#17 - Add RTX PRO 6000 benchmark YAMLs and baseline

[chloecrozier]

Partially addresses #17

RTX PRO 6000 benchmark YAMLs and results from our development system.

What this PR adds

• RTX PRO 6000 YAML configs for software loopback, dual-port NIC, and a generic template
• examples/rtx_pro_6000_baseline.md with measured results and hardware constraints

Results (no cable between ports)

Test	TX	RX
Software loopback	1580 Gbps	1580 Gbps
Real NIC, port 0 to port 1	382 Gbps	0
Reorder, software loopback	—	160 Gbps

Software loopback does not use the NIC. It confirms the GPUDirect build path but is not a wire-rate measurement. The NIC test shows transmit works at 382 Gbps; receive stayed at zero because the two ports are not connected to each other.

How to run

cmake -S . -B build -DCMAKE_CUDA_ARCHITECTURES=120 && cmake --build build -j
sudo ./build/examples/daqiri_bench_raw_gpudirect \
  ./build/examples/daqiri_bench_raw_sw_loopback_rtx_pro_6000.yaml --seconds 30

Hardware on the development system

• Five RTX PRO 6000 Blackwell GPUs
• Five ConnectX-7 / BF-3 NICs (400 Gbps class)
• One dual-port card at 61:00.0 and 61:00.1; both ports report link up
• No QSFP cable between port 0 and port 1

Tests completed without additional hardware

• Software loopback (GPUDirect validation)
• NIC transmit on port 0
• Reorder over software loopback
• Socket UDP kernel baseline

Next steps

Hardware needed	Planned work
QSFP cable or passive loopback between 61:00.0 and 61:00.1	Re-run the NIC YAML; add HDS and RoCE closed-loop configs; record TX, RX, and drops
Cable or switch between two NICs	Complete the generic template for the 800 Gbps cross-card target (GPU 0 TX, GPU 1 RX)
FFT and GEMM workload definitions in the repository	Application-level benchmarks required by #17

Outside #17, we will address these in the future:

HDS, RoCE closed-loop, FFT, GEMM, and Python bindings (#18).

[greptile-apps]

Greptile Summary

Adds four RTX PRO 6000 Blackwell benchmark YAML configs (software loopback, dual-port NIC, same-port experimental, and a generic placeholder template) together with a baseline results doc and the required doc-sync updates across AGENTS.md, raw_benchmarking.md, and configuration-walkthrough.md.

• Both commits carry a valid Signed-off-by trailer and the #17 - … commit-title prefix.
• The new YAMLs follow established conventions: kind: device/affinity for CUDA device targeting, angle-bracket placeholders for host-specific values (eth_dst_addr, PCIe BDFs in the template), and the parse_rx_configs/parse_tx_configs list-vs-map duality is handled correctly by the existing parser (verified in raw_bench_common.cpp).
• Two occurrences of a contributor name ("Cliff's 800 Gbps …") appear in the public-facing walkthrough doc and the template YAML header — swapping for neutral descriptive text is the only suggested change.

Confidence Score: 5/5

Safe to merge — this is a documentation and YAML-config-only addition with no library code changes.

No C++/CUDA or CMake library code is modified. The new YAMLs parse correctly under both the map and list forms handled by raw_bench_common.cpp, placeholders follow the project convention, doc-sync obligations are met, and both commits are signed off with the correct title format.

No files require special attention beyond the optional wording cleanup in docs/tutorials/configuration-walkthrough.md and examples/daqiri_bench_raw_tx_rx_rtx_pro_6000.yaml.

Important Files Changed

Filename	Overview
AGENTS.md	Benchmark table extended with four new RTX PRO 6000 YAML filenames; change is correct and minimal.
docs/benchmarks/raw_benchmarking.md	New RTX PRO 6000 tip block added alongside the existing DGX Spark tip; covers all four new YAMLs with hardware-constraint notes and run commands.
docs/tutorials/configuration-walkthrough.md	Four new RTX PRO 6000 bullet entries added; one bullet references "Cliff's 800 Gbps cross-card target" — a personal-name reference in public-facing docs.
examples/CMakeLists.txt	Four new YAML filenames appended to DAQIRI_BENCH_CONFIGS list; straightforward addition, no logic changed.
examples/daqiri_bench_raw_sw_loopback_rtx_pro_6000.yaml	Software-loopback smoke test for RTX PRO 6000; uses map-style bench_rx/bench_tx matching existing loopback config conventions; buf_size 8064 matches payload+header; looks correct.
examples/daqiri_bench_raw_tx_rx_rtx_pro_6000.yaml	Generic placeholder template for dual-NIC cross-card 800 Gbps target; uses list-style bench_rx/bench_tx (parser supports both forms); eth_dst_addr and PCIe BDFs correctly left as angle-bracket placeholders; comment references "Cliff's 800 Gbps vision" — personal name in a public file.
examples/daqiri_bench_raw_tx_rx_rtx_pro_6000_nic.yaml	Prefilled dual-port NIC config for dev-box PCIe BDFs; eth_dst_addr correctly left as placeholder; hardware limitations documented; list-style bench_rx/bench_tx parsed correctly.
examples/daqiri_bench_raw_tx_rx_rtx_pro_6000_nic_same_port.yaml	Experimental same-PF TX+RX config; clearly documented as failing daqiri_init; kept for follow-up; no functional concerns.
examples/rtx_pro_6000_baseline.md	Baseline results doc; hardware constraints, measured numbers, follow-up items all clearly stated.

_{Reviews (3): Last reviewed commit: "#17 - Doc sync and YAML fixes for RTX PR..." | Re-trigger Greptile}