That says more about Intel's shambolic efforts than it does about AMD's road to damascus moment.<p>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.<p>AMD have... turned out some decent gaming graphics cards and fine server CPUs but little else.<p>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.
AMD is doing 5 day RTO for almost everyone with 3 months of notice in an obvious attempt to get people to quit before vesting their 5x stock grants.<p>The rule includes >90% of "remote" workers who have discovered they were classified as hybrid (a classification now eliminated) and their manager didn't make it official. Now, remote approvals are SVP level.<p>The issue is, AMD still pays significantly less than any company with an RTO mandate, even with the stock appreciation. And the stock has gone down 25% in the past few weeks so it's even less competitive. The typical engineer was <i>not</i> given an equity-weighted package, even L9 only 30% of your TC is in stock.<p>It's a recruiting bonanza right now and everyone's getting calls offering 50% pay bumps from 5-day office semi corps like Broadcom or Qualcomm. The better people are getting offers at Google or Apple for 3-day RTO for even greater TC bumps.
I'm not surprised.<p>The board should've sacked the entire leadership team of Intel and AMD years and years ago.<p>The fact that you can sit through both a mining boom and now a generational AI boom and have little to show for it and still keep your job is a Yahoo!-level fumble.
They can't find new leaders! Sharon Zhou, VP who handles all our engagements with the frontier labs quit today to be an MTS at Anthropic. <a href="https://lnkd.in/p/g53S3-fG">https://lnkd.in/p/g53S3-fG</a><p>Sharon was the AI Chief of Staff brought in through an acquihire, managed all the high-visibility external/internal AI engagements, reported directly into Lisa Su, and was being groomed to be Lisa's successor. Even with all that, she couldn't make it 2 years at AMD.<p>I don't see any way out of this hole. We can't fire our execs and promote internally because our internal directors/VPs are mostly non-technical. We've been trying to pull leadership in with acquihires but we've been losing them too. Best win was replacing Alan Lee w/ Anush Elangovan which finally got ROCm working but there's a massive re-org coming he might not survive.
I can apply as a ceo if you accept master students. Having studied drama, I can definitely say I act better than pat gelsinger
How sad.<p>I remember buying an AMD K6-2 back in the day as it was cheaper and better. I've dropped in and out of the AMD sphere depending on who was winning at the given moment.<p>Now they're both losing.
> our internal directors/VPs are mostly non-technical<p>why not start with trying to fix that?
100% Lisa Su is a control freak and thought Zhou was a threat. Cult of personality CEOs do not generally end well for the company
Ryzen series chips might not have been released if AMD's entire leadership team was sacked many years ago.<p>The heavy investments into HMB between SK Hynix & AMD might also not have happened. They went down this road before HMB demand popped off and it was a risky bet.
Intel have gone through 3 leadership teams in 10ish years
They brought back some questionable former managers like the guy responsible for 14nm, Sanjay Natarajan, which was the first process to get stuck with a protracted period of low yields.
Gosh, you're right. I thought it was just one.
> The issue is, AMD still pays significantly less than any company with an RTO mandate<p>I'm sure they pay more than Intel ;-)
Interesting data point as a stock holder, thanks.
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.<p>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.
AMD has an entire division dominating in HPC and #2 close behind nVidia in datacenters with extremely good margins. Over half their revenue is from those segments. Not a lot comes from consumer devices right now. That's why they did so well against Intel right now. Intel is the one that is failing in those markets.
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.<p>It never made sense. It's slower than RAM, and what do I get for it? Persistence? I don't <i>want</i> 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.<p>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.<p>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.
Having a very low latency, durable write destination is a superpower for lots of database architectures (or things that share architectural details with databases, like some filesystems).<p>Optane had much lower latency than other types of durable storage, and also had very good endurance compared to NAND.<p>At a prior employer, we made extensive use of them as ZFS SLOG drives for large database servers. They made fsync very inexpensive.
Yep. I use a P4800X for ZFS special vdev. It's a monster. The thing has write endurance for many petabytes of write. The queue depth alone make for fast small reads; and the architecture means sustained reads and writes are... sustained. Try that with a normal SSD/NVME.
And instead of upgrading latency from NVMe SSDs to Optane, we downgraded to EBS.
