Intel's New Low: Commissioning Misleading Core I9-9900K Benchmarks [video]
youtube.com
youtube.com
In one case a third party who were demonstrating how much faster Intel Xeon Phi was for deep learning admitted that they were comparing highly-optimised code to unoptimised code in their results.
This doesn’t surprise me at all.
If you have simple if-statements and all branches are grouped together to be SIMD'd easily... then yeah. GPU threads kind of are like normal CPU threads.
But as soon as you have a serious degree of thread-divergence, your performance tanks. That's why things like Chess engines (which ARE parallel problems at heart), execute poorly on GPUs. Because even though it is massively parallel, chess has too many if-statements and can't extend to SIMD very easily.
--------
Raytracing algorithms are funny: they group rays together so that the GPU can SIMD over them more easily. But without the "re-grouping" step, its bad performance.
Ex: A bunch of rays start at the camera. Some might hit a diffuse surface like wood... some might hit a subsurface-scattering surface like skin, and others might hit a metalic surface. GPU-raytracing algorithms then save off all the rays, and then processes all "diffuse" rays together, to minimize divergence.
You can't just follow a singular ray in raytracing on a GPU. You gotta re-group to SIMD units for maximum performance.
FWIW, I have an AMD GPU with ROCm. HIP it's a lot like CUDA, so NVidea-focused tutorials ought to be fine. With the caveat that I'd have to be aware of hardware differences.
Lets use a simple example:
for(int i=0; i<10000000; i++){
if(someCondition)
break;
else
doStuff();
}
GPUs execute roughly 32-threads with the same instruction pointer. Lets say thread#0 had "someCondition" to be true. Then thread#0 will be set to "disabled", but otherwise, it will have to wait for the 31-other threads to be done with the loop before continuing.
Even if 31-threads have hit "someCondition" and have broken out of the loop, the 32nd thread will keep executing the loop until it is done (and threads 0-through-30 will "execute with" the 32nd thread, but will throw away the results).
That's the key with SIMD. Threads are run in groups of ~32ish at a time, at the same time. All 32-threads must execute if statements together and loops together.
In most cases, an if/else statement will be executed by BOTH threads (but the results "thrown out" by the GPU engine, through execution masks)
Actually, any "professional" journalism outfit would show the public the newsworthy results, not withhold this information from readers because of a desire not to piss off industry contacts.
It's a sad state of affair, but that's the reality of it.
Not sure I fully get it though, a scoop is a scoop
NDAs that try to restrict the content of a review and not just the publication date are reprehensible, but that doesn't seem to be what's going on here. Intel's just trying to soften the ground with their own misleading numbers, but still permitting the trustworthy sources to do their usual thing.
(Which, interestingly, means some sites that don't issue numeric scores are barred-by-default from those games. I think that's sometimes been used to spot these embargoes?)
In the computer hardware world, a review embargo that coincides with the product hitting the shelves does not carry any negative connotations about expectations for the product's reception.
the video doesn't dispute the benchmarks of the 9900, but rather they dispute the published compared benchmarks of the ryzen 7 2700X, which used default, unconfigured memory settings to push the benchmark down, while the 9900 benchmarks used configured memory settings (as it should)
It's also available for desktop parts to disable one ccx when game doesn't use more than 8 threads and you want to squize a couple more % of perf.
Game mode is disabled by default, requires AMD software to be installed to enable - they went extra mile to gimp competition..
One amusing thing I've noticed is that, playing the same game (Minecraft), Windows runs it ~25% faster in game mode than default NUMA.
In Linux the situation is reversed; it runs 10% faster with NUMA enabled, maybe because the Java garbage collector is NUMA-aware and I'm using enough memory that it's split across NUMA nodes anyway.
But either mode is faster in Linux than Windows.
At $499, the i9-9900k is almost competing against the 12-core Threadripper 2920x ($649, 12-core/24-threads, 4.4GHz clock, 60 PCIe lanes, quad-memory channels).
