Last I heard the Jupiter's CPU modules were to be of SiPearl’s design, namely Rhea1 ARM chips. Nvidia was chosen as one of the accelerator providers. There should eventually be others as well.
I suspect that there has been a sudden outbreak of realism, the HPC users have won out over the industrial strategists, and Jülich are buying a machine made of off-the-shelf boxes rather than trusting a smallish company with no track record with an order worth more than the company.
(that ARM has moved from 'clearly greatest tech company in the EU' to 'only NASDAQ listed' in a decade is probably also significant; at least part of SiPearl's purpose to exist was European self-reliance)
I'd not be entirely surprised if there were a small SiPearl machine next to Jupiter.
"This will boost the H200’s memory bandwidth from 80GB to 141GB, and memory bandwidth from 3.35TB/second to what NVIDIA is preliminarily expecting to be 4.8TB/second." [5th paragraph from top bottom part of paragraph]
Dis you mean memory capacity from 80GB to 141GB Instead?
Has anyone pushed nVidia on this bizarre 141GB figure? The only way I can get it to make sense is if they are accepting one broken 24Gbit die among the six eight-high stacks; I had thought HBM constructed its 1024-bit-wide bus as 128 bits from each die in the stack, if only to avoid routing 1024 fast lines from each die down to the controller at the bottom.
They've confirmed it's a yield/cost thing. I don't have any further details than that, but by reserving 3GB, they're able to ship more accelerators than they otherwise would have been able.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
14 Comments
Back to Article
SarahKerrigan - Monday, November 13, 2023 - link
"1 TFLOPS is just a bit under what 23,762 H100s can provide with FP64 tensor operations."Seems a little low, but what's a few orders of magnitude between friends?
Threska - Monday, November 13, 2023 - link
Now one can run CuLitho.https://youtu.be/HxyM2Chu9Vc
jvl - Tuesday, November 21, 2023 - link
Quietly corrected?thomasjkenney - Monday, November 13, 2023 - link
HAHA! I have 9 working GP computers (3 macs incl), and their combined RAM is still less than used in this H200 box. Crikey!Dante Verizon - Thursday, November 16, 2023 - link
An average high-end PC has almost 50gb of memory counting Vram and Ram.TeXWiller - Monday, November 13, 2023 - link
Last I heard the Jupiter's CPU modules were to be of SiPearl’s design, namely Rhea1 ARM chips. Nvidia was chosen as one of the accelerator providers. There should eventually be others as well.TomWomack - Tuesday, November 14, 2023 - link
I suspect that there has been a sudden outbreak of realism, the HPC users have won out over the industrial strategists, and Jülich are buying a machine made of off-the-shelf boxes rather than trusting a smallish company with no track record with an order worth more than the company.(that ARM has moved from 'clearly greatest tech company in the EU' to 'only NASDAQ listed' in a decade is probably also significant; at least part of SiPearl's purpose to exist was European self-reliance)
I'd not be entirely surprised if there were a small SiPearl machine next to Jupiter.
FWhitTrampoline - Monday, November 13, 2023 - link
"This will boost the H200’s memory bandwidth from 80GB to 141GB, and memory bandwidth from 3.35TB/second to what NVIDIA is preliminarily expecting to be 4.8TB/second." [5th paragraph from top bottom part of paragraph]Dis you mean memory capacity from 80GB to 141GB Instead?
Ryan Smith - Monday, November 13, 2023 - link
Yes I did. Thanks!Rοb - Monday, November 13, 2023 - link
Cost of Jupiter: €273M out of an operating budget of €500M for six years.Source: https://www.nextplatform.com/2023/10/05/details-em...
TomWomack - Tuesday, November 14, 2023 - link
Has anyone pushed nVidia on this bizarre 141GB figure? The only way I can get it to make sense is if they are accepting one broken 24Gbit die among the six eight-high stacks; I had thought HBM constructed its 1024-bit-wide bus as 128 bits from each die in the stack, if only to avoid routing 1024 fast lines from each die down to the controller at the bottom.Ryan Smith - Tuesday, November 14, 2023 - link
They've confirmed it's a yield/cost thing. I don't have any further details than that, but by reserving 3GB, they're able to ship more accelerators than they otherwise would have been able.StevoLincolnite - Tuesday, November 14, 2023 - link
But can it run Crysis?LordSojar - Tuesday, November 14, 2023 - link
Uhhh..............................................................................................................................................