Speculation and Rumors: Nvidia Blackwell ...

  • Thread starter Deleted member 2197
  • Start date
AFAIK (and I really need to double check this), NVIDIA’s L1 cache is still in-order for all misses. That is, if all of a warp’s load requests hit in the L1, it can reschedule earlier than an older warp reques that missed - but if a single thread misses in the L1, then it might have to wait until another warp’s data comes back from the far DRAM controller even though it hit in the L2 and could reschedule much sooner.

A lot of these patents therefore make little sense in the context of the current architecture and the bigger question imo is how their cache hierarchy will evolve in Blackwell.

The main benefit of that patent seems to be earlier submission of memory ops from the SMs perspective which is a pretty nice benefit even if the memory subsystem itself isn’t any faster.

The hardware cost for all of the scoreboarding required is likely not worth it though. I think we talked about this same paper a while back.
 
"This week we are sending out engineering samples of Blackwell all over the world they are under people's chairs right now," said Huang.

Sending samples of Blackwell B100 and B200 processors is a big deal for Nvidia, which plans to ship these products commercially in the fourth quarter. It is unclear whether Nvidia is beginning to sample Blackwell processors with hardware or software developers.

In fact, given that Nvidia's partners among hardware makers, such as Foxconn, Quanta, Wistron, Pegatron, and Asus have demonstrated their Blackwell-based servers at Computex, they may have been playing with Blackwell processors for a while now. Yet, so far not all software developers have had access to Nvidia's forthcoming processors for AI and HPC applications.
 
Rumors suggest there have been Blackwell "samples" in the hands of the Big Three server manufacturers for all of this calendar year if not even Q4 of '23...
 
Both statements could be true since we don’t have actual dates from either report. A 3 month slip for a Q3 to Q4 ramp would still be in 2H. No big deal either way.
Both statements could also be true if they're each referring to different segments - gaming Blackwell vs data centre Blackwell.
 
Someone did the math if the Blackwell rumor has any truth:
Here’s the math on the rumored B200 delay:

The Blackwell architecture was planned to start shipping in several varieties. There are the B100 and B200 chips, with the latter more powerful and in higher demand. Then there is the GB200, which is faster still because it packages two B200s with a third processor that all share data at high bandwidth.

Nvidia may be able to charge $30,000 to $40,000 for the B200, and $50,000 to $70,000 for the GB200, according to Mizuho analyst Vijay Rakesh.

One of the more detailed models out there for Nvidia product sales is from Timothy Arcuri at UBS. As of last month, he was projecting shipments of about 32,500 B200 chips and 43,400 GB200 modules in Nvidia’s quarter ending January 2025. That would be only about 7% of all the accelerators shipped in the quarter because the vast majority of units shipped would still be Hopper chips.

Combining both analysts’ intelligence, the January quarter contribution from B200 products would amount to $3 billion. That’s 9% of the $34.5 billion in total revenue expected for the quarter, per the consensus forecast.

But even if the delay materializes, the revenue pushout for Nvidia is unlikely to be as large as that. Customers still can’t get all the Hopper chips they want. Nvidia has nearly caught up with demand for the H100 version but the more powerful H200 remains backlogged.

So Nvidia could just crank out more Hopper chips and sell all they make.
 
Someone did the math if the Blackwell rumor has any truth

That seems to address impact of the rumor not its veracity. Any sign of weakness in Nvidia’s execution could send shareholders scurrying so it’s a big deal even if the actual impact is a nothing burger.
 
That seems to address impact of the rumor not its veracity. Any sign of weakness in Nvidia’s execution could send shareholders scurrying so it’s a big deal even if the actual impact is a nothing burger.
Did see someone that should be familiar with semi manufacturing say 3 months is a good estimate for a respin. So if it's a false rumor it's either lucky or from someone unusually knowledgeable about how semi works.

I'd calculate the financial hit to Nvidia different than just estimating cashflow. The AI balloon has a leak already, Nvidia is under anti trust investigation by complaints from both competitors and customers. A delay of 3 months piled atop a waitlist that customers already don't like could have worse effects than just missing some cashflow points, it's a breakpoint for customers to wonder if they're actually making a smart purchasing decision.
 
Last edited:
waitlist that customers already don't like could have worse effects than just missing some cashflow points, it's a breakpoint for customers to wonder if they're actually making a smart purchasing decision.
There is no alternative to what NVIDIA offers, B200 on it's own is immensely more powerful than anything else on the market, and if we talk about B200 server racks then it won't be matched for years.
 
Back
Top