A new report suggests that all three memory manufacturers — Samsung, SK Hynix, and Micron — have collectively sold through their 2027 memory manufacturing capacity to AI companies.
HBM and Vera Rubin are not useful for PC Gaming, at least not for current architectures. Main issue is that DDR5 production simply got cut to make more HBM. But there will be a huge amount of LPDDR, NVME / SSD drives and refurbished Hard Drives available, and hardware for running inference at home for pretty powerful models will become cheap as fuck, liberating us all from the hyperscalers; the open weights models from china are available for everyone.
I’ve never heard of uncut wavers hitting storage or the market. These things are made to be used for whatever is needed at the moment. If you hear this news regarding “Ram cap sold out”, it means all available wavers they expect to be produced for that year have been allocated to whoever paid for them. If 50% of the wavers go into HBM, 30% into DDR including LPDDR and 20% go into NAND storage (i do not have the exact numbers, but i know that HBM needs around 3x the amount of silicon compared to DDR), then you are done selling your capacity - whatever happens afterwards does not make DDR5 appear magically on the market.
Only thing that could change that would be that the AI bubble crashes before 2027, then the allocation numbers will be changed.
Not sure what you mean with the wafers hitting the market. I was only referring to the specific deal that OpenAI made with Samsung and SK hynix to purchase a metric boatload of 900,000 undiced DRAM silicon wafers per month.
If the memory wafers are largely HBM, so they can’t be used for gaming hardware.
If you’re thinking of HBM gaming GPUs, the chip design pipeline is too long. If Nvidia/AMD/Intel started making one right now, and rushed it, it would still be many years away.
…However, a lot of LPDDR5 is being made too.
Hence the rumors of AMD hedging their bets and using LPDDR for some future GPUs.
I’m not sure I worded this right before, but it takes ~half a decade to plan and produce a GPU.
They can’t just switch out the bus or memory on an existing design.
So they could make a HBM gaming GPU. But they would have had to have started ~2 years ago.
…That’s extremely unlikely.
It’s unlikely they’ll even plan one now.
AMD did indeed try making HBM gaming GPUs, and it turns out it’s supremely uneconomical; they immediately pivoted to the opposite extreme (narrow bus + larger cache).
However, LPDDR GPUs appear to already be in the pipe, so that will provide some GPU market relief.
HBM and Vera Rubin are not useful for PC Gaming, at least not for current architectures. Main issue is that DDR5 production simply got cut to make more HBM. But there will be a huge amount of LPDDR, NVME / SSD drives and refurbished Hard Drives available, and hardware for running inference at home for pretty powerful models will become cheap as fuck, liberating us all from the hyperscalers; the open weights models from china are available for everyone.
At least i pray it will be this way.
But aren’t the wafers uncut? It could be used for anything if they’re just in storage.
I’ve never heard of uncut wavers hitting storage or the market. These things are made to be used for whatever is needed at the moment. If you hear this news regarding “Ram cap sold out”, it means all available wavers they expect to be produced for that year have been allocated to whoever paid for them. If 50% of the wavers go into HBM, 30% into DDR including LPDDR and 20% go into NAND storage (i do not have the exact numbers, but i know that HBM needs around 3x the amount of silicon compared to DDR), then you are done selling your capacity - whatever happens afterwards does not make DDR5 appear magically on the market.
Only thing that could change that would be that the AI bubble crashes before 2027, then the allocation numbers will be changed.
Not sure what you mean with the wafers hitting the market. I was only referring to the specific deal that OpenAI made with Samsung and SK hynix to purchase a metric boatload of 900,000 undiced DRAM silicon wafers per month.
According to Tom’s Hardware it includes “DDR5 and LPDDR4/LPDDR5 as well as premium HBM.”
Honestly I’m not sufficiently familiar with how that works to understand if any of that is consumer grade stuff.
If the memory wafers are largely HBM, so they can’t be used for gaming hardware.
If you’re thinking of HBM gaming GPUs, the chip design pipeline is too long. If Nvidia/AMD/Intel started making one right now, and rushed it, it would still be many years away.
…However, a lot of LPDDR5 is being made too.
Hence the rumors of AMD hedging their bets and using LPDDR for some future GPUs.
According to Tom’s Hardware, it includes “DDR5 and LPDDR4/LPDDR5 as well as premium HBM.”
Could gpus use HBM? I thought AMD used it 10 years ago in theirs.
I’m not sure I worded this right before, but it takes ~half a decade to plan and produce a GPU.
They can’t just switch out the bus or memory on an existing design.
So they could make a HBM gaming GPU. But they would have had to have started ~2 years ago.
…That’s extremely unlikely.
It’s unlikely they’ll even plan one now.
AMD did indeed try making HBM gaming GPUs, and it turns out it’s supremely uneconomical; they immediately pivoted to the opposite extreme (narrow bus + larger cache).
However, LPDDR GPUs appear to already be in the pipe, so that will provide some GPU market relief.