AMD CDNA: MI300 & MI400 (Analysis, Speculation and Rumors in 2024)

LordEC911 said:
Really just depends on when they finalized this "strategy" and started planning the "backwards compatibility" into their architectures.
Click to expand...

While they have started teasing this only now, I am more in the camp “they wrapped up the planning and architectural designs for 2026 IPs” as the reason why they start talking about it.

They have had 5 years of insights from RDNA and CDNA products out in the wild. These ecosystem feedbacks and observations are not a random surge out of nowhere in this past few weeks.

Also given the CDNA-RDNA run has close to 5 years since first product launch by now, it is not a total surprise that the next multi-year architectural roadmap (incl. high level architectural features & big bets) is due or done by now.
 
Last edited:
trinibwoy said:
In-depth look at MI300X performance and user experience. Not a pretty picture.

semianalysis.com

MI300X vs H100 vs H200 Benchmark Part 1: Training – CUDA Moat Still Alive

Intro SemiAnalysis has been on a five-month long quest to settle the reality of MI300X. In theory, the MI300X should be at a huge advantage over Nvidia’s H100 and H200 in terms of specifications an…
semianalysis.com semianalysis.com
Click to expand...
I was in Taiwan this week visiting my supply chain and I can say that MI325 and MI350 are nowhere to be seen in the production pipeline. Yes AMD software stack is not pretty but hyperscalers can swallow it. What they can't swallow is the huge leap in performance and lower TCO that NVL72 racks bring to the table in real world reasoning inference. In fact, most tiers one customers are already updating their orders to B300 (Blackwell Ultra) when the delivery is set for end of the year. To the point that more than 70% of 2025 CoWoS capacity is taken by Nvidia (after AMD canceled orders due to low market interest). Scraps are between AMD, Broadcom and Marvel.
Jensen was right, the new GPU is a rack. And it has another domino effect, Nvidia also won the AI CPU battle this round, totally excluding intel and AMD. The vast majority of first 6 months production are with Grace. Only low volume DGX is shipped with Intel Xeon 8570.
I hope AMD will bring the fight back at the rack level with MI400 and whatever scale out solution they have because we are very close to a monopoly in AI.
 
trinibwoy said:
Well something is shipping as AMD’s datacenter numbers are trending up and at least one provider is offering MI325X.

blogs.vultr.com

Announcing the AMD Instinct™ MI325X Accelerator on Vultr Cloud | Vultr Blogs

Discover the power of the AMD Instinct™ MI325X GPU on Vultr Cloud. Ideal for AI model inference, HPC, and data analytics, with 256 GB HBM3E memory and 6 TB peak memory bandwidth.
blogs.vultr.com blogs.vultr.com
Click to expand...
Well I just checked and 2025 AMD Instinct revenue is expected to barely reach $10B (whole DC including EPYC CPUs and networking will be around $20B) when Nvidia aim at more than $160B for the same period. As I said, nearly nothing left for AMD...
 
xpea said:
Well I just checked and 2025 AMD Instinct revenue is expected to barely reach $10B (whole DC including EPYC CPUs and networking will be around $20B) when Nvidia aim at more than $160B for the same period. As I said, nearly nothing left for AMD...
Click to expand...
$10b for Instinct is the conservative number, they are just doubling the number from 2024.
MI350 is already ramping and has solid bookings for 2H '25.
 
xpea said:
This is not what I heard. In fact, AMD canceled some CoWoS capacity because of low demand.
Click to expand...
Thar rumor has happened at least 3 times the past year... the first time was after there was news about AMD looking for an additional partner/supplier.
Did you also believe the rumor that Nvidia, AMD, and Broadwell are cutting CoWoS capacity? Or did TSMC stating, "No customer is cutting CoWoS orders" clarify that at all?
If you read between the lines of the rumor, it appears to due to the new facilities that TSMC is bringing online in 2025.
They were originally going to also have CoWoS-S lines but now appears to be CoWoS-L only which seems to imply that Nvidia switching to CoWoS-L frees up enough CoWoS-S capacity in 2025.
 
You must log in or register to reply here.
Back
Top