Saturday, February 10th 2024
NVIDIA to Create AI Semi-custom Chip Business Unit
NVIDIA is reportedly working to set up a new business unit focused on designing semi-custom chips for some of its largest data-center customers, Reuters reports. NVIDIA dominates the AI HPC processor market, although even its biggest customers are having to shop from its general lineup of A100 series and H100 series HPC processors. There are reports of some of these customers venturing out of the NVIDIA fold, wanting to develop their own AI processor designs. It is to cater to exactly this segment that NVIDIA is setting up the new unit.
A semi-custom chip isn't just a bespoke chip designed to a customer's specifications. It is co-developed by NVIDIA and its customer, using mainly NVIDIA IP blocks, but also integrating some third-party IP blocks the customer may want; and more importantly, approach semiconductor fabrication companies such as TSMC, Samsung, or Intel Foundry Services as separate entities from NVIDIA for their wafer allocation. For example, a company like Google may have a certain amount of wafer pre-allocation with TSMC (eg: for its Tensor SoCs powering the Pixel smartphones), which it may want to tap into for a semi-custom AI HPC processor for its cloud business. NVIDIA assesses a $30 billion TAM for this specific business unit—that's all its current customers wanting to pursue their own AI processor projects, who will now be motivated to stick to NVIDIA.
Source:
Reuters
A semi-custom chip isn't just a bespoke chip designed to a customer's specifications. It is co-developed by NVIDIA and its customer, using mainly NVIDIA IP blocks, but also integrating some third-party IP blocks the customer may want; and more importantly, approach semiconductor fabrication companies such as TSMC, Samsung, or Intel Foundry Services as separate entities from NVIDIA for their wafer allocation. For example, a company like Google may have a certain amount of wafer pre-allocation with TSMC (eg: for its Tensor SoCs powering the Pixel smartphones), which it may want to tap into for a semi-custom AI HPC processor for its cloud business. NVIDIA assesses a $30 billion TAM for this specific business unit—that's all its current customers wanting to pursue their own AI processor projects, who will now be motivated to stick to NVIDIA.
42 Comments on NVIDIA to Create AI Semi-custom Chip Business Unit
The original XBOX used nvidia (a modified geforce 3), the Playstation 3 used nvidia (a modified 7800), arcade hardware is still big in some areas of the world and back in the PS3 era the Taito Type X-2 (which games such as Street Fighter 4 were developed on and for) used a 7900 gs and later a 9800. In each of these cases nvidia burned the hardware makers pretty hard and all of them went AMD immediately after having dealt with nvidia.
While the details of what went down are obviously stuff the parties can't talk about what is known is it wasn't about price or profits. But something happened for them to burn the bridge with Xbox, Playstation, and Taito within short order and that none of those companies want to deal with them anymore. There's whatever happened with apple as well. Currently nvidia is in the Nintendo Switch and that seems to have worked as nvidia is going to be in the next Nintendo switch as well from what we currently know. The thing is that these are corporations and corporations like trust and stability. They'll pay more or make less to deal with a reliable partner where they don't have to worry about any headaches or distruptions cropping up. So far AMD has proved to be an extremely reliable partner and nvidia has proved to be a giant pain in the ass.
The kicker is that console makers don't need nvidia at all. On the other hand AI, enterprise, and professional users need nvidia. Gamers want nvidia. Of these three groups nvidia only needs AI, enterprise, and professional users which is the market AMD actually wants.
Nvidia's ace in the hole if it wants gaming is it's entire product stack for when PC gaming makes it's final big push to get everything up into the cloud.
But all this goes to show nvidia can be defeated. Of course the result of that is not cheaper nvidia, it's pushing them out of a market completely. And since PC gamers only want cheaper nvidia and don't care who it hurts to get that and aren't willing to just push them out of the market what PC gamers want and say has no impact at all. It's just children stomping their footsies. Where MS and Sony just flat out got rid of them.
The solution for cheaper stuff is to stop wanting advances in graphics, physics, and all of that so all the new advances could focus on making what can be done now cheaper. But the moment you want something better prices go back up again and no amount of screaming "negreedia!!!! muh g4m1ng PC!!!! Master Race!!!!" is going to change that.
Console makers aren't special snowflakes like PC gamers though. They go into this knowing they are going to take a loss on each unit produced and subsidize the cost to make it all back later through games, fees, and licensing. What they care about is a partner that they can deal with reliably, honestly, and consistently. The moment they get screwed on this they will tell another party to go get fucked and find a new supplier.
Custom chips can mean a lot of things as well. Nvidia isn't just GPUs.
