[paulpan]

Agree with OP that Intel was probably too deep into its downward spiral. While it seems Pat tried to make changes, including expanding into GPUs, it either wasn't enough or too much for the Intel board.

Splitting Intel is necessary but probably infeasible at this point in the game. Simple fact is that Intel Foundry Services has nothing to offer against the likes of TSMC and Samsung - perhaps only cheaper prices and even then it's unproven to fab any non-Intel chips. So the only way to keep it afloat is by continuing to fab Intel's own designs, until 18A node becomes viable/ready.

[mnau]

He failed on GPU. The product was substandard.

That means either he knew and allowed it to happen, which is bad, or he didn't know and allowed GPU division to squander the resources, which is even worse. Either way, it was an adventure Intel couldn't afford.

There is a similar story in other areas.

[paulpan]

Disagree on "failed on GPU" as it depends on the goal.

Sure Intel GPUs are inferior to both Nvidia and AMD flagship offerings, but they're competitive at a price-to-performance ratio. I'd argue for a 1st gen product, it was quite successful at opening up the market and enabling for cross-selling opportunities with its CPUs.

That all said, I suspect the original intent was to fabricate the GPUs on IFS instead of TSMC in order to soak up idle capacity. But plans changed along the way (for likely performance reasons) and added to the IFS's poor perception.

[7speter]

The issue with the GPUs is that their transistor to performance ratio is poor. The A770 has as many transistors as about a 3070ti but only performs as well as a 3060 (3060ti on a good day).

So with that, they are outsourcing production of these chips to TSMC and using nearly cutting edge processes (battlemage is being announced tomorrow and will use either TSMC 5 or 4), and the dies are pretty large. That means they are paying for dies the size of 3080s and retaling them at prices of 3060s.

[ryao]

The A770 actually has more transistors than the RTX 3070 Ti:

RTX 3070 Ti: 17,400 million transistors

A770: 21,700 million transistors

https://www.techpowerup.com/gpu-specs/geforce-rtx-3070-ti.c3...

https://www.techpowerup.com/gpu-specs/arc-a770.c3914

It has taken Nvidia decades to figure out how to use transistors as efficiently as it does. It was unlikely for Intel to come close with their first discrete GPU in decades.

That said, it is possible that better drivers would increase A770 performance, although I suspect that reaching parity with the RTX 3070 Ti would be a fantasy. The RTX 3070 Ti has both more compute and more memory bandwidth. The only advantage the A770 has on its side is triple the L2 cache.

To make matters worse for Intel, I am told that games tend to use vendor specific extensions to improve shader performance and those extensions are of course not going to be available to Intel GPUs running the same game. I am under the impression that this is one of the reasons why DXVK cannot outperform the Direct3D native stack on Nvidia GPUs. The situation is basically what Intel did to AMD with its compiler and the MKL in reverse.

In specific, information in these extensions is here:

https://gpuopen.com/amd-gpu-services-ags-library/ https://developer.nvidia.com/rtx/path-tracing/nvapi/get-star...

Also, I vaguely recall that Doom Eternal used some AMD extension that was later incorporated into vulkan 1.1, but unless ID Software updated it, only AMD GPUs will be using that. I remember seeing AMD advertising the extension years ago, but I cannot find a reference when I search for it now. I believe the DXVK developers would know what it is if asked, as they are the ones that told me about it (as per my recollection).

Anyway, Intel entered the market with the cards stacked against it because of these extensions. On the bright side, it is possible for Intel to level the playing field by implementing the Vulkan extensions that its competitors use to get an edge, but that will not help it in Direct3D performance. I am not sure if it is possible for Intel to implement those as they are tied much more closely with their competitors’ drivers. That said, this is far from my area of expertise.

[timschmidt]

> He failed on GPU. The product was substandard.

I will never understand this line of reasoning. Why would anyone expect an initial offering to match or best similar offerings from the industry leader? Isn't it understood that leadership requires several revisions to get right?

[mnau]

Oh, poor multi billion company. We should buy its product with poor value, just to make it feel better.

Intel had money and decades of integrated GPU experience. Any new entrant to the market must justify the value to the buyer. Intel didn't. He could sell them cheap to try to make a position in the market, though I think that would be a poor strategy (didn't have financials to make it work).

