Intel's flagship Arc Alchemist GPU has lackluster performance for the first time in leaked benchmarks

  • 2021-11-22 22:16
  • |
  • 55
  • |
  • Recompile By Brian - CPU Review
  • |
  • Comments

what to expect: Intel's upcoming Arc flagship has appeared for the first time in an online benchmark database. An unnamed Intel partner sampled GPU engineering in the Geekbench OpenCL test suite using the somewhat old i5-9600K validation platform, and here's what we learned...

Overall, the GPU performs poorly. The four runs achieved an average OpenCL score of about 67,000 points and its peak score of 69,000 points. For comparison, the AMD Radeon RX 6800 XT released about a year ago is in the 150,000 to 180,000 ball point range.

But don't be discouraged. Consider this: according to the metadata collected by Geekbench during the run, the GPU oscillates at around 1.3 GHz. At that frequency, with 512 EUs, the GPU has a theoretical 11 TFLOPS performance.

AMD's Radeon RX 6600 XT has the same 11 TFLOPS and scores between 60,000 and 70,000 points. Therefore, Intel's architecture has the same performance per TFLOPS in OpenCL as AMD's.

So the question is: can Intel match AMD's high-end TFLOPS numbers?


512 is the magic number that says this is the top, but that's 512 EUs, not CUs.

According to Geekbench, the GPU has a rated clock speed of 2.1 GHz. If it ran at that speed, and not at 1.3GHz, it would have 17 TFLOPS, which is on par with the RX 6800 (not the XT). So the answer to that question is, yes, if 2.1 GHz can be achieved.

And to be honest, 2.1 GHz is not that high. If the rumors are true and Intel is using the TSMC N6 node, then they are likely to push the clocks high enough to break into the performance territory of the RX 6800 XT and RTX 3080.

Keep in mind that this is pure conjecture, but we're not the first to claim that Intel's GPUs will land somewhere with the RTX 3070 or RTX 3080. Intel itself also hinted that its flagship they will cost more than the latter at $740 to $790. However, as today's OpenCL scores suggest, Intel may still be working to get there.

Comments