Samsung to more than double HBM4 and HBM4E output next year, lifting glass carrier cleaning volume to 50,000 sheets a month
- Samsung will raise outsourced glass carrier cleaning volume to 50,000 sheets a month next year from 20,000 this year, a 2.5-fold jump that industry analysts read as HBM4 and HBM4E output at least doubling.
- Samsung's overall HBM wafer input is projected to grow nearly 40% to roughly 250,000 wafers a month next year from about 180,000, with the HBM4 family climbing from around 40% of shipments to about 80%.
- Samsung started mass production shipments of HBM4 in February using 10-nanometer-class 1c DRAM and a 4-nanometer base die, and supplied 12-layer HBM4E samples to customers including Nvidia in May.
- A glass carrier is a support temporarily attached to the underside of an HBM DRAM wafer to prevent bending or cracking while the wafer is ground thin and drilled, and demand ran at 10,000 sheets a month last year before doubling this year.
- The HBM4 and HBM4E parts Samsung is scaling up center on 12-layer and higher stacks, where wafer thinning and warpage control matter more as stack counts rise.
Hacker News opinions
Shame that this probably just means consumer DRAM prices get even worse.
Why would more RAM supply push consumer prices up? More supply is supposed to mean cheaper.
Is that going to be enough for AI's appetite?
Not even close.
Samsung's already building another fab, the one that got suspended years ago when memory prices were in the toilet. Hope that when it comes online in a few years it isn't all HBM for AI.
It'll go live right around when AI runs fine on consumer hardware and the data center buildout collapses.
That collapse only happens after the industry gets severe indigestion, and that doesn't look far off.
Remember Covid: Intel announced new fabs to fix the chip shortage, and the shortage was over before they broke ground. The projects got canceled.
Bought some DDR4 or DDR5 two years ago, checked the price a few months back and it was over 3x. That's insane. Governments need to do something about this system AI amplified.
What abuse? Responding to supply and demand isn't abuse, it's the market working properly.
Which government, and do what exactly? The demand is global. The US can't conjure world class fabs overnight, Korea and Taiwan don't have the demand to justify the investment on their own, and China doesn't have the IP yet. Nobody holds all the levers.
3x? I paid 300 eur for 128GB DDR5 SODIMMs. Now it's 4000 eur.
Ignoring the price, what actually blocks HBM from being the main memory in consumer electronics?
Nothing and everything. It's DRAM, but the bandwidth needs mean it's paired to the processor, so no DIMMs. On a MacBook or the Framework tower where RAM is fixed for the device's life, there's basically zero tradeoff.
There's no blocker so much as a cost: an HBM package eats about 3x the manufacturing capacity for the same storage as regular DRAM. You're trading total bytes for bandwidth.
HBM is built to sit in the same package as the CPU, so you lose DIMM sockets entirely. Latency is higher too.
Why would you want it on a desktop? The whole advantage is bandwidth, and that needs a 2048-bit bus against the 128 bits of a normal CPU, plus a processor that can eat 2 to 3 TB per second. Only a GPU wants that.
Now go make some DDR5.
HBM is DDR5, just in a different package.
Is this the start of the bubble popping?
It's RAM for exactly the datacenter systems AI needs, so no?
How do you even get that from this headline and article?
Related thing I read yesterday: the real bottleneck on Chinese AI accelerators is HBM, not processors or ASML equipment. Huawei's Ascend volume is limited by CXMT's HBM capacity. HBM needs die thinning, via drilling, plating, alignment, and that gear is sanctioned for everyone else too, so the Chinese built it themselves with low yields, even on shorter stacks.