Hacker News Reader

Comment by 2OEH8eoCRo0 | Hacker News Reader

I think we overestimate desktop GPU relevance. Are gaming GPUs really that lucrative?

the value proposition of Intel's graphics division wasn't in the current generation gaming GPUs, it was the growth of talent internally that could target higher and higher end chips at a much lower price than Nvidia until they were knocking on the door of the A100/H200-class chips - the chips that Nvidia produces for $2k and then sells for $40k.

Not to mention, Intel having vertical integration gave Intel flexibility, customization, and some cost saving advantages that Nvidia didn't have as much of, Nvidia being a fabless designer who are themselves a customer of another for-profit fab (TSMC).

If TFA is true, this was an anticompetitive move by Nvidia to preemptively decapitate their biggest competitor in 2030's datacenter GPU market.

No. It used to be more even between datacenter and gaming for NVIDIA, but that's not been the case for a few years. Gaming has brought in less money than networking (mellanox) since '24 Q4.

https://morethanmoore.substack.com/p/nvidia-2026-q2-financia...

But the same thing that makes GPUs powerful at rendering is what AI needs - modern gaming GPUs are basically supercomputers that provide Hw and Sw to do programmable embarrassingly parallel work. That is modern game rendering but also AI and crypto (and various science engineering) which is the second revolution that Intel completely missed (the first one being mobile).

I don't think anybody is using gaming GPUs to do serious AI at this point, though.

But you can use a gaming card to do AI and you can use H100 to game. The architecture between them is quite similar. And I expect upcoming edge AI applications to break down and end up using GPUs more than having dedicated AI accelerator HW because A) you need something to do display anyway B) the fixed function DSPs that have been called "AI accelerators" are worse than useless for running LLMs.

AI (apparently) needs much lower precision in training and certainly in inference than gaming requires though. A very very large part of the die on modern datacenter GPUs is effectively useless for gaming

I disagree that HW blocks for lower precision take up that much die space. Data center GPUs are useless for gaming because it's tuned that way. H100 still has 24 raster operating units (4050 has 32) and 456 texture mapping units (4090 has 512). That's because there's only so much they can tune the HW architecture to one use-case or the other without breaking some fundamental architecture assumptions. And consumer cards still come with tensor units and support for lower precision. This is because the HW costs and unit economics are such that it's much more in favor of a unified architecture that scales to different workloads vs discrete implementations specific to a given market segment.

They've also not bothered investing in SW to add the H100 to their consumer drivers to work well on games. That doesn't mean it's impossible and none of that takes away from the fact that H100 and consumer GPUs are much more similar and could theoretically be made to run the same workloads at comparable performance.

Depends if one cares about a PlayStation/XBox like experience, or Switch like.

If they weren't why would Nvidia keep making them? They do seem like an increasingly niche product, but apparently not enough that Nvidia is willing to just exit the market and focus on the datacenters.

They aren't just for gaming, there's also high-end workstations, but that's probably even more niche.

Have you seen the latest generation of Nvidia gaming cards? They are increasingly looking like an after thought.

I'm honestly curious why they keep making them. As far as I can tell, NVIDIA can sell literally as many datacenter AI chips as they can produce, and that would probably continue to be true even if they significantly increased prices. And even without increasing prices, the datacenter products are considerably higher margin than the consumer GPUs. Every consumer GPU they sell is lost revenue in comparison to using that fab capacity for a datacenter product.

The only reason I can imagine for them leaving the money on the table is that they think that the AI boom won't last that much longer and they don't want to kill their reputation in the consumer market. But even in that case, I'm not sure it really makes that much sense.

Maybe if consumer GPUs were literally just datacenter silicon that didn't make the grade or something, it would make sense but I don't think that's the case.

This is classic short term thinking. Whether or not AI is a bubble like the dot com bubble remains to be seen. But gamers have been buying Nvidia since before the dot com bubble and it is a market demand that has existed for a long time and will continue indefinitely. It doesn't make sense to cede this market to AMD.

I purposefully compare AI boom with the dot com bubble because we all knew how important the internet became eventually, but investments in it were way ahead of its time.

I pretty explicitly mentioned that possibility. I'm just a bit skeptical that, even if they completely abandoned the consumer GPU market, that they wouldn't be able to get back into it in 5 years or so when/if the bubble bursts. The longer they are out, the harder it would be to get back in with non-trivial market share (since the longer they are out the more their brand would have eroded), but also, the longer they are out, the more money they would have left on the table by staying in for so long.

They pretty much are doing the bare minimum in the gaming space already.

It is a false dichotomy. They can spend the bare minimum to stay in the game card market while fabing AI cards. At this point that is just an insurance premium.

How would one learn to be a marketable AI dev/researcher if not playing with the ecosystem/tooling on a consumer hardware? If nobody is exploring AI at home, the influx of fresh minds ceases, the development of the field slows down or stops entirely, market gets disillusioned and the field eventually disappears.

They're a market entry point. CUDA became popular not because it was good, but because it was accessible. If you need to spend $10k minimum on hardware just to test the waters of what you're trying to do, that's a lot to think about, and possibly tons of paperwork if it's not your money. But if you can test it on $300 hardware that you probably already own anyway...

they kept nvidia in business for a long time until their datacenter breakthrough.

Gaming GPUs make up 7% of Nvidia's business, 93% is datacentre. So, no.

Intel has always pursued agglomeration into the main CPU. They sucked up the math co-processor. They sucked up the frontside bus logic. They sucked up the DDR controllers more and more. They have sucked in integrated graphics.

Everything on-die, and with chiplets in-package, is the Intel way.

Default, average integrated graphics will continue to "statisfice" for a greater and greater portion of the market with integrated graphics continuing to grow in power.

Intel made fun of AMD for "taping chips together". Intel did everything on a monolithic die for about way too long.

The smaller the node the smaller the yield, chiplets is a necessity now (or architectural changes like Cerebras).

Running tests and then fusing off broken cores or shared caches helps to recover lots of yield for bigger chips. Certain parts of the silicon is not redundant, but Intel's designs have redundancy for core pieces and chunks that are very large and hence probabilistically more prone to a manufacturing error.

Yep, cerebras takes that thing to the next level with their "wafer chips". A common technique is killing defective cores entirely (how all cheaper CPUs are made).

But reducing size will still increase yield since you can pick and choose.

Which is ironic, given that Intel had to glue two Pentium D dies together to compete with the monolithic Athlon 64 X2: https://en.wikipedia.org/wiki/Pentium_D#Presler

Reply on news.ycombinator.com