Discussion Qualcomm Snapdragon Thread

[SarahKerrigan]

?
ARM definition of "HPC" is everything not 30W DPU.

ARM counts "cloud computing" separately. They care about the supercomputer market. "Neoverse V-Series include high-performance floating-point and vector instruction support, with features like SVE/SVE2, Bfloat16, and Int8 MatMul delivering strong HPC and AI/ML application performance" isn't talking about general-purpose server.

My argument style is just me being real.
GA is a very specific thing and none of the V2 parts (Amazon, GOOG, etc) are there yet.

Grace is. You can buy it off the shelf. End users are buying it in quantity. But we're right back to "HPC doesn't count because reasons."

(Oh, and "z tier availability"? Come on. Far more Grace has moved in the last quarter than IBM sells Z systems in an entire three-year generation.)

Christ, this is tiring. I'm done.

 

[adroc_thurston]

ARM counts "cloud computing" separately.

It's also Neoverse V.
See the recent bracketing of IPs from the V3 announcement.

Grace is.

That's a funny toy with no mem capacity.

You can buy it off the shelf.

Where can I get a PowerEdge with it?
I get it, you hate AMD, but no need to twist reality like that.

(Oh, and "z tier availability"? Come on. Far more Grace has moved in the last quarter than IBM sells Z systems in an entire three-year generation.)

?
Not as the Grace-Grace CPU, rofl.

Christ, this is tiring. I'm done.

Good for you.

 

[SarahKerrigan]

That's a funny toy with no mem capacity.

Neoverse V2 is still not GA anywhere. Doesn't exist.

Hmm.

Where can I get a PowerEdge with it?

By that standard, Altra never hit GA, and therefore doesn't exist.

I get it, you hate AMD, but no need to twist reality like that.

I don't care enough about AMD to hate them. (Or Intel, ARM, Nvidia... I care about IBM stuff a fair bit, which is why I do hate IBM.)

You seem upset though.

 

[adroc_thurston]

Hmm.

Yeah.
That's not GA.
Call me when AWS lets me rent it with arbitrary DRAM allotments!

I don't care enough about AMD to hate them.

You spend half the last page lashing on Zen4.
Again, great entertainment value but you're adding nothing to the discourse.

 

[Mahboi]

OEMs pretty obviously disagree.

That would be MS honestly.
They're pushing it all into the pit to got on The Ride™.

 

[jpiniero]

It's a cute marketing point but SOC eval has many more points of contention.

Right... there's AI, AI, AI and AI.

 

[adroc_thurston]

Right... there's AI, AI, AI and AI.

Nope.

 

[FlameTail]

Some parts are a meme

Are you a meme?

/s

 

[adroc_thurston]

Are you a meme?

No, I'm a magical girl (or a Scub Coral researcher, depends on your POV).

 

[xpea]

Where can I get a PowerEdge with it?

it's coming

Loading…

www.dell.com

• The Dell PowerEdge XE9680 server is being enhanced to support NVIDIA H200 Tensor Core GPUs, as well as the newest NVIDIA B100 and B200 Tensor Core GPUs, delivering a clear performance path to sustain demanding organizations and multiple use cases.
• Dell is also announcing support for the NVIDIA GB200 Grace Blackwell Superchip as a standalone platform for specific deployments. More information will be available in the 2H of this year.

BTW last quarter Nvidia sold more Grace CPUs in DC than AMD Epyc. But Grace2 is where the fun really starts...

 

[xpea]

That's a funny toy with no mem capacity.

One can apply the same reasoning and say that Epyc is a "funny toy with no bandwidth capacity"

 

[adroc_thurston]

it's coming
https://www.dell.com/en-us/blog/power-the-enterprise-ai-experience/

These are not CPU systems.
Please learn to read before shilling.

One can apply the same reasoning and say that Epyc is a "funny toy with no bandwidth capacity"

?
Grace with usable capacities is 512b@6.4GT/s, nothing all that useful.

 

[jpiniero]

This has a useless GPU tumor

But AI...

 

[adroc_thurston]

But AI...