Optane was cheaper than DRAM per bit.<p>If Intel had volume production of Optane and had been able to quickly reconfigure it to produce a device that could stream out model weights fast enough for inference or could have gotten a read-write device that could hold KV caches effectively, they could have made an absolute killing in the current market.<p>HBF is slowly going after the former market. I'm not sure flash has a great answer for the latter due to endurance issues.<p>I can't think of any company that is as good as Intel at having a technology that's ready or almost ready at or before the time that the market is ready for it and then completely failing to sell the product into the newly developed market.
> what do I get for it? Persistence?<p>Density. PCM is closer in bits-per-unit-area to Flash than to DRAM.
Huh. I wrote a whole thing here about serving a niche in serving local LLMs, but turns out, any niche it could have served is probably better served (per dollar) by either more RAM or modern SSDs.<p>There's an argument to be made for the persistent memory being a cheaper substitute for more RAM, but imo the gap isn't big enough to justify it when considered as a % of the rest of the system you're building around it and the performance tradeoff.<p>Important reminder to check your biases. I was sure Optane had missed its mark by dropping out just before LLMs became a primary use case for high $ systems.
IIRC when it came out it was faster than the SSDs you could buy. It never really got that much better and then SSD mostly caught up.
The other answers are hinting at it, but Optane was fundamentally much much lower latency than NAND flash used in SSDs.<p>SSDs use lots of tricks to hide this latency (using RAM). However, when the data is super critical, you can't just throw some data in RAM and hope that it gets written out.<p>Enterprise flash drives have battery backups and keep the system alive for a moment to flush out the writes, but Optane doesn't need special tricks to get this sort of durability, it just has it. The Optane can keep accepting writes, eventually the RAM of the SSD is filled, and then the write performance falls off a cliff. However, if your workload gives the storage device a breather to write out to flash, you wouldn't notice the difference.<p>If you have a large stream of critical data flowing in that has to be stored and can't be lost- Optane suddenly looks way better than SSDs
You can have very large MoE models run on small amounts of ram (or vram), if you can transfer the experts you need on the fly from ram to disk.<p>SSDs are good, but an improved Optane would've been better.
Sometimes, things are just cool.
It somewhat (maybe) had sense for HDD caching but I don't think HDDs in 2026 are used in cases where separate cache is needed.
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.<p>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.
And they are making great progress on the software side. ROCm 10 is significantly better than 7.x
Honest question, couldn't one just ask Claude to make a better driver for it? And reimplement pytorch on top of that or something?
The community is basically doing this now with the kickflips in driver/rocm/etc "stuff". It's gotten so good and well-embraced that even AMD is backing some of these community efforts.
Is there something wrong with PyTorch? I’ve heard people complain a lot about their GPUs. I’ve only used MI250x in an HPC, but PyTorch worked fine.
It's due to ROCm, think equivalent of CUDA lower level drivers, which is well problematic.<p>There's huge upswing in using Vulkan Compute which works across NVIDIA, AMD, and Intel (and even others), and which is apparently much easier and stable to run on AMD
>AMD have... turned out some decent gaming graphics cards<p>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).
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.
What's more interesting to me is how blisteringly evident it is that GPUs depreciate a lot slower than the analysts assumed. But Nvidia GPUs depreciate slowly because a lot of thought and resources were dedicated into building that ecosystem.<p>AMD has always been an also ran in this space and it shows to this day, but the true believers cannot be dissuaded of their faith so whatevs.<p>I honestly think Nvidia needs a worthy nemesis right now more than ever and I wish it existed but I don't think it does... Yet...
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?
> Where's AMD in all of this? Nowhere.<p>They're around, selling AMD Strix Halo (up to 128GB unified memory) on the low, equivalent to DGX Spark, level; and MI* chips for enterprise AI inference.<p>The problem they have and they've always had is that everyone uses CUDA. ROCm works, but everyone's toolchain is CUDA exclusive or CUDA first, so it's much harder to use AMD hardware.<p>Using higher level tooling (vLLM, llama.cpp) I have DGX Spark comparable performance for 1/3 the price (but without the same software out of the box experience nor the clustering network options).
"road to damascus"<p>Great Hope and Crosby movie.
>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.
> Optane<p>Optane really died before LLMs went big. I was at an enterprise using a TON of them, often for SAP HANA workloads.<p>> AMD have... turned out some decent gaming graphics cards and fine server CPUs but little else.<p>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.
Intel pirated Optane and bailed when they got caught, seemed unrelated to AI, just Intel being Intel.
Ahh TIL! I used ChatGPT to find out what your "..road to damascus moment." expression meant, it was quite complimentary to you;<p>> 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.