Re:Intel demonstrates a 28-core processor running at 5GHz
by ajross on Tuesday June 05, 2018 @10:58AM
(#17238248)
They're already shipping them. 28 core Skylake-SP parts have been available under the Xeon Platinum branding for almost a year now. This is presumably just a consumer paint job on the existing silicon. And yes, they're very expensive: $10k+ per chip.
[ Parent
| on HN ]
Re:Intel demonstrates a 28-core processor running at 5GHz
by ucha on Tuesday June 05, 2018 @12:00PM
(#17238941)
Yes, the Xeon Platinum 8180 had 28 cores clocked at 2.5Ghz with a TDP of 205W. I don't see how they could pull off 5Ghz this year, apart from using LN2...
Re:Intel demonstrates a 28-core processor running at 5GHz
by ajross on Tuesday June 05, 2018 @12:11PM
(#17239076)
You've gotten confused about how the numbers interact. The 2.5GHz is a base frequency calibrated to be under the TDP under all operating conditions[1], actual clocks are allowed to be higher (in particular when only a subset of cores are used), but subject to thermal throttling. The chip turbos to 3.8GHz in single core operation. Surely the 5GHz number is likewise a turbo rate.
This sounds extraordinarily doable to me. I'd be very shocked if this was anything but an off-the-shelf SKL-SP die with consumer fusing.
[1] And given that it's a server part, this calibration is inherently more conservative than you see in desktop stuff.
Re:Intel demonstrates a 28-core processor running at 5GHz
by loeg on Tuesday June 05, 2018 @02:28PM
(#17240611)
That Cinebench score reflects a multi-threaded benchmark, which is not a workflow where the CPU would (or should, anyway) turbo a single core.
Also, while not AVX-intensive, CB does use some AVX, which usually draws more power (and thus limits clock more) than pure integer or even ordinary FPU computation.
Jumping from 3.8 GHz single core turbo to 5 GHz 28-core turbo is a big jump in power.