Not useful.
I'm not in business of swindling venture capitalists out of money.

 

[jpiniero]

I'm not in business of swindling venture capitalists out of money.

That's Silicon Valley in a nutshell.

 

[adroc_thurston]

That's Silicon Valley in a nutshell.

Not quite, but during any hype cycle, yes. Indeed.

 

[poke01]

Not useful.
I'm not in business of swindling venture capitalists out of money.

What? You don’t like selling people “AI” powered Pins. Your too nice…

 

[adroc_thurston]

What? You don’t like selling people “AI” powered Pins. Your too nice…

So true, gonna use MS-mandated Recall™ rootkit.
Long live Satya Nutella.

 

[podspi]

Anyway...

Tbh, I am very excited for mainstream ARM systems. If Microsoft manages to get a multi-ISA platform off the ground we're potentially entering an era of amazing competition in client cpus.

I pre-ordered an Omnibook X w/ 32gb of RAM. Really conflicted over whether the NPU will actually be useful VS. It won't be ENOUGH. Seems like we're entering a period of really rapid improvements in INT4/8 matrix operations throughput. I got spoiled, being above to use my Haswell era desktop for 10 years...

What is everybody's thoughts on the different elite skus? Most of the laptops I've seen seen to be the 78 variant, particularly if battery life is a focus of the model. I have a hunch these are TDP limited, and this is just OEMs choosing the reasonable part for the purpose. I've used throttlestop for years to do the same on my 1065g7. It's incredible the extra battery life I was able to achieve without any real (subjective) performance loss.

 

[adroc_thurston]

What is everybody's thoughts on the different elite skus

"Qualcomm can't bin".
Fin.

 

[podspi]

Maybe so. The models with the higher elites usually boasted ~45W TDPs. I've never seen that highlighted before, but I think architecture +TDP will be more useful going forward.

 

[adroc_thurston]

The models with the higher elites usually boasted ~45W TDPs

That's cooler capacity iirc.

but I think architecture +TDP will be more useful going forward.

QC is always annoying with disclosures like that.

 

[xpea]

These are not CPU systems.
Please learn to read before shilling.

?
Grace with usable capacities is 512b@6.4GT/s, nothing all that useful.

You wrote exactly:

Where can I get a PowerEdge with it?

Where did you say "CPU" systems? None. But yeah I know, it's your typical Goal Post Moving tactic when you (often) loose the argument

?
Grace with usable capacities is 512b@6.4GT/s, nothing all that useful.

You know very well what I mean. The Epyc puny bandwidth to the accelerator with the garbage tiers PCIe Gen5 x16 at 128 GB/s vs 900GB/s on Grace. You know, the exact reason why nobody uses Epyc+MI300X for training because this pathetic off die bandwidth that kills all the performance...
And regarding how useful Grace total mem capacity is, well the market has spoken and customers disagree with you. Let me say it again: Nvidia sold more Grace CPUs in DC than AMD sold Epyc CPUs last quarter. That says it all...

 

[adroc_thurston]

Where did you say "CPU" systems? None. But yeah I know, it's your typical Goal Post Moving tactic when you (often) loose the argument

We're talking V1/2 CPU IP.
Please do the shilling bit correctly if you want to entertain me.

You know very well what I mean. The Epyc puny bandwidth to the accelerator with the garbage tiers PCIe Gen5 x16 at 128 GB/s vs 900GB/s on Grace

?
CPU mem attach b/w?

You know, the exact reason why nobody uses Epyc+MI300X for training because this pathetic off die bandwidth that kills all the performance

No one uses GH200 for training either, it's all H100 8GPU HGX boxes with 400G networking of various flavours.
Please do the green menace shilling bit correctly if you want to entertain me.

 

[podspi]

Exactly, that's my point. Like mobile chipsets, there is no model-TDP map, it's whatever oems want to do. The M4 makes this really clear. Give it enough cooling and it'll run much faster.

I hope this marks the beginning of sane power limits in windows laptops. Don't get me wrong, nothing wrong with a DTR, but at least offer the option to reign in power use to claw back some battery life.