NVIDIA's old Tesla V100 GPUs outpace new rivals in AI benchmarks
A modder put four NVIDIA Tesla V100 GPUs—originally launched in 2017—on a single board using NVLink and custom cooling, then put the setup through AI inference tests against much newer hardware. The result: this Frankenstein quad-V100 rig beat both Apple’s M5 Ultra and two DGX Spark units in several benchmarks, sometimes by over 200%. Still, pulling it off required hardware hacking and space for a hefty, power-hungry system.
- Quad-V100 rig used NVLink and custom heatsinks
- Beat M5 Ultra by up to 230% in prompt benchmarks
- Surpassed DGX Spark by up to 167% performance
- V100 GPUs with 32GB sell for around $744 each on eBay
- Setup needed significant modification and cooling
Sources covering this
In this story
More in Chips
Microsoft unveils Maia 300 AI chip, plans to scale deployments
Microsoft just unveiled its Maia 300 AI chip, the third version of its in-house silicon, and said it will shift most of its AI services…
AMD unveils Helion rack for agentic AI with up to 32,768 cores
AMD just announced the Helion rack, a new open design for building AI server systems.
Intel launches D-Series AI GPUs with up to 1.3 PFLOPS
Intel has officially named its Crescent Island chip the D-Series GPU and revealed new technical specs.
Nvidia and Cisco announce ultra-dense 400Gb/s USFP optical modules
Nvidia and Cisco unveiled a new optical module called USFP, designed to deliver 400Gb/s per lane and up to 12.8Tb/s of network capacity.