Nvidia Calls Out Intel For Cheating In Xeon Phi vs GPU Benchmarks (arstechnica.com)

An anonymous reader writes: Nvidia has called out Intel for juicing its chip performance in specific benchmarks -- accusing Intel of publishing some incorrect "facts" about the performance of its long-overdue Knights Landing Xeon Phi cards. Nvidia's primary beef is with the following Intel slide, which was presented at a high performance computing conference (ISC 2016). Nvidia disputes Intel's claims that Xeon Phi provides "2.3x faster training" for neural networks and that it has "38 percent better scaling" across nodes. It looks like Intel opted for the classic using-an-old-version-of-some-benchmarking-software manoeuvre. Intel claimed that a Xeon Phi system is 2.3 times faster at training a neural network than a comparable Maxwell GPU system; Nvidia says that if Intel used an up-to-date version of the benchmark (Caffe AlexNet), the Maxwell system is actually 30 percent faster. And of course, Maxwell is Nvidia's last-gen part; the company says a comparable Pascal-based system would be 90 percent faster. On the 38-percent-better-scaling point, Nvidia says that Intel compared 32 of its new Xeon Phi servers against four-year-old Nvidia Kepler K20 servers being used in ORNL's Titan supercomputer. Nvidia states that modern GPUs, paired with a newer interconnect, scale "almost linearly up to 128 GPUs."

Comments

Hmmm.

[Anonymous Coward]

So Intel is the Volkswagon of CPUs?

Re:Hmmm.

[oic0]

Not quite. This is deceptive advertising. You can deceive customers all you want so long as you have enough fine print don't outright lie.

Re:Hmmm.

[Mashiki]

Only if you think this is new. Intel has been doing shit like this for year and keeps getting caught, there were two lawsuits against them from AMD a few years ago where they ended up paying AMD around $7B USD for doing things like this and other forms of anti-competitive behavior which resulted in multi-billion dollar fines. Then again, nvidia has been caught doing the same. Probably the best example most recently is with their "Hairworks" API, which is likely going to land them in hot water again. Nvidia got nailed a few years ago for anti-competitive behavior over shaders.

Re:Hmmm.

[Luthair]

It wasn't benchmarks, it was changing their commercial (as in users pay for it) compiler to ignore CPU flags for non Intel parts and not documenting it, then various illegal behaviour with effectively paying OEMs to not use AMD parts.

Re:Hmmm.

[bluefoxlucid]

It's not so much "older" as "different" in the artificial benchmarking world. Real-world loads don't tend to follow benchmarks religiously, and the newer benchmark might favor a configuration that's not as good in real-world loads.

The classical marketing maneuver is to select from multiple sets of up-to-date benchmarks and pick the ones that favor your particular product. CPUBoss usually shows that one CPU outperforms another consistently (except for single-core vs threaded with dissimilar cores or SMT--fast clock wins single-core, many-cores wins threaded); and frequently shows the same benchmark tool using different strategies and rating each CPU faster than the other based on how it was configured, or shows that one benchmark favors one CPU and another favors the other.

This goes all the way up to real-world functional tests, where you select games which perform better because of some feature or strategy of your GPU and CPU. You have better shaders? Pick a shader-heavy game. Heavy parallelism? Pick a game that meshes with that. You've got fewer parallel operations, but a higher clock? Avoid games that work best with 387-core GPUs and pick ones that like that 1185MHz clock. Show off 6 or 7 games running at freakishly-high 292fps.

Re:Hmmm.

[PhrostyMcByte]

Intel optimized per-architecture, not per feature. This had the end result of AMD chips taking the generic path and being slower, but I wouldn't call this tactic dirty. Why would Intel go out of their way to optimize for a competitor?

CPUs have a wide variety of timing and pipeline limitations, and optimizing purely for feature set will never get you peak performance -- this is why GCC has the exact same per-architecture optimization support.

Re:Hmmm.

