
@app.function(gpus="l40s:1") # 1x L40S
@app.function(gpus="rtx6000:1") # 1x RTX PRO 6000
@app.function(gpus="a100:4") # 4x A100
@app.function(gpus="h200") # 1x H200 (count defaults to 1)When you pass an integer like gpus=8 without a type, urun defaults to H100.
For guidance on choosing GPUs for your workload, see Compute.