Most companies aren't Google or Tesla though.
And like I said, it's likely that different solutions will appeal to different use cases. I bet that aside from those custom chips, Google and Tesla are also using GPU's for some things.
Yeah, you need to be doing enough of the work to justify your own design team and fabrication costs. Most companies aren't on that level or doing that kind of specialized work.
I'm not surprised to see such companies going their own way though. NVidia has managed to alienate just about every major partner they've ever done business with and some of those people have no doubt gone on to work with other companies and have steered them away.
But they don^t need a mainstream dGPU for that. ponte vecchio is so different from alchemist, I doubt selling alchemist helps much to distribute R&D costs.It's not just fab filler, they presumably want to enter the same professional markets that Nvidia is dominating.
Nvidia sells the same basic architecture to both consumers and professionals.But they don^t need a mainstream dGPU for that. ponte vecchio is so different from alchemist, I doubt selling alchemist helps much to distribute R&D costs.
Nvidia sells the same basic architecture to both consumers and professionals.
That last sentence already means that I'm right, but for their data center GPU's, they sold Ampere and Turing cards as well. So it was not just the same basic architecture, but the exact same chips (although still less gimped on the data center GPUs).Not really. The AI and HPC focused chips like the V100 and A100 are a different architecture from the consumer chips. Only the Quadros are the same.
AI chips are relatively simple. Much easier than designing a full GPU or even full CPU cores.
You still would need a few dozen engineers, which is going to be a few million dollars in costs, possibly into the tens of millions depending on the size of your team.
Then you need a set of masks for the chips, which can again be millions of dollars, and after that the wafers to make the chips and someone to handle the packaging. At least that cost depends on volume to a degree.
The problem is that if you realistically only need something equivalent to 100 of NVidia's H100 GPUs, you'd have to be able to design and deliver your own for less than $3.5 million dollars before it makes financial sense.
Most companies just don't have those needs. They will certainly find another company with a less expensive solution, but they're unlikely to roll their own. You probably need to be requiring at least another order of magnitude of computational power before you even start to consider switching.
Yes, apart from all those architectural differences they're totally the same! XDThat last sentence already means that I'm right, but for their data center GPU's, they sold Ampere and Turing cards as well. So it was not just the same basic architecture, but the exact same chips (although still less gimped on the data center GPUs).
Hopper does have different chips, but they are clearly fairly similar in basic architecture and such. Although they gimped Ada in a variety of ways (CUDA 8.9 instead of 9, PCIe 4 vs 5, no NVLink, far fewer NVENCs, GDDR vs HBM, etc). But none of that makes the architecture fundamentally different.
It's not been the exact same chip since the P100 all the way back in 2016. V100 was the first one to actually be different on an architectural level though.@NTMBK
Except that for the previous two generation it was the exact same chip. And most of the changes for Hopper are either just some upgrades to the API, which clearly is an evolution, not a revolution, or just features tacked on there. Again, I never claimed that it was the same chip for this generation, just that it is the same basic architecture. Not that I never said that it is the exact same architecture.
It's like taking a car with a V4 and putting a V6 in there with a super charger. And then adding launch mode to the computer. Plus adding a tail hook. That doesn't change the basic architecture of the engine, or of the car, like what happens when you create a self-driving electric car or such.
Yes, apart from all those architectural differences they're totally the same! XD
It's not like the old days where a Tesla card was just a fully enabled Geforce 580. It's different chips, with differently balanced resources, different instruction sets, different memory systems, different interconnects etc. Yes there is a lot of commonality, but it's not the same architecture.
Maybe the A770 8GB and lower tier cards. The A770 16GB should maintain its value relative to even the 4060 Ti 8GB when that is released. I wouldn't entertain anything less than 16GB in the market right now. Well, maybe the 4070 but my mind keeps kicking the thought out due to "only" 12GB VRAM.Otherwise they will have to basically give them away.
It's not that simple. There are older games that still have bad performance, crash, have graphical issues, or fail to run on ARC. It can't match the competition for recording and streaming as of yet. The Intel version is a nightmare for servicing down the road when it needs new paste. There is uncertainty concerning the products future.The A770 16GB should maintain its value relative to even the 4060 Ti 8GB when that is released.
It sure would be interesting to know how many cards they built. If you saw no end to the mining craze and wanted to capitalize on the higher ASP of graphics cards before the mining proposition went bust, how many would you make, considering you have the sort of money that Intel has?If you can't get it all back, at least get a big install base, some great publicity, and inspire some warm feels that hopefully carry over to Gandalf.
I have no speculation to share about their strategy for Gandalf. I do recall reading they planned to ship over 4 million ARC. That included mobile too I think. That probably doesn't count ASRock and Acer. Getting millions of PC gamers using ARC before Gandalf gets here would be good juju IMO.It sure would be interesting to know how many cards they built. If you saw no end to the mining craze and wanted to capitalize on the higher ASP of graphics cards before the mining proposition went bust, how many would you make, considering you have the sort of money that Intel has?
Their mistake was underestimating the driver work and arriving at least 6 months too late, allowing lowered prices of AMD/Nvidia cards to put a dent in Intel's Grand GPU loot ambitions. I think our wish for an ARC firesale will come true, if and when Battlemage hits the market. Then the conundrum will be, do they let ARC 16GB be their new low end and only release two Battlemage SKUs targeting the high end? Or do they wait a while, then release low end Battlemage that offers better performance than ARC A770, to entice the A770 gamers to upgrade again and throw more money into Intel's coffers? I would hate to be their product strategist coz I'm already getting a headache.
They really need some worthy word-of-,mouth pr from "gamers in the streets". If they have any long-term hope of surviving for the longer-term in the GPU market.Getting millions of PC gamers using ARC before Gandalf gets here would be good juju IMO.
Most of us were saying to avoid ARC unless you like to troubleshoot and be challenged simply to use hardware normally. For n00bs hell no.They really need some worthy word-of-,mouth pr from "gamers in the streets". If they have any long-term hope of surviving for the longer-term in the GPU market.
I would say that at least with NVidia cards, MANY of them were purchased by gamers, because of word-of-mouth of OTHER gamers, not just because of flashy marketing and PR. (Of course, that can help too, especially for 1st-time buyers.)
My only point is the simplicity compared to GPU/CPU design making this business more open to disruption. Could you imagine Tesla Designing it's own GPU/CPU cores? Or a bunch of startups getting funded for GPU design? Or Intel who struggles at the back of the GPU pack, also has an A100 beating DL chip.
Anyone with scale or even a business case to sell to others who have scale can do their own AI chips.
Most of us were saying to avoid ARC unless you like to troubleshoot and be challenged simply to use hardware normally. For n00bs hell no.
Now, it is good enough overall. I'd still say no for n00bs. The pace Intel is on now, that will be okay by the holidays as a recommendation.
You mean, like the RX 6700 non-XT? 10GB GDDR6, and $320 minus many different promo codes at time.A $300 GPU with at least 10 GB of VRAM could clean house right now.