AMD bought MEXT, a startup whose software makes cheap NAND flash act like DRAM, so a server runs on far less of the expensive stuff. With DRAM contract prices up ~90%+ this year, the logic isn't subtle. How predictive memory tiering actually works:
https://www.buysellram.com/blog/amd-mext-flash-as-dram/

#AMD #MEXT #DRAM #NANDFlash #MemoryTiering #AIInfrastructure #DataCenter #ServerHardware #SSD #ITAD #Semiconductors #HBM

AMD just spent money to make memory cheaper, not faster. Its acquisition of MEXT, a small software company, is a bet on a simple idea: most of the DRAM in a server sits idle, and cheap NAND flash can stand in for it if the software is smart enough.

MEXT's Predictive Memory Engine watches how an application uses memory, predicts which pages are about to be needed, and moves them from flash up to DRAM before the program asks. Hot data stays fast; cold data lives on flash that costs a fraction as much. The application never notices.

Conventional DRAM contract prices jumped roughly 93-98% in a single quarter as makers shifted capacity to HBM for AI accelerators. When DRAM is this scarce, stretching it with flash stops being a niche trick.

One caveat: flash being ~50x cheaper per bit doesn't mean a 50x cut to a memory bill. The realistic payoff is closer to half the DRAM cost. Still, the direction is clear, and it changes what aging server memory and SSDs are worth on the way out.

https://www.buysellram.com/blog/amd-mext-flash-as-dram/
#AMD #MEXT #DRAM #NANDFlash #MemoryTiering #AIInfrastructure #DataCenter #ServerHardware #SSD #ITAD #Semiconductors #HBM #tech

At Computex 2026, NVIDIA's RTX Spark put a 20-core Arm Grace CPU and a Blackwell GPU inside mainstream Windows laptops from Dell, HP, Lenovo, Asus, MSI and Microsoft. The same week, Arm's first-ever data-center processor — the AGI CPU, with 136 Neoverse cores — headed into mass production, with Meta, OpenAI and Oracle Cloud already on board.
But here's the part most coverage skips: Arm vs. x86 isn't a "who's better" contest. Architecture doesn't decide performance — the software ecosystem and the workload do. x86 isn't going anywhere soon; it still anchors the enterprise.
What is changing is the mix. As hyperscalers shift new AI capacity to Arm, recent-generation x86 servers get parted out — flooding the secondary market with enterprise RAM, SSDs, CPUs and GPUs.

https://www.buysellram.com/blog/from-rtx-spark-to-agi-cpu-arms-two-front-challenge-to-x86/

#Arm #x86 #DataCenter #RTXSpark #NVIDIA #CPU #ITAD #ServerHardware #AGICPU #WindowsOnArm #AIInfrastructure #ITAssetDisposition #technology

Arm challenged it on two fronts at once. At Computex 2026, NVIDIA's RTX Spark put a 20-core Arm CPU + Blackwell GPU inside mainstream Windows laptops, while Arm's first data-center chip — the AGI CPU — hit mass production with Meta, OpenAI and Oracle aboard. So what happens to x86 servers and parts out there now?

https://www.buysellram.com/blog/from-rtx-spark-to-agi-cpu-arms-two-front-challenge-to-x86/

#Arm #x86 #DataCenter #RTXSpark #NVIDIA #CPU #ITAD #ServerHardware #AGICPU #WindowsOnArm #AIInfrastructure #tech

At Computex 2026, NVIDIA's RTX Spark put a 20-core Arm Grace CPU and a Blackwell GPU inside mainstream Windows laptops from Dell, HP, Lenovo, Asus, MSI and Microsoft. The same week, Arm's first-ever data-center processor — the AGI CPU, with 136 Neoverse cores — headed into mass production, with Meta, OpenAI and Oracle Cloud already on board.
But here's the part most coverage skips: Arm vs. x86 isn't a "who's better" contest. Architecture doesn't decide performance — the software ecosystem and the workload do. x86 isn't going anywhere soon; it still anchors the enterprise.
What is changing is the mix. As hyperscalers shift new AI capacity to Arm, recent-generation x86 servers get parted out — flooding the secondary market with enterprise RAM, SSDs, CPUs and GPUs.

https://www.buysellram.com/blog/from-rtx-spark-to-agi-cpu-arms-two-front-challenge-to-x86/

#Arm #x86 #DataCenter #RTXSpark #NVIDIA #CPU #ITAD #ServerHardware #AGICPU #WindowsOnArm #AIInfrastructure #ITAssetDisposition #technology

Arm challenged it on two fronts at once. At Computex 2026, NVIDIA's RTX Spark put a 20-core Arm CPU + Blackwell GPU inside mainstream Windows laptops, while Arm's first data-center chip — the AGI CPU — hit mass production with Meta, OpenAI and Oracle aboard. So what happens to x86 servers and parts out there now?

https://www.buysellram.com/blog/from-rtx-spark-to-agi-cpu-arms-two-front-challenge-to-x86/

#Arm #x86 #DataCenter #RTXSpark #NVIDIA #CPU #ITAD #ServerHardware #AGICPU #WindowsOnArm #AIInfrastructure #tech

To me, #NVIDIA plan has been always been vertical ingration.

They don't want to have the best GPU for AI, they want to offer the full stack: GPU + CPU + Software.

They got the GPU part and the Software (CUDA) part ready. Their CPU is the shaky part, and how much they can push ARM is the key to their success.

Eventually they'll sell complete GPU+CPU servers, leaving out AMD and Intel for scraps.

#Technology #AI #GPU #CPU #Hardware #Servers #ServerHardware #AMD #Intel #CUDA #ARM

RT @digitalix: Neue "GPU" mit 144 GB HBM – AMD Instinct MI350P, ABER sie nutzt PCIe!

mehr auf Arint.info

#AMD #GPU #HBM #InstinctMI350P #PCIe #ServerHardware #arint_info

https://x.com/digitalix/status/2052414183963086957#m