NVIDIA’s 9-Year-Old Tesla V100 Still Packs Some Inference Muscle; 4 GPUs On An NVLink Board Can Outperform M5 Ultra, DGX Spark In Various AI Benchmarks By Up To 230%
NVIDIA’s 9-Year-Old Tesla V100 Still Packs Some Inference Muscle; 4 GPUs On An NVLink Board Can Outperform M5 Ultra, DGX Spark In Various AI Benchmarks By Up To 230%
摘要: Hardware
NVIDIA’s 9-Year-Old Tesla V100 Still Packs Some Inference Muscle; 4 GPUs On An NVLink Board Can Outperform M5 Ultra, DGX Spark In Various AI
正文
Hardware
NVIDIA’s 9-Year-Old Tesla V100 Still Packs Some Inference Muscle; 4 GPUs On An NVLink Board Can Outperform M5 Ultra, DGX Spark In Various AI Benchmarks By Up To 230%
Omar Sohail
Oct 12, 2026 at 08:01am EDT
Comments
Add to Preferres Sources
Looking for an affordable inference monster setup? NVIDIA’s Tesla V100 shows that it’s still got it / Image credits - Jun Yamog
The Volta architecture, introduced in 2017, can still hold its own across a multitude of AI-related workloads, with the Tesla V100 and its 32GB of HBM2 memory and 900GB/s bandwidth sufficient to take on modern-day solutions. Of course, it took an NVLink carrier board and some heatsink modifications to ensure the hardware was mounted and working properly, but once that obstacle was cleared, it was off to the races. Against the
M5 Ultra
DGX Spark
, four V100s can deliver up to a 167 percent performance uplift.
In Qwen3.8-Flash-Next, the Tesla V100 can effortlessly hold its own against the competition, as long as you’re willing to accept the heftier size and additional modifications
Without an NVLink carrier board, it looked impossible to house four separate V100 units, but YouTuber Jun Yamog got his hands on the additional hardware that can deliver 300GB/s of bandwidth, with a total of six NVLink 2.0, resulting in the increased speeds. Even if mounting each of the four Tesla V100 GPUs wasn’t an arduous task, getting them to cool properly would have been.
Related Story
AMD Rumored To Launch RDNA 5 In Second Half Of 2027; Also Quashes Rumors Of Next-Gen Using 4 GB GDDR7 Memory
Image credits - Jun Yamog
This is why Jun modified NVIDIA’s A100 GPU heatsink by milling it down to accommodate the V100, but when the process was complete, the giant cooler could effortlessly dissipate 300W of power being generated from each chip. With a couple of passthrough adapters, the YouTuber connected the carrier board via a PCIe x8 connection.
Image credits - Jun Yamog
Coming to the benchmarks