Question Fujitsu A64FX successor Monaka

[soresu]

Successor to Fujitsu's custom supercomputer ARM CPU A64FX named as Monaka, with ARMv9, SVE2 and 150 cores scheduled for 2026/27 release, apparently targetted at datacenters rather than supercomputer use cases.

Width of the SVE2 units is not declared, but it will be built on TSMC's N2 node.

Link.

 

[soresu]

Anandtech posted an article about it back in March this year.

 

[moinmoin]

This is quite some way out still. The text in the slide, "Providing overwhelming energy efficiency compared to CPUs release by 2027 from competitors", is quite the optimistic bet. Looking forward to seeing how that plays out.

 

[DrMrLordX]

Strange, why the shift in focus from HPC to datacentre?

 

[moinmoin]

Strange, why the shift in focus from HPC to datacentre?

Fujitsu A64FX was mainly built and sold for the Fugaku supercomputer, so my only guess is that a supercomputer project of comparable size doesn't exist (yet?) for Monaka so Fujitsu looks to sell more of it through regular channels.

 

[DrMrLordX]

Fujitsu A64FX was mainly built and sold for the Fugaku supercomputer, so my only guess is that a supercomputer project of comparable size doesn't exist (yet?) for Monaka so Fujitsu looks to sell more of it through regular channels.

A bit unfortunate. Fugaku/A64FX achieved what Intel never could with Larrabee and its successors. The simplicity of the programming model for Fugaku vs. other supercomputers alone should have made such a successor appealing to data scientists.

 

[soresu]

A bit unfortunate. Fugaku/A64FX achieved what Intel never could with Larrabee and its successors. The simplicity of the programming model for Fugaku vs. other supercomputers alone should have made such a successor appealing to data scientists.

I feel like Fugaku was also innovating on the ARM server concept itself with Fujitsu partnering with ARM to drive the development of SVE's first iteration among other things.

Now that those hurdles are already crossed they can just put out new chips based on A64FX, and whoever wants to use them can do so just as with Intel and AMD CPU options.

I think this is more like Fujitsu saying "stock ARM isn't the only option if you want moar SIMD grunt" to the datacenter crowd, while Monaka can probably still be used for exactly the same purpose as Fugaku.

Essentially this is Fujitsu trying to monetise all the hard work they put in here while the ARM server market is still somewhat in flux.

 

[moinmoin]

In some surprising twist Fujitsu is partnering with AMD to bring Instinct and ROCm to Monaka and ARM.

Fujitsu and AMD to begin strategic partnership to develop more sustainable computing infrastructure intended to accelerate open-source AI initiatives

Fujitsu and AMD to begin strategic partnership to develop more sustainable computing infrastructure intended to accelerate open-source AI initiatives

www.fujitsu.com

https://www.amd.com/en/newsroom/press-releases/2024-11-01-amd-and-fujitsu-to-begin-strategic-partnership-to-develop-more-s.html

 

[soresu]

In some surprising twist Fujitsu is partnering with AMD to bring Instinct and ROCm to Monaka and ARM.

Fujitsu and AMD to begin strategic partnership to develop more sustainable computing infrastructure intended to accelerate open-source AI initiatives

Fujitsu and AMD to begin strategic partnership to develop more sustainable computing infrastructure intended to accelerate open-source AI initiatives

www.fujitsu.com

https://www.amd.com/en/newsroom/press-releases/2024-11-01-amd-and-fujitsu-to-begin-strategic-partnership-to-develop-more-s.html

Makes sense, after all there's no gain in turning their nose up at server/datacenter opportunities based on non x86 CPUs when Instinct is one of their biggest revenue drivers.

 

[soresu]

Chips and Cheese article about Monaka and a Youtube video....

 

[soresu]

Looks like it has a similar $ bottom, CPU top die stacking layout to Zen 5 V cache SKUs, albeit with combined SRAM/IO dies like RDNA3, and a silicon interposer tying the stacks together.

 

[DrMrLordX]

Chips and Cheese article about Monaka and a Youtube video....

Thing looks pretty beastly. It'll be a long time before we hear of it though.

 

[soresu]

Found the ARM µArch layouts link, but it doesn't seem to show how many NEON/SVE2 units each one has.

AFAIK A7xx retains the same 2x 128 bit unit configuration that A78 had, can anyone confirm or refute this?

 

[DrMrLordX]

Found the ARM µArch layouts link, but it doesn't seem to show how many NEON/SVE2 units each one has.

AFAIK A7xx retains the same 2x 128 bit unit configuration that A78 had, can anyone confirm or refute this?

https://documentation-service.arm.com/static/65f1692987f147198672973b?token=

Looks like 2x128b. Check page 14 aaaaaaand 12.

 

[soresu]

https://documentation-service.arm.com/static/65f1692987f147198672973b?token=

Looks like 2x128b. Check page 14 aaaaaaand 12.

Ah ye thanks, didn't think to check the optimisation guides!

 

[DrMrLordX]

Ah ye thanks, didn't think to check the optimisation guides!

No problemo. Google pointed me to it when poking around for that info actually. So guess I should thank the behemoth for its help.

 

[soresu]

@DrMrLordX Just checked the A725 optimisation guide, and it looks like it also retains the 2x 128 bit ALU config of A720.

Be curious to see if A730 makes any headway there.

 

[DrMrLordX]

Hmm. That would depend on whether ARM has plans for wider SIMD on the A7xx series.

 

[FlameTail]

@DrMrLordX Just checked the A725 optimisation guide, and it looks like it also retains the 2x 128 bit ALU config of A720.

Be curious to see if A730 makes any headway there.

Qualcomm's Oryon medium core is also 2×128b.

Apple has moved on to 3×128b though.

 

[soresu]

Apple has moved on to 3×128b though.

Yeah I would be surprised if A730 doesn't pump it up a notch.