I think most people will find more use out of +4 cores (granted, on a NUMA platform) than higher clocks. Cores for compiling code, rendering, video editing, etc. etc.
Pretty much only gamers want +Clock speeds, and more and more games actually use all-cores these days (Doom, Battlefield, etc. etc.)
-----------
That's the thing. The i9-9900k isn't even a "high end chip" anymore. Its at best, "highest of the mid-range" since the HEDT category (AMD Threadripper, or Intel-X) has been invented.
Once you start getting into 8-cores/16 threads, I start to worry about dual-memory channels and 16x PCIe lanes + 4GB/s DMI to the Southbridge. Its getting harder and harder to "feed the beast". A more balanced HEDT system (like Threadripper's quad-memory channels + 60PCIe lanes) just makes more sense.
I wish. We use a commercial path-tracer that scales very well to many cores, GPUs and entire clusters when it's chewing away at a single fixed scene or animation.
But in interactive mode many scene modifications are bottlenecked on a single or few threads and locks until it gets back into the highly optimized rendering code paths. So a lot of work goes into quickly shutting down as many background threads as possible to benefit from high turbo-boost clocks on Xeon Gold processors so the user doesn't have to wait long and then ramp them back up when it's just rendering the fixed scene.
For that you can get a Ryzen 2700X, a nice motherboard and a 256GB SSD. Performance delta shouldn't be more than 15% deficit for the Ryzen for a few specific games.
So maybe you could translate it as "quantified difference".
Cloudflare ran into some trouble with this: https://blog.cloudflare.com/on-the-dangers-of-intels-frequen...
You need the "X-series" to get AVX512.
AVX2 is ~2x, Ryzen/AMD fakes AVX2 instructions with multiple SSE instructions.
Some AVX2 instructions downclock but not much, I see very close to 2x speedup over SSE2 with some workloads. Some of the downclock loss is made up for because there are more instructions available (gather, etc)
AVX512 might hit more than 4x improvement over SSE on some workloads despite the downclocks, due to all of the masking features. I have seen results consistent with this, 2nd hand. (I don't own an AVX512 cpu)
Anyway all of these things depend on workload, cpu, compiler etc. But it does happen!
Not quite. Everything is fake, because everything is encoded into micro-ops first.
Ryzen's internal micro-op engine is 128-bit wide. But it has 4x pipes... each handling 128-bits at a time. So any 256 bit instruction will simply use two-pipes at a time.
-------
So the 256-bit instruction does in fact, execute at once.
The difference is that Intel has 3-pipelines, each of which can do a 256-bit instruction by itself.
-------
In effect: Ryzen is 4x128-bit pipelines, with the ability to merge pipelines together to do a 256-bit instruction.
Intel is 3 x 256-bit pipelines, with the ability (on Skylake-X) to merge pipelines together to do a 512-bit instruction
In any case, Intel has wider pipes than Ryzen. Intel Skylake is effectively a 256-bit CPU, while Ryzen is only a 128-bit CPU.
> I don’t have too much of an issue with Intel commissioning the report itself, and the Principled Technologies report is very transparent as they clearly state how they tested the games and configured the hardware. The results and testing methods are heavily biased, but they haven’t attempted to hide their dodgy methods. You can dig into the specs and find all the details, it’s still dodgy but it’s a paid report, so it’s somewhat expected.
Vast majority of buyers and sadly huge chunk of tech press won't be able to tell looking at the settings if our how much was benchmark skewed in Intel's favor.
And it plenty skewed.
Ryzen 5 2500U 256GB SSD 16GB RAM
for only $859. I'm perfectly happy with the speed of the CPU, and the integrated GPU is fantastic (can even play Destiny 2 on low settings at 720p). That, and the build quality is absolutely superb.
Keep in mind, I bought this for programming at home, not gaming. But I'm glad that it works as a portable light gaming rig when needed.
AMDs share price is up almost 1400% in 2 years, what on earth are you on about?
I have a feeling it'll make its way back up though.