Nvidia Blackwell Architecture Speculation

  • Thread starter Deleted member 2197
  • Start date
pharma said:
August 4, 2024
www.semianalysis.com

Nvidia’s Blackwell Reworked – Shipment Delays & GB200A Reworked Platforms

MGX GB200A NVL36, B102, B20, CoWoS-L, CoWoS-S, GB200A NVL64, ConnectX-8, Liquid Cooling vs Air Cooling, NVLink Backplane, PCB, CCL, Substrate, BMC, Power Delivery Nvidia’s Blackwell family is encou…
www.semianalysis.com www.semianalysis.com
Click to expand...
Were there any previous rumours of B102 being HBM and/or the highest-end gaming chip not being B102? Feels like someone is wrong, and while not perfect, I’d trust semianalysis more than the random claims I’ve seen so far. I also think we are late enough without enough leaks that September is impossible but October/November is not impossible yet - but I have no intuition about whether that’s likely or if it is indeed early 2025 at this point.

Regarding the B100 delays - it sounds like it would be more of a ramp delay than a shipment delay (but it is both), which might be explained by effectively losing 3 months of CoWoS-L capacity at TSMC with that equipment being idle (or at least not used for -L) while waiting for the respin. And maybe the real delay is slightly less than the respin delay because everything else wasn’t 100% ready to go, but it’s still really significant, and if anything I’m surprised NVIDIA didn’t drop more today…
 
Last edited:
Revenue guidance will be interesting. Even if they spin up alternative SKUs in record time their customers will need time to pivot assuming they decide to move forward with Nvidia’s stuff at all.
 
I actually think any impacted customer will likely stay the course and wait the extra 3 months for the revised gpu. The Blackwell arms race with other LLM competitors is effectly put on hold for everyone. Considering different hardware for a three month lapse might only be feasible in very low volume deployments and assuming no software dependencies (ie, Omniverse, Clara, Digital Twins, Cuda, etc...).
 
Arun said:
And maybe the real delay is slightly less than the respin delay because everything else wasn’t 100% ready to go, but it’s still really significant, and if anything I’m surprised NVIDIA didn’t drop more today…
Click to expand...

It's the sunk cost fallacy in investor's brains desperately battling the realization that a stock price can't go up forever, the main question is "when" that battle is finally lost.
 
Arun said:
Were there any previous rumours of B102 being HBM and/or the highest-end gaming chip not being B102? Feels like someone is wrong, and while not perfect, I’d trust semianalysis more than the random claims I’ve seen so far. I also think we are late enough without enough leaks that September is impossible but October/November is not impossible yet - but I have no intuition about whether that’s likely or if it is indeed early 2025 at this point.
Click to expand...
B102 is not GB202. Different products
 
pharma said:
I actually think any impacted customer will likely stay the course and wait the extra 3 months for the revised gpu
Click to expand...
What choice do they have? Competition doesn't have CUDA (very important), B200 is far ahead of the competition in performance (foe both training and inference), MI350X (the closest match to B200) will be Q4 2025 at best, Intel is nowhere to be seen, there is really no story here.

Server makers are still going ship in Q4.

https://twitter.com/x/status/1820643228611387622

Analysts don't think it matters.

"Worries over a potential delay of Nvidia’s latest AI chips appear overblown as demand continues to rise and major clients (cloud/hyperscalers) continue to grow their capex outlooks, Reuters reports, citing Bernstein analyst Stacy Rasgon, who said, “Nvidia’s competitive window is so large right now that we don’t think a 3-month delay will cause significant share shifts"

https://twitter.com/x/status/1820643525735985593
 
Last edited:
"We have not confirmed any design flaws or timing issues with B100 in calendar Q4, B200 in Q1 of calendar 2025, or GB200 in Q2 of calendar 2025. Speculation has centered on the B200's delay from late 2024 into early 2025. Note that the B200 is a different platform and a higher-performance part vs. the B100 (which has the same infrastructure as the current H100/200). Our current model has always assumed that NVDA would release the B200 in early 2025. We are comfortable with our Blackwell assumptions and the timing of the various SKU ramps for a current sampling of parts without any issues or delays"

https://twitter.com/x/status/1820789497833168902
 
These analysts are just rubbish - if they assumed no B100 revenue in Q3 and no B200/GB200 revenue in Q4, then they didn’t pay close enough attention. *If* that’s when the first revenue comes, there *is* a delay compared to what NVIDIA was hoping for 3 months ago.

Dylan Patel of semianalysis seems to think it won’t matter for NVIDIA’s revenue as hyperscalers will just buy more Hoppers. Maybe. I suppose Hopper was always going to be the majority of revenue until sometime in 2025 anyway given the ramp rate, so it’s not as big a shift as it might seem, but I am not super convinced yet.
 
Is the AI gold rush still so strong that companies literally dont care what Nvidia AI processor they even buy? That they'd rather just buy hundreds or thousands of H100 processors on short notice instead of waiting a few more months for Blackwell as they originally planned?
 
Seanspeed said:
Is the AI gold rush still so strong that companies literally dont care what Nvidia AI processor they even buy? That they'd rather just buy hundreds or thousands of H100 processors on short notice instead of waiting a few more months for Blackwell as they originally planned?
Click to expand...
I suspect a distinction that has been missed is a lot of the extra Hopper is probably H200 which is using supply-limited HBM3e that was otherwise going to Blackwell - i.e. lower/delayed Blackwell shipments probably allows NVIDIA to increase supply for H200 (which otherwise was also supply limited forcing customers to buy H100s when they wanted H200s). For inference if you don’t use FP4/FP6, H200 is possibly still very attractive at the right price.

