Holy cow the v100s go for like $1k on ebay now... They were selling for $100 or less a couple years ago.

Given the rush to dig MI50s and V100s out of the trash and press them into service, 2024 seems like future space alien technology.

&gt; I&#x27;m starting to believe in QualcommSomeone drop the YouTube link of their robot presentation

&gt; tri-channel on-package (!) memory good for 228 GB&#x2F;sWhy is on-package memory impressive&#x2F;better in this case? It&#x27;s one channel more than your typical consumer x86 system (192-bit bus vs 128-bit) which is something but you could technically have quad channel with two LPCAMM2 sockets?

Thats relieving!I&#x27;m starting to believe in Qualcomm! The Snapdragon X2 Elite Extreme is alas only available in a single system, with it&#x27;s tri-channel on-package (!) memory good for 228 GB&#x2F;s. But it&#x27;s a beast. The core is just good, even in the mainstream form in other systems: some of the best single core performance you can get, shy of Apple.And they are upstreaming their stuff much better! They hired a bunch of really good known FOSS developers and it seems to be working!Now they are talking up High Bandwidth Compute as their next big thing. And I want to believe, even if it makes me nervous that they too intend to leave consumer computing completely in the dust. <a href="https:&#x2F;&#x2F;www.qualcomm.com&#x2F;news&#x2F;onq&#x2F;2026&#x2F;11&#x2F;qualcomm-hbc-validated-silicon-ai-inference" rel="nofollow">https:&#x2F;&#x2F;www.qualcomm.com&#x2F;news&#x2F;onq&#x2F;2026&#x2F;11&#x2F;qualcomm-hbc-valid...</a>This is, as commented, at M4 Max levels of GB&#x2F;s. It is not impressive. It has some uses! But if this was their modern opening salvo I would not be impressed.

If the specs seem outdated, it’s because this chip launched in 2024. It’s available at cloud providers including Cirrascale

Not if the customers cannot rent a H200 (because they have been told not to rely on foreign gpus)

I guess they get their customers on scale, can you get hundreds of h200?

At $5 per hour these are absolutely terrible value. You can rent an H200 for that price and an H200 is much better in every way. Weird

Buy for the low, low price of - nope. Not for you.<pre><code> Cirrascale (8-card Ultra server, 128 GB each, 512 GB system RAM): 
 $4,699&#x2F;mo (monthly term) or $3,759&#x2F;mo (annual term)

 Cloud GPU rental (per-GPU): from $5.15&#x2F;GPU&#x2F;hr
</code></pre>
ref: <a href="https:&#x2F;&#x2F;cloud-gpus.com&#x2F;qualcomm-ai-100-ultra&#x2F;" rel="nofollow">https:&#x2F;&#x2F;cloud-gpus.com&#x2F;qualcomm-ai-100-ultra&#x2F;</a>

Less than I expected for this level of GPU power.It&#x27;s a trend I like.

8,750 in stock at 1 cent, too. The more expensive 47 and 63 cent variants are all sold out.

Yeah, I clicked that too. But knowing qualcomm they obviously cost an arm and a leg. Still interesting how much.

According to their own website, they cost 1 cent.

Honestly, let me buy one, it&#x27;s many times faster than the minipc I run local ai on so if the price is right I&#x27;m in.

Curious how the memory bandwidth actually holds up under sustained inference loads compared to the spec sheet numbers.

Yeah, this thing is unusual. It resembles a consumer GPU but using normal DRAM with more channels, so then you get more RAM but less memory bandwidth than GDDR&#x2F;HBM, which is the same thing Apple does. And the TDP is pretty attractive for what it is.But what is that amount of SRAM for? It simultaneously seems like a lot, because SRAM is so expensive, but still not very much, because what can you fit in 0.5GB? Is there a specific workload that needs something that looks like this?

Sort of interesting design It has half a terabyte of SRAM on the accelerator with a 128 GB of RAM replacing HBM with SRAM but keeping regular RAM.

