alt Hacker NewsThat says more about Intel's shambolic efforts than it does about AMD's road to damascus moment.
Honestly both of them have demonstrated that they have little interest in AI. Intel gave up on Optane right as (a much-improved successor) would've helped with MoE offsetting or even as a highly durable drive for KV caching.
AMD have... turned out some decent gaming graphics cards and fine server CPUs but little else.
People are out there solding SMDs on old Nvidia mining rig GPUs to use them for inference; their hardware is not depreciating at the normal rate you'd expect from GPUs. DGX spark is finally useful now that AI can write multi-machine inference tooling optimised to it. Where's AMD in all of this? Nowhere.
Replies
AMD has their Strix/Gorgon Halo as a competitor to DGX Spark, which is at least in the ballpark in terms of bandwidth, and being x86 has some compatibility benefits.
If you wanted a GPU with 32GB of memory for AI tasks, the AMD AI Pro R9700 was (up until about a month ago) reasonably priced at ~$1300. It's not a RTX 5090 in terms of bandwidth and compute, but for the price and being actually purchasable it's a reasonable pick.
Not to forget AMDs dominance in performance per dollar in CPUs of any variety up until very recently on basically all fronts. Consumer, handheld/low power, server, you name it.
Intel finally has really great low power/budget answers, but still loses greatly to 3d cache chips on most workloads.
AND GPUs are competitive on a price/performance standpoint. Their software stack is where they fall behind.
>AMD have... turned out some decent gaming graphics cards
that's quite a large market and their only competitor has moved on to greener pastures.
Persistent memory will be back, it's just too compelling to give up on, and thanks to the research industry it spawned we have a pretty good idea of how to build systems that effectively exploit it. I can't wait until every DB can use pmem to achieve the holy grail of low-latency transactional durability (which is impossible today even with NVMe).
Doesn't Intel have near cutting edge fabs tho? Even if their own AI chips aren't great, aren't they getting contracts to make Nvidia and Apple chips?
Intel pirated Optane and bailed when they got caught, seemed unrelated to AI, just Intel being Intel.
>Honestly both of them have demonstrated that they have little interest in AI. Awesome? Maybe this means supply wont be ruined and I'll be able to buy CPUs.
Huh? Instinct is basically the only real alternative to Nvidia in the market. And they have the only real rack scale competition with Helios. This revenue is not because of gaming gpus and server cpus. This literally is AMDs road to damascus moment with data center GPUs.
"road to damascus"
Great Hope and Crosby movie.
> Optane
Optane really died before LLMs went big. I was at an enterprise using a TON of them, often for SAP HANA workloads.
> AMD have... turned out some decent gaming graphics cards and fine server CPUs but little else.
I'm on a AMD desktop, using an AMD gpu, next to a gaming console using an AMD CPU and GPU, making large DC purchasing decisions where intel hasn't even shown up as a factor.
Ahh TIL! I used ChatGPT to find out what your "..road to damascus moment." expression meant, it was quite complimentary to you;
> And honestly, "Intel's shambolic efforts" alongside "AMD's Road to Damascus moment" is rather delicious phrasing. The author has managed to deliver quite a sophisticated insult in a single sentence.
I never understood why some fraction of technologists seem convinced that Optane was this amazing thing that would've been awesome if only Intel hadn't killed it.
It never made sense. It's slower than RAM, and what do I get for it? Persistence? I don't want persistent RAM, I like the ability to reboot to solve problems, and storing stuff on a non-volatile mediums has all sorts of legal implications which it's easier to avoid.
It's more expensive and has less storage than SSDs, even though the main downside of SSDs is their high price per gigabyte and relatively low capacities.
It's going in the wrong direction for both use cases. I guess it could have a very niche use case as a super fast persistent caching layer in front of spinning rust but that's kinda it.