Based on the reports, I expect it’s the B100/B200 used for inference that are effectively dead/low volume, with total GB200 NVL36/72 used for training mega-clusters only delayed by ~1.5 months with more of the (smaller) available capacity shifted to that.

So, I expect hyperscalers will *not* shift their training megacluster orders to Hopper and will just delay them slightly, but they will shift inference server orders mostly to H200. Again I have no idea how you go from this to it resulting in *more* revenue for NVIDIA overall as some analysts are saying, that feels like one magic trick too far to me.
 
DavidGraham said:
Analysts don't think it matters.

"Worries over a potential delay of Nvidia’s latest AI chips appear overblown as demand continues to rise and major clients (cloud/hyperscalers) continue to grow their capex outlooks, Reuters reports, citing Bernstein analyst Stacy Rasgon, who said, “Nvidia’s competitive window is so large right now that we don’t think a 3-month delay will cause significant share shifts"

https://twitter.com/x/status/1820643525735985593
Click to expand...

If analysts were good at their job they'd be called day traders.
 
Arun said:
Based on the reports, I expect it’s the B100/B200 used for inference that are effectively dead/low volume, with total GB200 NVL36/72 used for training mega-clusters only delayed by ~1.5 months with more of the (smaller) available capacity shifted to that.
Click to expand...
According to JPMorgan, NVIDIA's GB200 "Blackwell" AI servers are expected to see an annual output of 500,000 units in 2024, a decrease from the anticipated 600,000 units.

wccftech.com

NVIDIA's Blackwell GB200 AI Servers To See A Significant Drop In Production Figures, 500K Units Expected In 2024

It looks like NVIDIA Blackwell GPUs have become a victim of production foes as the firm now expects lower units, topping out at 500K in 2024.
wccftech.com wccftech.com


It remains clear that demand levels continue to rise, with all major hyperscalers continuing to grow their capex outlooks. Nvidia's competitive window is so large right now that we don't think a three-month delay will cause significant share shifts.
- Stacy Rasgon via Reuters
Click to expand...

"Despite all the noise recently regarding the ramp of the new AI GPUs, our checks suggest yields at TSM are stable even for advanced packaging and that wafer shipment forecasts for all applications, including AI/Blackwell, remain unchanged."
-Susquehanna on $NVDA
Click to expand...

https://twitter.com/x/status/1820838944310993048
 
TrendForce reports that NVIDIA is still on track to launch both the B100 and B200 in the 2H24 as it aims to target CSP customers. Additionally, a scaled-down B200A is planned for other enterprise clients, focusing on edge AI applications
Click to expand...

TrendForce reports that NVIDIA will prioritize the B100 and B200 for CSP customers with higher demand due to the tight production capacity of CoWoS-L. Shipments are expected to commence after 3Q24
Click to expand...

www.techpowerup.com

NVIDIA's New B200A Targets OEM Customers; High-End GPU Shipments Expected to Grow 55% in 2025

Despite recent rumors speculating on NVIDIA's supposed cancellation of the B100 in favor of the B200A, TrendForce reports that NVIDIA is still on track to launch both the B100 and B200 in the 2H24 as it aims to target CSP customers. Additionally, a scaled-down B200A is planned for other...
www.techpowerup.com www.techpowerup.com
 
It is estimated that the next generation of products will have a relatively large improvement, because it uses the symmetry of DDR7, and the performance has been greatly improved.

Because NVIDIA next-generation products have greatly reduced the power consumption of the GPU, for example, the highest one is 175, and the next-generation products, one only needs a maximum of 140 watts, and then it is 115, the maximum is 115, the best is 25 watts, which is the reverse power consumption, so what?

Right now with 40 series, you need 150-140, and then add full in the product and reduce 122 watts or so 20 watts, so, so, such a future happy product plus product, then the power consumption of the entire system will be reduced by 40 to 50 watts of many products, its performance is only 150 watts, now it can far exceed the current product performance.
Click to expand...

wccftech.com

NVIDIA GeForce RTX 5060 "Laptop" GPU To Consume Lower Power Than Ada, Features GDDR7 & RTX 4070-Level Performance

NVIDIA's GeForce RTX 5060 "Blackwell" laptop GPU could end up being a major upgrade over RTX 4060 according to Chinese manufacturer, Hasee.
wccftech.com wccftech.com

TLDR: Laptop 4060 max TDP 140W goes to Laptop 5060 max TDP 115W with the latter performing like Laptop 4070.
 
Last edited:
Why does the laptop 4060 need 140W to hit lower clocks than the 115W desktop version? Seems there’s a bit of sandbagging going on.
 
trinibwoy said:
Why does the laptop 4060 need 140W to hit lower clocks than the 115W desktop version? Seems there’s a bit of sandbagging going on.
Click to expand...

4050, 4060, and 4070 laptop are all 35~115W. 4080 is 60 ~ 150W and 4090 is 80~150W.
 
They aren't saying it pulls 140W continuously, nor that 140W is the limit everyone uses. THey're saying (and they're correct) the SKU permits limiting the GPU across a range of power, from 35W (I didn't look it up, I'm trusting pcchen on this number) to 140W. The actual power limit profile is determined by the manufacturer of the laptop and programmed into the firmware. This power limit is typically immutable, so the end user can't up the power slider in MSI Afterburner on a laptop as they might on a desktop part.

The new SKU apparently caps that range to 115W, therein lies the "interesting" story apparently.

Hope that clears it up.
 
You must log in or register to reply here.
Back
Top