AMD, Broadcom and Google Intensify Anti-Nvidia Offensive as AI Semiconductor Landscape Faces Potential Realignment
Posted by sr_local@reddit | hardware | View on Reddit | 21 comments
Dense-Elephant5048@reddit
NVIDIA has a 63% net profit margin and a near monopoly in many products. Make sense that strong players need find a way to curb the finantial hemorrage that only favours a single supplier. AMD, being a competitor, needs to increase its market share.
elkond@reddit
nvidia has near monopoly because they invested in CUDA since forever. amd has ROCm that has a hardware support with less shelf life than my cat's food in open sun
Dense-Elephant5048@reddit
CUDA can get the GPU to excel in games, cryptos, AI and what not. Slowly but surely AMD will catch up.
elkond@reddit
that's why companies like roblox came to intel to migrate to oneapi because rocm is "catching up"?
Dense-Elephant5048@reddit
KIV...
iluvchromosomes@reddit
Those "strong players" intentionally gave up the way, years ago, for short term profits.
This is what happens when everyone outsources their chips to TSMC and doesn't invest in their own infrastructure.
Intel might be spinning up new fabs. But they suck. And TSMC is still making all of Intel's current best chips (panther lake).
Out of all the deplorable capitalists outsourcing to TSMC, Nvidia had the most desirable chips, and now can pay the most to get priority from TSMC.
It's a shitty game and all the other players deserve to be in their current situation.
RHINO_Mk_II@reddit
Because Nvidia is fabbing their own chips? Hmm.
techraito@reddit
Yea, this almost reads like a skill-issue from other companies lol. I don't like monopolies, but competitors just aren't competing at the same level.
Google's own Tensor gotta be a waste of TSMC fabs :/
cotdt@reddit
Amd needs to make CPU that has integrated GPU and can run ai models locally
kwirky88@reddit
That's unecessary for cloud llm providers
pearlgreymusic@reddit
Isn’t that what Strix Halo is?
Awkward-Candle-4977@reddit
Llama 8b int4 runs at 15 tps in my 7840u igpu
pcgameshardware@reddit
"Anti-Nvidia offensive" sounds a bit over dramatic. This looks more like hyperscalers trying to avoid being locked into one supplier. Nvidia still has the strongest ecosystem, but AMD, Broadcom and Google make sense as pressure points in very different parts of the market...
- Jacky
Allfeelings0Logic@reddit
You'd think Amazon and friends would join this trio
PastaPandaSimon@reddit
Once you have a model family, and you're huge, you gain massively from creating your own environment and hardware, and paying Nvidia inevitably becomes an unnecessarily high cost, which becomes competitive disadvantage.
Awkward-Candle-4977@reddit
Google, aws and azure will use their in house ai chips more and more.
I did rough calculation on google tpu. It's at least 80 percent cheaper for google than buying nvidia chips
hanshotfirst-42@reddit
Does this mean we’ll get GPUs equal to a RTX-4060 in 2035 instead of 2040?
RealPjotr@reddit
AMD skipped a generation, the plan was for RDNA5 high end again. But with AI, one has to wonder if we will see any new high end consumer boards from either manufacturer this decade. 🤷🏼♂️
j_osb@reddit
I mean for compute, chiplets have scaled very well. MI300X and MI350/355X are chiplet based and incredible value compared to NVIDIA solutions. The only thing they’re lacking in right now compared to NVIDIA is interconnect and training, which they want to change with the Helios rack.
Luckily, as RDNA5 or 6 is supposed to be UDNA, we can actually expect higher end models; as they’ll want to have rocm occupy the same niche as CUDA (currently, rocm between RDNA and CDNA is fragmented, which makes development for it harder) in terms of accessibility, so they’ll want to drop a deal of powerful personal cards as well.
Mental-At-ThirtyFive@reddit
imho chiplet design for ai gpus is necessary to accelerate parallel compute even more and AMD has the compute-cache experience to make it happen. wait and see. AMD should prioritize their igpus for edge ai
ryanvsrobots@reddit
This has nothing to do with consumer GPUs