[hairyfeet]

I guess "Windows isn't done unless Lotus won't run" by your logic is completely reasonable behavior, it WAS their OS...right?

Or maybe if you'd stop waving your little Intel flag as hard as your squeeing fangirl ass can you'd know they didn't "not optimize" for other chips, they purposely designed their compiler to put out broken code on other chips so badly in fact that you could take a Via CPU (the only CPU that allows you to change the CPUID in software) and by simply changing the CPUID from "Centaur Hauls" to "Genuine Intel" you magically got a 30% performance boost...wow, the power of of CPUID huh?

Of course what it really was was a classic case of "Windows isn't done unless Lotus won't run" and this kind of behavior is typical of Intel, hence why they had to shell out 1.4 billion for market rigging and anti competitive behavior in the EU just 2 years ago. Would you like a quote from the judgement?

"The Commission demonstrated to the requisite legal standard that Intel attempted to conceal the anti-competitive nature of its practices and implemented a long term comprehensive strategy to foreclose AMD from the strategically most important sales channels. ... The General Court considers that none of the arguments raised by Intel supports the conclusion that the fine imposed is disproportionate. On the contrary, it must be considered that that fine is appropriate in the light of the facts of the case.

In other words the exact same shit MSFT got busted for and frankly they should get no less than what MSFT did, 10 years of being monitored by the courts to keep them from pulling shit like this again.

Here's the real reason for Nvidia's complaints

[CajunArson]

The real reason that Nvidia is bitching up a storm is that KNL has received a very positive reception in the HPC world.

Oh, and KNL is actually an absolute bargain in comparison to the requirements to get a high-end Pascal system setup, not only because you can buy an entire KNL system (not just a GPU card) starting at only $5000, but because it's self-hosting and doesn't need a high-end Xeon CPU just to feed the GPU. To put it in perspective, you could build a cluster of 26 KNLs for the price of one of those 8-way systems Nvidia is selling.

http://www.colfax-intl.com/nd/...

Re:Here's the real reason for Nvidia's complaints

[Creepy]

Yeah, but their beef isn't about the cost, it is about the speed comparisons. Intel never has tried to compete in the GPU performance space - they are happy with being in the low cost space. If you just compare what you get for a certain cost I have no idea, but I'm guessing having so many more Intel chips in your cluster will add significant power and space requirements at the very least. You may actually be better off with the nVidia solution in the long run.

Re:Here's the real reason for Nvidia's complaints

[Guybrush_T]

Yes, the HPC world is waiting for KNL because they don't want to port their old codes to CUDA. But that's just the expectation : people are starting to realize that running a Xeon code on KNL is by no mean immediate and you won't get much performance boost without a serious application rewrite ... just like porting to GPUs, maybe slightly easier though.

But on the performance side, it is very clear that KNL performance is terrible. The fact that Intel only shows scaling figures is quite funny : it is very easy to make a slow code scale, because computation times are high compared to communication times. To have good scaling, you can either have a faster interconnect or a slower CPU. Since they're never showing performance comparison but only "scaling", I'd bet it is the latter.

To illustrate, say the speed of your code is 1 on 1 CPU, and 32 on 32 CPUs, scaling is perfect. If the speed is 100 on one GPU, and 2400 on 32 GPUs, the scaling is not perfect and you can show the scaling curve from Intel saying "hey, we scale better !". That's ridiculous.

Worlds smallest violin for Nvidia

[Moheeheeko]

After they spent the last decade going out of their way to force "improvements" into new games (that either never work or cause severe issues for pc games) just to make their overpriced cards look better than AMD, they can go fuck themselves.

Re:Worlds smallest violin for Nvidia

[thegarbz]

If by improvements you mean optimsations, and by overpriced cards you mean get what you pay for, and by better than AMD you mean better than AMD then yeah you're 100% right.

Now you can repeat the same statement for AMD.
And for Intel
And for ARM
And for every other chip manufacturer who targets a specific market with specific products.