[timschmidt]

I think you misunderstood me. I wasn't calling for people to purchase a sub-par product, rather for management and investors to be less fickle and ADHD when it comes to engineering efforts one should reasonably expect to take several product cycles.

Honestly, even with their iGPU experience, Arc was a pretty impressive first dGPU since the i740. The pace of their driver improvement and their linux support have both been impressive. They've offered some niche features like https://en.wikipedia.org/wiki/Intel_Graphics_Technology#Grap... which Nvidia limits to their professional series.

I don't care if they have to do the development at a loss for half a dozen cycles, having a quality GPU is a requirement for any top-tier chip supplier these days. They should bite the bullet, attempt to recoup what they can in sales, but keep iterating toward larger wins.

I'm still upset with them for cancelling the larrabee uarch, as I think it would be ideal for many ML workloads. Who needs CUDA when it's just a few thousand x86 threads? I'm sure it looked unfavorable on some balance sheet, but it enabled unique workloads.

[mnau]

> I don't care if they have to do the development at a loss for half a dozen cycles,

And here is the problem. You are discussing a dream scenario with unlimited money. This thread is about how CEO of Intel has retired/was kicked out (far more likely) for business failures.

In real world, Intel was in a bad shape (see margins, stock price ect) and couldn't afford to squander resources. Intel couldn't commit and thus it should adjust strategy. It didn't. Money was wasted that Intel couldn't afford to waste.

[timschmidt]

Well, seeing as GPU is important across all client segments, in workstation and datacenter, in console where AMD has been dominant, and in emerging markets like automotive self-driving, not having one means exiting the industry in a different way.

I brought up Intel's insane chiplet [non-]strategy elsewhere in the thread as an example where it's clear to me that Intel screwed up. AMD made one chiplet and binned it across their entire product spectrum. Intel made dozens of chiplets, sometimes mirror images of otherwise identical chiplets, which provides none of the yield and binning benefits of AMD's strategy. Having a GPU in house is a no-brainer, whatever the cost. Many other decisions going on at Intel were not. I don't know of another chip manufacturer that makes as many unique dies as Intel, or has as many SKUs. A dGPU is only two or three of those and opens up worlds of possibility across the product line.

[lmm]

Pulling out of a vital long-term project because it can't deliver a short-term return would be a bigger waste. Unless you think Intel is already doomed and the CEO should be pursuing managed decline?

[dralley]

It's worth mentioning that IIRC the team responsible for the Arc GPU drivers was located in Russia, and after the invasion of Ukraine they had to deal with relocating the entire team to the EU and lost several engineers in the process. The drivers were the primary reason for the absolute failure of Arc.

Intel deserves a lot of blame but they also got hit by some really shit circumstances outside of their control.

[mnau]

He was CEO. Chief executing officer. It's literally his job to execute, i.e. fix that stuff/ensure it doesn't happen. Get them out of Russia, poach new devs, have a backup team, delay the product (i.e. no HVM until good drivers are in sight). That's literally his job.

This only reinforces my previous point. He had good ideas, but couldn't execute.

[rvba]

Executive, not executing.

On a side note getting people in russia write your drivers sounds a bit insane. Yea lower cost and probably ok quality, but the risks...

[jpadkins]

CEO stands for Chief Executive Officer.

a chief executive officer, the highest-ranking person in a company or other institution, ultimately responsible for making managerial decisions.

Maybe you mean COO?

[Wytwwww]

> shit circumstances outside of their control.

They chose to outsource the development of their core products to a country like Russia to save costs. How was that outside of their control? It's not like it was the most stable or reliable country to do business in even before 2022...

[kgeist]

Russia is reliable when it comes to software engineering. I've met a few guys from Intel Russia, bright folks. The politics, though...

[Wytwwww]

Individual Russian software developers might be reliable but that's hardly the point. They should've just moved them to US or even Germany or something like that if they were serious about entering the GPU market, though...

e.g. There are plenty of talented engineers in China as well but it would be severely idiotic for any western company to move their core R&D there. Same applied to Russia.

[kgeist]

Well, Intel Russia opened in 2000 back when USA and Russia were on good terms, and Putin was relatively unknown. Sure it was a mistake in hindsight...

[Wytwwww]

> He failed on GPU. The product was substandard.

Weren't they pretty good (price/performance) after Intel fixed the drivers during the first year or so after release? The real failure was taking so long to ship the next gen..