The best "solution" for nvidia from a PC gaming standpoint would just be to ignore them. But that's not going to happen as you can see whenever AMD releases a card the first words out of peoples mouths are hoping this will cause nvidia to lower their prices. It's pathetic to the point of watching the fat ugly kid constantly pining after the hot chick who's off with the good looking guys having fun. Console makers aren't like that though. People buy consoles for the games on it and could give a rats ass who made the hardware or what's in it apart from teenage boys arguing on the internet over it. So all it takes is someone like nvidia driving the wrong people at Sony up the wall and causing them headaches and blamo nvidia is out and AMD is in. For an entire hardware generation, if not multiple generations. As if AMD delivers and the people at Sony are happy with it and still remember nvidia screwing them over the next time it comes to design something AMD is already in the door and everyone is still angry with nvidia. This is something Valve got early on. They went with a reliable and easy to deal with partner. Not only that Steam Deck OS is the killer advantage it has over all the competition not the hardware.
Console makers aren't dumb and neither is nvidia. There's no nvidia sticker on the switch because nobody cares. Mario, Zelda, Metroid, and allow Nintendo to sell hardware and thus stay in business no what's in the box. Slapping nvidia stickers on Xboxs and Playstations and talking about DLSS, RTX, and all that isn't going to move sales at all and will just generate eyerolls and "keep that PC bullshit out of here please" comments at best. So nvidia can scream all they want "we are the desired brand, just look at the PC gamers" and console makers can howl with laughter and say "our customers aren't that stupid and desperate" and that will be that.
While inflation has played a role, costs of 5nm wafers has played a role, having to put hi quality components to handle 300-600W has played a role, the main reason for prices going up is that now companies like AMD, Nvidia and Intel need to make enough income by selling fewer and at the same time more expensive cards. So they try to increase their profit margin from every card they sell. A $200 card will be sold for $300, a $600 for $1000, a $1000 card for $2000. The fact that they can start selling at a loss is a positive for Nvidia, if Nvidia can convince them that using Nvidia tech will mean much higher console sales, more income through "games, fees, and licensing".
Nvidia is going "FULL ARM". They waited and remained mostly quiet to get ARM, but others weren't fooled. Now Nvidia is coming out on the open, going full ARM. Just a speculation from my part. They might end up doing nothing. But $30 billions does look serious. It's not just pathetic, it's also boring. I see this mentality everywhere the last 20 years.
Forget SONY. I don't expect them to jump camp. Watch out for MS.
As for Valve and Steam OS, don't forget that Nvidia is also primarily a SOFTWARE company. Always was. MS also. Those two can build a very good UI for a gaming console.
Consoles like XBOX wouldn't get an Nvidia sticker either, but the reviews and comparisons with the PS6 could be extremely favorable if AMD does an RDNA3 mistake again with RDNA4. AMD made RDNA for gaming, somewhere after RDNA2 they decided to increase compute performance with RDNA3, instead of RayTracing performance for example and ended up losing a generation. I hope their RDNA4 will be an architecture focusing again in gaming, or they will definitely lose one of the consoles.
On Switch you would never see an Nvidia logo for two reasons. Nintendo would NEVER EVER agree to it, the Nvidia SOC is too slow, so even Nvidia wouldn't want it's logo on the Switch. If you have a quick look at tech press, they NEVER review pathetic hardware from Nvidia, either that is in the form of the GTX 1630 or the RX 3050 6GB. Nvidia probably doesn't want it and tech press plays along.
OpenAL understands (like IBM did with the original PC) that being beholden to a single-source manufacturer for the core of your business is a bad plan, mainly because Nvidia could hold the hardware for ransom and charge exorbitant amounts. Pretty much exactly what they're already doing.
TLDR: Well-funded AI companies will spend literal billions in R&D to not have to deal with Nvidia.
They'll be cheaper and they can get exactly what they want. No middle man.
Like Apple did with their CPUs.
Apple did not pull the M3 SoC out of a hat like some magic trick.
Remember that the first SoC with CPU cores designed in house was the A4 (2010). That means earlier iPhones used more standard Arm core designs. And yet it was still a modified version of the Cortex-A8 design. It would take years before Apple fully designed their own CPU cores.
Same with the graphics cores which for many years were Imagination PowerVR designs. On the GPU side, they transitioned gradually, introducing custom shader cores with the A8 (2014) but still retaining the PowerVR design. It wasn't until the A11 SoC (2017) that Apple ditched PowerVR.
So even a company with their own ASIC design team may take close to ten years to realize an original design from the moment of commitment.
Buttery smooth skin and ai rendering
we are so fucked
I agree with @john_, it's not unreasonable to expect improvements for GPUs.