The NVIDIA Tesla V100 are 370% more powerful than the Radeon Instinct MI60
It is discovered that AMD lied and tampered with the AMD Radeon Instinct MI60 benchmarks that they can't with NVIDIA's seedier Tesla.
Despite the fact that Halloween has been around for a few days and that 'trick or treat?' it seems that there is a brand that has used it. If Intel or an outside company recently manipulated the benchmark data, now it is AMD's turn. A few days ago I present to great fanfare the AMD Radeon Instinct MI60 which is based on the Vega 20 architecture and 10nm lithography. According to that benchmark, Instinct MI 60 was very close to the NVIDIA Tesla benchmarks, although it was a lie.
AMD manipulates the Radeon Instinct MI60 benchmarks to get closer to the NVIDIA Tesla
The software where the fraud has been shown is the Resnet-50 Training, a benchmark specifically designed for Artificial Intelligence. We saw how the initial data from AMD showed the Instinct MI60 with 334, just below the NVIDIA Tesla V100 PCIe that gave 357 points.
The situation is more ridiculous now since we know that the NVIDIA graphics was working at 1/3 of its capacities and even so, with this trap, it did not manage to beat it in performance. AMD stupidly insists on the Vega architecture that it had to put HBM2 gaming memory to so that it wouldn't be a hideous ridiculous and would not fall behind the RX 580. GCN is design more than dead and the company needs to innovate, but in a hurry .
NVIDIA logically has not been silent at any time and has come out saying that the NVIDIA Tesla T4 benchmark that consumes 70W in that same benchmark offers 395 frames per second thanks to the Tensor Core, which represents almost 60 points more than the Radeon Instinct MI60 that consumes 300W.
NVIDIA statement: “The NVIDIA Tesla T4, with a 70W power consumption and its active Turing Tensor Cores, offers more training performance than the 60W Radeon Instinct MI300. The Tesla V100 can deliver 3.7x times more training performance with Tensor Cores and Mixed Precision (FP16 / FP32), enabling faster solution time while converging neural networks to required precision levels "
And so friends is how magic is done in a benchmark. The FP32 is still an outstanding measuring stick and here it is in the only field the Instinct MI60 is equated with the Tesla V100. But it is that in the other tests the difference is 370%, which is a bloody madness thanks to the Tensor Cores from NVIDIA.
AMD has given its unfortunate version of events: “Regarding the comparison, our footnotes for that slide clearly pointed out the modes, so there are no problems. The reason is that the FP32 measurement is used in most cases so that FaceID has an accuracy of 99.99%, for example, in banks and other instances that require high levels of precision ”
Pathetic explanation for an action that had no possible justification when Intel did it and that does not have it at this time with AMD. Vega, let me speak in silver, it is a shitty architecture that is worthless. AMD bet on Ryzen leaving Radeon to its fate and now this division is a decade behind NVIDIA.
Yes, there are no games for Ray Tracing and DLSS, but Artificial Intelligence is a reality and NVIDIA has the Tensor Cores and RT Cores. What does AMD have? Absolutely nothing.
Source: wccftech
