They might even do such a good job that AMD are left further behind (when both are running OpenCL).

But you also have external developers outside of AMD and Nvidia working on running native CUDA applications on the latest Intel gpu hardware. This Phoronix article today talks about running the Geekbench CUDA tests on Intel Xe/UHD gpus with equivalent OpenCL performance. Interesting development. So someday you might be able to run native CUDA applications on AMD hardware.
https://www.phoronix.com/scan.php?page=news_item&px=Intel-ZLUDA-CUDA-For-GPUs

You are a bit behind the curve, Intel are in the process of introducing cards that can be inserted into the system of one's choice in the shape of the Intel Xe series. It will be interesting to see how they stack up against the AMD and nVidia price equivalent offerings.

That said, both AMD & Intel have for many years offered "all-in-one" CPU & GPU chips with varying degrees of success and performance, as you imply these tend to have lower performance than (in the case of AMD) their discrete cousins, or (at least in the early days) the CPU taking a performance hit. These both have found a niche in business laptops, where their low price and "better than nothing" performance is attractive to many IT bosses. (I do note that they will often buy much higher spec computers including a discrete GPU for their own use, while they won't allow the mere mortal user have such a laptop - to much muttering from CAD-monkies.)

You mean a normal PCI-E graphics card? If not, I tend to stay clear of Intel's weird ideas. What was that weird hybrid disk thing they came up with, Optane memory I think, you might aswell just buy an SSD.

Something I only discovered the other day is that some at the top of AMD's design tree have been "headhunted" by Intel (Raja Koduri the driving force behind AMD's Radeon Technologies Group, followed by chip architect Jim Keller and graphics marketer Chris Hook and a few more, to name just a few.) It will be interesting to see if the move of these people from AMD to Intel will have any negative impact on AMD and a positive impact on Intel.

You mean a normal PCI-E graphics card?

Yes.

Even projects that use more DP, still use single precision.

Maybe so, but the problem with the projects that require Double Precision, is that you cannot use a Single Precision GPU, at all. So then the issue is moot whether the project only uses DP or also (in part) SP.

Which modern GPU has only single precision?
All modern GPUs do DPP, including modern Intel IGPs nowadays...

don't be so narrow minded to compare two cards without accounting for what they are designed for. gaming by and large does not need high DP performance, so the manufacturer has nothing to gain by making a gaming card with good DP performance, theyve realized this and shifted DP to being a feature on the more professional line of products. AMD is not immune to the DP performance reduction and they have been steadily reducing DP capability in their consumer cards as well.

if someone wants to spend "several grand" on a single card and DP is important, the will not buy an RTX 20x0/30x0 or even the new AMD cards, they will buy an Nvidia TitanV, which will absolutely wreck your old/cheap card. If they want to be a bit more frugal and can handle the increased power consumption, then they might go for a Radeon VII.

you get more compute density with the Titan V though, and it's much more power efficient than the AMD cards. There's a reason that the top host at Milkyway is using TitanVs.

but there's only a single project where DP is even relevant, Milkyway.

---

why speak in absolute terms for things that are not absolute? The world is constantly in flux.

AMD is leading Intel... NOW. but it wasnt the case for over a decade before last year.

AMD performance per dollar better than Nvidia? well that depends on which performance metric you chose and what variables you decide to ignore.

Gaming? not really, several of the RTX 30-series cards are a better value than AMD options on a FPS/$dollar metric.

Interesting that you bring up Primegrid, because Nvidia is dominating that leaderboard too... even if you ignore the 10x systems from Syracuse University with 10x RTX 6000 each, all of the top "normal" users are using Nvidia. I had to go all the way down to rank 60+ to find an AMD card. this is a project with both a CUDA (nvidia) and openCL (AMD) application, and the nvidia cards significantly faster on cards in the same price bracket (RX 5700XT vs RTX 2070) when using the same application (PPS Sieve).

so yeah, you can't just say AMD wins price/performance all the time. because that's objectively false.

---

this is a project with both a CUDA (nvidia) and openCL (AMD) application, and the nvidia cards significantly faster on cards in the same price bracket (RX 5700XT vs RTX 2070) when using the same application (PPS Sieve).

Now we have another variable, how well were the OpenCL and Cuda versions of PPS Sieve written?

I can't say for PPS Sieve, but for the HPC design project I was involved in a couple of years back OpenCL on AMD was considerably slower than CUDA on nVidia. At the same time the Intel server CPU offerings were ahead of the AMD server CPU offerings in terms of performance, and at the same time the bus-mastering capabilities of ARM beat them both and the project ended up with an interesting hybrid network of blocks of nVidia Qudros supported by Intel CPUs doing the data splitting and recombination, all buses controlled by ARM CPUs, meanwhile the power & thermal control was managed by a number of ARM CPUs. (I'd done the initial sketches of these systems on an Arduino as I had one sitting on the bench looking sad, lonely & underused and it did the job.)

This was a horses and courses and costs be blowed project.

That just means people have chosen Nvidias, it doesn't mean they were correct to do so.

they chose them for the fact that nvidia/cuda combo is faster and more power efficient. when both apps exist, Nvidia/CUDA has always been faster than AMD/OpenCL. but very very few projects have both app types available anymore, so we don't get many opportunities for true apples to apples comparisons. a lot of projects either don't have devs capable of coding well in CUDA, or don't have the resources (or will) to develop and maintain two different app versions.

but take a look at Folding@home too. ever since they came out with CUDA apps, Nvidia cards have dominated their PPD charts too.

https://folding.lar.systems/gpu_ppd/overall_ranks

They might have bought them for gaming too.

but i thought your argument was that AMD was better for gaming too? bit of a contradiction and cognitive dissonance there.

but this is a stretch at best, most of the guys topping the leaderboards are there because they run BOINC 24/7, not because they're doing a significant amount of gaming. So it makes more sense to reason that their buying choice was influenced by BOINC/project performance more than anything else.

Now we have another variable, how well were the OpenCL and Cuda versions of PPS Sieve written?

since the apps were created by the same project devs, it's likely they are coded with the same skill/capability.

---