Another day, another delay. This time it looks like it’s Nvidia’s Blackwell AI GPUs that may have fallen victim to last minute delay-itus, as reports now suggest they’ll be launching in the first half of 2025 instead.
While engineering samples for the B100 and B200 GPUs have already been delivered, it seems enterprise customers including Microsoft, Meta and xAI might have to wait for their orders. According to Tom’s Hardware, two anonymous sources who worked on the chips first released news of the pushback, which was later allegedly corroborated with an anonymous Microsoft source by Bloomberg.
Previously, it was thought that Nvidia was aiming for an end of 2024 release. With multiple high-profile customers placing huge orders for Nvidia’s latest mega-GPUs, and prices estimated to run at $70,000 per “superchip”, significant delays are not likely to be well received.
While these sources have pointed at design flaws as a reason for the pushback, Dylan Patel, Chief Analyst at independent research company SemiAnalysis, has posted a tweet telling a different story.
According to an update sent out to SemiAnalysis’ clients on July 22, CoWoS-L (Chip on Wafer on Substrate) ramping issues are primarily to blame, with TSMC’s AP3 packaging fab needing to be shut down to convert from CoWoS-S production to CoWoS-L.
This process is not expected to be completed until the second quarter of next year. The tweet also mentions rumours about overheating and firmware issues, but characterises them as “overblown”. While the update mentions a variety of “teasing pains to be fixed”, SemiAnalysis seems to think that they are not the primary cause of the delays.
People keep asking about Blackwell delaysWe already sent out the our update 2 weeks ago, July 22nd to our hyperscaler, semiconductor, and investor clientsSell side and media are quite lateThey still have multiple things wrong tooVolumes, ASP, etc belowhttps://t.co/frahauIDRt pic.twitter.com/Oap8CSdZBrAugust 3, 2024
Whatever the cause, any significant delay in shipping could potentially shake confidence in Nvidia’s ability to deliver its costly AI-processing GPUs en masse and on time.
That being said, Nvidia are unlikely to be worried that any significant delay will result in lost customers. While AMD has its own competing AI GPU, the Instinct MI300, AMD CEO Dr. Lisa Su has recently indicated that it’s currently selling as many as it can make, and that “the overall supply chain is tight and will remain tight through 2025.”
Given Nvidia’s dominance in the sector and the reported potential performance of the Blackwell series AI GPUs, it’s unlikely any major customer will switch over to other hardware providers in the face of a potential delay, especially if the top candidates are in the same short supply boat.
Once again it seems, TSMC holds the key to the world’s chipmaking capabilities, and if it needs to reconfigure to keep up with demand, the rest of the world’s top tech companies may just have to stand in line and wait. Or perhaps there really are design issues with Nvidia’s latest AI hardware holding things back.
Still, at least it doesn’t seem to be a mistake at the printers, ey?
