Tuesday, May 13th 2025
Report: Customers Show Little Interest in AMD Instinct MI325X Accelerators
AMD's Instinct MI325X accelerator has struggled to gain traction with large customers, according to extensive data from SemiAnalysis. Launched in Q2 2025, the MI325X arrived roughly nine months after NVIDIA's H200 and concurrently with NVIDIA's "Blackwell" mass-production roll-out. That timing proved unfavourable, as many buyers opted instead for Blackwell's superior cost-per-performance ratio. Early interest from Microsoft in 2024 failed to translate into repeat orders. After the initial test purchases, Microsoft did not place any further commitments. In response, AMD reduced its margin expectations in an effort to attract other major clients. Oracle and a handful of additional hyperscalers have since expressed renewed interest, but these purchases remain modest compared with NVIDIA's volume.
A fundamental limitation of the MI325X is its eight-GPU scale-up capacity. By contrast, NVIDIA's rack-scale GB200 NVL72 supports up to 72 GPUs in a single cluster. For large-scale AI inference and frontier-level reasoning workloads, that difference is decisive. AMD positioned the MI325X against NVIDIA's air-cooled HGX B200 NVL8 and HGX B300 NVL16 modules. Even in that non-rack-scale segment, NVIDIA maintains an advantage in both raw performance and total-cost-of-ownership efficiency. Nonetheless, there remains potential for the MI325X in smaller-scale deployments that do not require extensive GPU clusters. Smaller model inference should be sufficient for eight GPU clusters, where lots of memory bandwidth and capacity are the primary needs. AMD continues to improve its software ecosystem and maintain competitive pricing, so AI labs developing mid-sized AI models may find the MI325X appealing.
Source:
SemiAnalysis
A fundamental limitation of the MI325X is its eight-GPU scale-up capacity. By contrast, NVIDIA's rack-scale GB200 NVL72 supports up to 72 GPUs in a single cluster. For large-scale AI inference and frontier-level reasoning workloads, that difference is decisive. AMD positioned the MI325X against NVIDIA's air-cooled HGX B200 NVL8 and HGX B300 NVL16 modules. Even in that non-rack-scale segment, NVIDIA maintains an advantage in both raw performance and total-cost-of-ownership efficiency. Nonetheless, there remains potential for the MI325X in smaller-scale deployments that do not require extensive GPU clusters. Smaller model inference should be sufficient for eight GPU clusters, where lots of memory bandwidth and capacity are the primary needs. AMD continues to improve its software ecosystem and maintain competitive pricing, so AI labs developing mid-sized AI models may find the MI325X appealing.
22 Comments on Report: Customers Show Little Interest in AMD Instinct MI325X Accelerators
That being said, as someone that would like to actually get a GPU in the next year or two, I for one would love to see the slow death of the AI GPU industry. I'd like to see the next generation of GPU's be actually in stock consistently in stores, unlike the current generation.
Anyways, after reading the link below, I was reminded that there are places with strong bias that will buy only from certain vendors, regardless of others offering something equal or better for less money.
Company that I worked did this with intel CPUs, just because the clueless managers still believed the old saying "Nobody got fired by buying Intel".
Anyways, interesting read here.
spectrum.ieee.org/ai-inference
Their new article is just showing improvements AMD made to software.
Another thing is that deploying those products is really painful and AMD is really lacking in this regard as well, albeit there have been some improvements (as shown in the link that I quoted above).
You can read more about those hurdles here:
semianalysis.com/2024/12/22/mi300x-vs-h100-vs-h200-benchmark-part-1-training/
They are improving a lot, but they still fall behind Nvidia quite a lot, specially when new features take way too long to be ported to AMD (if at all). AMD is playing the catch up game against a target that's still running pretty fast.
Nonetheless, from your own link:It's still behind the H200, which is the previous gen, and AMD suffers from not having a proper ecosystem/platform to scale out to multiple nodes as seamlessly as Nvidia.
The rest of your comment is spot on, they need a lot of work, but they also seem to be advancing quickly.
Without a robust dev environment, there's no way AMD is going to catch up with Nvidia here. We've seen this for 10+ years in the consumer GPU side and gaming drivers.
The other main metric for compute/datacenter is performance-per-watt. If AMD can take a significant advantage here (right now it does not), there might be a chance where some potential customers will make the switch to endure AMD's weak software to take some power savings (total cost of operation over the lifetime of the deployed hardware).
This is not applicable for enthusiast PC gaming because Joe Gamer really doesn't care how much power their graphics card is drawing when they're playing some poorly-optimized AAA title.
The fact that we're even comparing it to a H200 and not the Blackwell products already shows that it's not in great light.
Gaming / Art / AI everywhere, everywhere the same story.
In Gaming Nvidia has better features and actively investing in R&D of new features that has never been heard before.
In Art Nvidia monopoly is not figurative it's literall, Renderers only support Nvidia. Literally only Nvidia works ( excluding just a couple exceptions)
As outlined by this news report apparently same situtation persists in AI segment as well.
Once the genie is out of the bottle...
The zluda dev had to revert all of the work done during the time AMD sponsored the project, and is now basically starting from scratch at a way slower pace.
I don't think they need a CUDA layer, they are making great strides with ROCm. It's still bad, but it's improving every day and is in way better shape than it was 1 year ago.
Apple managed to get MPS out there in lots of frameworks, AMD surely can sort their shit out and become truly competitive (software-wise) as well.