News for hackers. Stuff that matters.
Re:Basic Blackbird Bundle with 4-core IBM Power9 CPU by dragontamer on Monday November 26, 2018 @11:16AM (#18534352)
Power9 is one of the highest instructions-per-clock architectures in the world.

Power9 just has relatively weak vector units: only 128-bit. (Even AMD Threadripper has support for 256-bit. Power9 128-bit is executed per super-slice, so its really 2x64-bit, to compare against AMD's 2x128-bit support) Otherwise, its throughput is downright insane. 6-issues per clock (Skylake only has 6 if its in the uOp cache: otherwise the decoder only pushes 4 uOps), 10MB of L3 per pair of cores. That's 20MB of L3 on the 4-core, and 110MB of L3 cache on the 22-core.

-----

As far as price/performance goes, the 18-core / 72-thread chip seems best. But the 4-core / 16-thread or 8-core / 32-thread chips aren't bad, as far as I can tell.

The main issue is that most compute-heavy code these days use those vector units (2x128 on AMD, 2x256 on Intel). But if you got a big GPU for SIMD / vectorized ops, then that disadvantage kinda goes away. GPUs are way fatter than Intel or AMD.

The other issue is that AMD is offering cores at stupidly low prices. AMD Threadripper and EPYC are the machines to look at for raw CPU power these days. But I'd argue that the 18-core Power9 ($1,050) is at a good price/performance level. If you can find any application that benefits from the stupid-high 90MB of L3 cache on the 18-core, then you're more or less golden.

I'd bet that databases love that L3 cache.

Re:Basic Blackbird Bundle with 4-core IBM Power9 CPU by dis-sys on Tuesday November 27, 2018 @09:18AM (#18542421)
thanks for the detailed info.

phoronix.com just published some Power9 vs. Xeon vs Epyc benchmarks[1]. It is actually not on par with Xeon/Epyc. also note that the motherboard used by those Power9 has a price tag of $2499, when a dual socket LGA3647 motherboard from a first tier vendor can be purchased at $400 delivered.

[1] https://www.phoronix.com/scan.php?page=article&item=power9-x...

Re:Basic Blackbird Bundle with 4-core IBM Power9 CPU by dragontamer on Tuesday November 27, 2018 @11:06AM (#18543467)
A lot of the "weakness" there is due to the weak 4x64-bit vector units equipped on the Power9. Again: 4x64-bit vector units (Power9) simply has less throughput than 4x128-bit (EPYC) units, which has less throughput than 3x512-bit (Xeon Gold) vector units.

On the few tests which don't have anything to do with vector units, Power9 does incredibly well. See 7-Zip Compression, Stockfish, Fhourstones, LLVM compiles. Power9 is very, very far ahead of the game here.

When vector units (ie: OpenMP or intrinsics) are involved: such as Parboil, Gimp, x264, Rodinia, etc. etc. the Power9 does very poorly.

-------------

Something else I've noticed: Power9 has a lot of latency behind its units: it is always 2-cycles or 3-cycles of latency (while EPYC or Skylake have many operations with 1-cycle of latency, like add or subtract). Power9 relies upon instruction-level parallelism to get things done.

So there are a few cases like the Rust Mandelbrot benchmark, which surely hits that latency problem, while the Rustlang Prime benchmark shows off the ILP goodness the system offers.