(I know it&#x27;s you comparing it to the M4 Max, not them, but:) it&#x27;s Qualcomm, so it&#x27;ll be compared to the M4 Max, while being slower than a M2 Max.

So it&#x27;s an M4 Max Mac Studio, RAM-wise, with a GPU&#x27;s worth of compute?

Seems like a reasonable general-purpose card.

&gt; I’m not sure how much effect you get from higher TOPS with such low bandwidth though.Raised arithmetic intensity. It would be quite good for LLM prefill, and decode would want to adapt by pushing parallelism higher (saving on memory bandwidth by reusing the same layer weights for multiple streams) or, if unable to do so (e.g. due to KV cache constraints) by resorting to speculative decoding (where the expensive &#x27;validate&#x27; step handles tokens in parallel within a single stream). It does not make the platform inherently unviable. It would also do well at other tasks such as image&#x2F;video generation, which is already quite compute bound.

This could be very fast at running embedding models (if it supports the right modes)

Considering that memory bandwidth is the most important for inference for most neural networks, that matters more.

Apple M6 has 80-100 TOPS.Qualcomm Cloud AI 100 Ultra has 800 TOPS, almost 10 times more.

Modern inference stacks use different GPUs for decode versus prefill. This is a GPU suitable for prefill.

Ok so it’s significantly worse than Apple M6 and we don’t know if it’s cheaper or notEdit: Ah sorry you’re right, the bandwidth is lower but the TOPS is a lot higher. I’m not sure how much effect you get from higher TOPS with such low bandwidth though.

If you are going down this path, MI350P and RTX PRO 6000 Blackwell is what you would want. A lot more compute a lot more vram bandwidth.

But there are already plenty of products for most purposes.

For most purposes, this seems reasonable? This looks like a single-height card, so fitting four of these in a workstation or server would give 512 GB RAM and 2 TB&#x2F;s total bandwidth. That&#x27;s pretty attractive for a lot of current LLMs, even if one wishes for 1 TB capacity at 5 TB&#x2F;s bandwidth.

&gt; I would think players would like to bring something new to the table instead of bringing what was okay two years ago or so.Speaking of... from [0]:<pre><code> Introducing Qualcomm Cloud AI 100 Ultra
 High performance and cost-optimized cloud AI inferencing for generative AI
 Written by
 Rashid Attar
 Nov 13, 2023

 Today we announce the Qualcomm Cloud AI 100 Ultra, a new member of our portfolio of cloud artificial intelligence (AI) Inference cards, purpose-built for generative AI and large language models (LLMs). ...

</code></pre>
[0] &lt;<a href="https:&#x2F;&#x2F;www.qualcomm.com&#x2F;news&#x2F;onq&#x2F;2023&#x2F;11&#x2F;introducing-qualcomm-cloud-ai-100-ultra" rel="nofollow">https:&#x2F;&#x2F;www.qualcomm.com&#x2F;news&#x2F;onq&#x2F;2023&#x2F;11&#x2F;introducing-qualco...</a>&gt;

You can&#x27;t do quality inference on such a toy. It&#x27;s a myth.

5 year old GPUs are still at crazy prices.Plenty of businesses want in local inference and are priced out.

This is all the CPU companies playing catch up with Nvidia and AMD.The thing you are describing is much closer to AMD and Nvidia latest generation server offerings.See MI455X for the absolute cutting edge and MI350P for something you can use outside of a datacenter.

It may be a semantic thing but I don&#x27;t think bandwidth pooling works like that? Unless your blocks are distributed perfectly between cards there&#x27;s going to be data passed between each card and from VRAM to compute. Correct me if I&#x27;m wrong though

8 of those cards give you 1TB of ram and 4.4TB&#x2F;s

Aren&#x27;t these specs already outdated? Seems meh. I would think players would like to bring something new to the table instead of bringing what was okay two years ago or so.Give me 1TB RAM and 5TB&#x2F;s or something that exciting.

Qualcomm Cloud AI 100 Ultra – 150W AI Accelerator Card with 128GB at 548GB/s