What if the plan for the hyperscalers and cloud builders is not just to build their own stuff, but to keep you and OEM partners from building an alternative? That is what happens when the enterprise, government, and academia sectors get too small.

nextplatform.com/2024/04/17/wha…

Ampere Computing is cranking up the vector-enhanced cores and memory bandwidth on its Arm server CPUs to attack the AI inference opportunity in the datacenter and at the edge.
nextplatform.com/2024/04/16/amp…

Some HPC workloads are driven by the cheapest memory bandwidth per core – and this is especially true at Los Alamos National Laboratory, which has just fired up its Venado Grace-Hopper supercomputer.
nextplatform.com/2024/04/15/los… $NVDA $HPE

With MTIA v2, Meta Platforms is one step closer to having its own AI training chip – but not quite there yet.
nextplatform.com/2024/04/10/wit… $META $NVDA $AMD $INTC

With Google added to the pack, Arm server CPUs are going to literally give X86 chips from Intel and AMD a run for the money in the datacenter. And they are probably going to win a lot of races. $GOOG $INTC $AMD $NVDA $MSFT $AMZN $ARM
nextplatform.com/2024/04/09/goo…

With Gaudi 3, Intel has a real chance to sell some AI accelerators in a GenAI market where GPUs are still in short supply. $INTC $NVDA $AMD
nextplatform.com/2024/04/09/wit…

Intel has a plan to be the second biggest chip foundry in the world by 2030, and one that makes profits, too. But for now, Intel Foundry is bleeding red ink like a startup.
nextplatform.com/2024/04/04/int… $TSM

Microsoft and OpenAI are reportedly going to spend over $100 billion on a series of AI supercomputers culminating in a machine called “Stargate” in 2028, Here is what that program might look like.
nextplatform.com/2024/04/01/mic… $NVDA $AMD

Imagine a GPU chiplet compute complex with 24 or 32 banks of HBM stacked memory. Or even a CPU outrigged this way. Eliyan has a way to make one, and we think you will want one.
nextplatform.com/2024/03/28/how…

The ODMs have captured nearly 15 percent share of the datacenter Ethernet market. How long until they give Cisco Systems and Arista Networks a serious run for the money? $CSCO $ANET $HPE $NVDA
nextplatform.com/2024/03/25/how…

The “Ceiba” supercomputer being built by Nvidia at AWS for its own use is getting a Blackwell reboot, with close to $1 billion in GB200 CPU-GPU compute engines and NVSwitch interconnect.
nextplatform.com/2024/03/20/aws… - $NVDA $AMZN #HPC

If Nvidia is teaching the world anything with its Blackwell GPU platform, it is that memory and network are just as important as compute. You can’t skimp on any of this if you want to do AI.
nextplatform.com/2024/03/19/how…

So, what if there are no FP64 and FP32 CUDA cores on the 'Blackwell' datacenter GPU? Wouldn't that make it a TPU?

nextplatform.com/2024/03/18/wit…

If you want to take on Nvidia on its home turf of AI processing, then you had better bring more than your A game. Blackwell GPUs and their systems take AI scale to new heights.
nextplatform.com/2024/03/18/wit…

With its new WSE-3 compute engine, Cerebras can build a 256 exaflops beast that can handle 24 trillion parameters and probably costs on the order of $6 billion – which believe it or not, is a lot cheaper than a similar machine build from GPUs.
nextplatform.com/2024/03/14/cer… $NVDA $AMD

Meta Platforms gives us a glimpse into its ginormous GPU cluster buildout as it pursues artificial general intelligence.
nextplatform.com/2024/03/13/ins… $META $NVDA $ANET

A distributed computing system would seem to need a distributed operating system. Luminaries Michael Stonebraker and Matei Zaharia have created one called DBOS.
nextplatform.com/2024/03/12/the…

Researchers play Tetris with chiplet topologies to come up with the Goldilocks arrangement, called HexaMesh. Not to be confused with TexMex, which also comes with chips.
nextplatform.com/2024/03/11/hex…

Broadcom is gearing up for PCI-Express 6.0 fabrics to drive AI servers and other kinds of high performance clusters. $AVGO
nextplatform.com/2024/03/06/pus…