231–240 of 722 posts
How does this connect to Gelsinger's retirement, announced yesterday? The comments on that news were all doom and gloom, so I had expected more negative news today. Not a product launch. But I'm just some guy on HN, what do I know?
Earlier quoted context omitted.
I think we don't really disagree, I just think that this shape isn't really a gpu its just a cpu because it isn't very good for graphics at that point.
That's why I said "basic GPU". It doesn't have to be too fast but it should still be way faster than a regular CPU. Intel already has Xeon Phi so a lot of things were developed already (like memory controller, heavy parallel dies etc.)
Earlier quoted context omitted.
M3/M4 Max MacBooks with 128GB RAM are already way better than an A6000 for very large local LLMs. So even if the GPU is as slow as the one in M3/M4 Max (<3070), and using some basic RAM like LPDDR5x it would still be way faster than anything from NVidia.
The M4 Max needs an enormous 512bit memory bus to extract enough bandwidth out of those LPDDR5x chips, while the GPUs that Intel just launched are 192/160bit and even flagships rarely exceed 384bit. They can't just slap more memory on the board, they would need to dedicate significantly more silicon area to memory IO and drive up the cost of the part, assuming their architecture would even scale that wide without hit…
We'd be happy to pay 5000 for 128gb from Intel.
Earlier quoted context omitted.
Just how "basic" do you think a GPU can be while having the capability to interface with that much DRAM? Getting there with GDDR6 would require a really wide memory bus even if you could get it to operate with multiple ranks. Getting to 128GB with LPDDR5x would be possible with the 256-bit bus width they used on the top parts of the last generation, but would result in having half the bandwidth of an already mediocre…
What if they put 8 identical GPUs in the package, each with 1/8 the memory? Would that be a useful configuration for a modern LLM?
Earlier quoted context omitted.
HBM3E memory is at least 3x the price of DDR5 (it requires 3x the wafer as DDR5) and capacity is sold out for all of 2025 already... that's the price and production bottleneck. High speed, low latency server grade DDR5 is around $800-$1600 for 128GB. Triple that for $2400 - $4800 just for the memory. Still need the GPUs/APUs, card, VRMs, etc. Even the nVidia H100 with "only" 94GB starts at $30k...
Nvidia's $30,000 is a 90% margin product at scale. They could charge 1/3 that and still be very profitable. There has rarely been such a profitable large corporation in terms of the combo of profit & margin. Their last quarter was $35b in sales and $26b in gross profit ($21.8b op income; 62% op income margin vs sales). Visa is notorious for their extreme margin (66% op income margin vs sales) due to being basically a…
Too late, and it has a bad rep. This effort from Intel to sell discrete GPUs is just inertia from old aspirations, won't really help noticeably to save it, as there is not much money in it. Most probably the whole Intel ARC effort will be mothballed, and probably many more will.
No reviews and when you click on the reseller links in the press announcement they're still selling A750s with no B-Series in sight. Strong paper launch.
Why don't they just release a basic GPU with 128GB RAM and eat NVidia's local generative AI lunch? The networking effect of all devs porting their LLMs etc. to that card would instantly put them as a major CUDA threat. But beancounters running the company would never get such an idea...
The deeper problem is that the market for this is probably incredibly niche.
Earlier quoted context omitted.
give me 48gb with reasonable power consumption so I can dev locally and I will buy it in a heartbeat. Anyone that is fine-tuning would want a setup like that to test things before pushing to real GPUs. And in reality if you can fine-tune on a card like that in two days instead of a few hours it would totally be worth it.
I would love too, but you can't just add the chips, you need the the bus too.
Earlier quoted context omitted.
They are probably held back by same reason thats preventing AMD and nVidia from doing it either.
The reason is AMD and Nvidia don't is that they don't want to cannibalize their high end AI market. Intel doesn't have a high end AI market to protect.
As far as I understand it, it gives you 64 GiB of HBM per socket.
Earlier quoted context omitted.
100% - this could be Intel's ticket to capture the hearts of developers and then everything else that flows downstream. They have nothing to lose here -- just do it Intel!
They could lose a lot of money?
At some point they should make a stand, that's the whole meta-topic of this thread.