Intel Core Ultra 7 270K Plus Performance In 340+ Linux Benchmarks Review
Posted by Artoriuz@reddit | hardware | View on Reddit | 33 comments
Posted by Artoriuz@reddit | hardware | View on Reddit | 33 comments
Artoriuz@reddit (OP)
For context, in the Techpowerup review, the 270K roughly matches the 9950X in application performance. Here, it is slightly behind even the 9900X. The difference probably comes from the machine learning tests, as these greatly benefit from things like AVX-512, which the 270K lacks.
kuddlesworth9419@reddit
Not having AVX-512 is a pretty big problem. I assume the next Intel chips will support it at least. Normally software is behind hardware but in this case software is ahead of Intel.
jaaval@reddit
For scientific computing yes. For average user it’s not a problem.
Kryohi@reddit
The software used by an average user can very well use libraries that benefit a lot from AVX512/AVX10 under the hood. It doesn't only accelerate FP HPC stuff, see e.g. simdjson
jaaval@reddit
See e.g. simdjson, how much faster is it going to make average user’s computer. 5ms faster in an operation he does twice a day? It seems avx512 can make json parsing 10-30% faster. Who regularly parses gigabyte size json files except for large data analysis work or in servers running databases?
As we can see in the benchmarks in the article, the actual effect of AVX512 is very selective. It is useful, I used to need it a lot in my previous work. But currently I have next to no benefit from large vector operations. And I am still much more likely to need it than the average user.
Kryohi@reddit
What does the average user's computer do anyway? If you're saying only JavaScript performance matters, then yes, big simd instructions won't make a difference.
But for anything that has to do with audio, videos, images and media in general for instance avx512 can help a lot. Simdjson is often cited only because it is benchmarked by phoronix.
Artoriuz@reddit (OP)
Exactly. Nobody should be buying a 270K if the only use case is browsing Reddit.
Intel is clearly marketing it for productivity, and AVX-512 is great for productivity as seen in these benchmarks.
VenditatioDelendaEst@reddit
In the geomean, but if you look at the per test breakdown the geomean appears to be driven by a few high-multiplicity test suites that benefit from AVX-512, where AMD has a large lead.
Noreng@reddit
Michael at Phoronix tends to get the systems running, then runs through the tests without any tweaks. In the past this has caused CPUs with E-cores to single-threaded tests on the E-cores rather than the P-cores. There might be some (lack of) scheduling shenanigans going on here, I'm not going to say it's guaranteed by any means though.
Artoriuz@reddit (OP)
Intel decided to ship CPUs with asymmetric cores, so it's up to them to make sure scheduling works as expected. The user should not need to tweak anything.
Noreng@reddit
I absolutely agree on that part, but default settings might not give optimal performance.
For example on my laptop with a 258V, the P-cores are practically always powered down completely, because it's defaulting to low-power mode. This is fantastic for portable use, but performance obviously suffers.
nanonan@reddit
That's still on Intel to fix, not Phoronix. Should a benchmark of your laptop be measured on default behaviour or should it be hand tweaked to deliver different results from what a regular user would encounter?
Noreng@reddit
Right, and I suppose the absolutely horrible battery life on Linux with Strix Point is on AMD to fix then?
gamebrigada@reddit
This is unfortunately the problem with Intels architecture of different sized cores. The CPU's realistically cannot maintain TDP with P cores doing much.
Noreng@reddit
It's nothing to do with power draw, but rather that the scheduler doesn't move anything off the LPE-cluster.
gamebrigada@reddit
And it does this because..... It's trying to limit power draw....
Scheduling on mismatched cores has been implemented in Windows for years. Its intentional.
Euler007@reddit
And it's 200$ cheaper.
RhubarbUpper@reddit
In Canada they are marking up the 270 heavily, PC parts are just insane these days
Euler007@reddit
I see it 180$ cheaper than the 9950x on ça.pcpartpicker.com
vegetable__lasagne@reddit
Are they running at different clocks/power limits? Average power in multithreaded tests for phoronix is ~200W so it's probably clocking lower than at TPU.
Artoriuz@reddit (OP)
Michael reported an average of 154\~157 W and a peak of 259 W. TPU has their average at 140.4 W and their multi-threaded power consumption at 248 W with the default power limit and 265 W without them.
ElementII5@reddit
Oh boy, considering Zen 5 is on a 7nm and 5nm node and the Intel chip on a 3nm node AMD is going to wipe the floor with Zen 6.
Rekt3y@reddit
Zen 5 dies are on TSMC's N4P, and Zen 5c dies are on N3E for desktop, Threadripper and Epyc
Kryohi@reddit
There is no desktop or threadripper part on N3E/zen5C
Rekt3y@reddit
Epyc?
SmashStrider@reddit
9950X3D is on an enhanced N4P for the compute die vs N3B for the 270K, not simply '7nm and 5nm vs 3nm' as you say, which misrepresents the true capability of the nodes being used. Also, Arrow Lake lacks AVX-512 that Zen 5 has, something that will be compensated for again in Nova Lake. So the competition may be more than a simple "Zen 6 will wipe the floor against Nova Lake" as you seem to believe.
But yeah, keep misrepresenting the capability of hardware products and frame it in your own way to support your AMD stock options, as indicated by your post history.
Physical_Brief4935@reddit
Intel is most definitely striking back. They have a processor priced to compete aggressively with the flagship year old AMD processor.
Unfortunately for Intel, they are sacrificing their margins on the 270K plus at a time were PC build costs are so high that it will not translate to as much market share recovery as it typically would. AND they are going to have to claw back some mindshare.
Worse - for both AMD and Intel - they are about to get their worlds rocked in the laptop space by Apple - whose Neo laptop is absolutely crushing it.
Also anyone thinking that AMD stock value is tied to 9950X3D sales is delusional - the current AMD share price is based on datacenter traction and speculation about the MI450's competitive. Intel doesn't currently have a strong contender in this space.
31c0c3@reddit
yeah with this logic we should be including 5nm and 6nm for the gpu and soc tiles as well, not just saying 3nm
steve09089@reddit
Ehh, I think it has more to do with the 270K’s lack of AVX512, which Nova Lake will fix.
In other tasks that don’t benefit from AVX-512, Intel is competitive with the 9950X.
ElementII5@reddit
The intel part supports AVX2 though.
Messing around with the exported results via .csv.
Completely removing AVX tests --> AMD pulls ahead even further.
Leaving AVX2 but removing AVX512 test --> basically on par.
Leaving us with the earlier point I made. 7nm and 5nm class nodes vs. a 3nm class node and on par does not a good look make.
Geddagod@reddit
Extremely surface level analysis like this is exactly why you ended up being so wrong on your prediction that Turin is going to have 2x the perf/watt as GNR.
ElementII5@reddit
Hi, geddagod! Got that link somewhere to show the class?
We can come back to this thread when Zen6 eventually did wipe the floor with Nova Lake.
Artoriuz@reddit (OP)
Nova Lake will finally support AVX10, which I'm hoping will allow Intel to bridge the gap in these compute-heavy benchmarks.