Comments on: Meta Platforms Crafts Homegrown AI Inference Chip, AI Training Next https://www.nextplatform.com/2023/05/18/meta-platforms-crafts-homegrown-ai-inference-chip-ai-training-next/ In-depth coverage of high-end computing at large enterprises, supercomputing centers, hyperscale data centers, and public clouds. Mon, 29 May 2023 04:29:04 +0000 hourly 1 https://wordpress.org/?v=6.7.1 By: skierpage https://www.nextplatform.com/2023/05/18/meta-platforms-crafts-homegrown-ai-inference-chip-ai-training-next/#comment-209237 Sun, 28 May 2023 01:33:02 +0000 https://www.nextplatform.com/?p=142405#comment-209237 @Timothy,
“Which begs the question: …”
No it doesn’t, it RAISES the question. “Begging the question” is a rare logical fallacy. The phrase seems to give your question more rhetorical heft, but it’s wrong.

It’s crazy that keeping us scrolling through an endless feed of divisive inflammatory posts can pay for tens (hundreds?) of millions in custom chip design and fabrication to optimize Deep Learning Recommendation Models. Just show me what my friend have been doing!

]]> By: amanfromMars https://www.nextplatform.com/2023/05/18/meta-platforms-crafts-homegrown-ai-inference-chip-ai-training-next/#comment-208892 Sun, 21 May 2023 12:49:55 +0000 https://www.nextplatform.com/?p=142405#comment-208892 Do you think the likes of a western Apple, or an oriental Baidu or Huawei are, for some time ago now, pioneering with ethereal champions expert and experienced in their respective novel fields leading AI Interference with AI Inference Training for Chipped Platforms in Linked Computerised Networks ……. in their own inimitable stealthy unilateral way, failsafe secure and proprietary intellectual property protected within their sealed gardens and almighty fortresses of virtual delights …… and making ready to lay immaculate waste to any and all wilfully negative and punitive competition or similarly disposed opposition with a vast and comprehensive series of controlled releases for subsequent universal deployment/employment/enjoyment?

And you surely do realise if you think not, the likes of an Apple or a Baidu or a Huawei streak ever further away, and way out ahead in front, with nothing to hinder or stop progress leading the Internetworking of Things with their way of doing all things.

Do you realise that is where you and they, and IT and AI, are presently at?

]]> By: Mark Sobkow https://www.nextplatform.com/2023/05/18/meta-platforms-crafts-homegrown-ai-inference-chip-ai-training-next/#comment-208872 Sun, 21 May 2023 06:12:22 +0000 https://www.nextplatform.com/?p=142405#comment-208872 I think a lot of people are going to be rather surprised at where RISC-V pops up and how quickly it evolves to compete head to head with “top end” processors from the Arm and x64 incumbents. From a programmer’s perspective (outside the device driver and compiler writer realms), programmers don’t really care all that much about the details of the CPU instruction sets nowadays. Anyone who thinks they do is invited to compare code base sizes between the 80s and the 2020s – modern code is not written to the metal any more.

]]> By: EC https://www.nextplatform.com/2023/05/18/meta-platforms-crafts-homegrown-ai-inference-chip-ai-training-next/#comment-208796 Fri, 19 May 2023 16:58:30 +0000 https://www.nextplatform.com/?p=142405#comment-208796 >>NNP And GPU Can’t Handle The Load With Good TCO<< TCO = TOTAL Cost of Ownersnip. And to that I'd wonder if the green shaded Meta bean counters have factored in development, deployment, software, support and maintenance costs into their counting machines? Core business semiconductor guys spread these costs over thousands of customers while Meta has one. With late generation devices reaching into the $Bs to properly develop and deploy, I'm not sure the math pencils out.

]]> By: HuMo https://www.nextplatform.com/2023/05/18/meta-platforms-crafts-homegrown-ai-inference-chip-ai-training-next/#comment-208758 Fri, 19 May 2023 01:37:59 +0000 https://www.nextplatform.com/?p=142405#comment-208758 I’ve been wondering about proper use cases for RISC-V for a while (aisde from quadfurcated pet-rock droop engines), and this might be one, maybe. ARM and POWER seem much closer to the sweet spot between x86/64-CISCs and very simple RISCs, for generic workloads, in terms of the balance between performance and efficiency. Preemptive multitasking, virtual machines, and automated garbage collection (gc), as found in Java programs running on contemporary OSes (for example), require some solid CPU oomph, like that provided by ARM/POWER and x86/64. But, single-tasking machine code, or even C, running on bare metal, with manual memory allocation, is probably fine performance-wise on RISC-V. Patterson’s perspective (in letting go of SOAR and SPUR in favor of RISC-V) may have been “exactly” this, that a computer would have a sufficiently large number of simple cores (and spare ones) to run each (simultaneous) process on a separate physical core, without any need for context-switching, without virtualization, without “simultaneous multi-threading”, and possibly without gc too. To do this well (it seems to me) the OS should be adapted to remove most of its advanced modern bits that enable these features, and focus more (or only) on dispatch of ready-processes to free cores, or the system (or device) could run without an OS at all. The use of RISC-V in this Meta MTIA seems to fit this perspective, as I doubt that this particular chip is meant to ever run an OS (similarly to a GPU). Hopefully they find a way to fit that paunchy-portly rotund-yummy dragon-stout fan in there somewhere!

]]>