AIO Workstation Combines 128-Core Arm Processor and Four NVIDIA GPUs Totaling 28,416 CUDA Cores

[AleksandarK]

All-in-one computers are often traditionally seen as lower-powered alternatives to traditional desktop workstations. However, a new offering from Alafia AI, a startup focused on medical imaging appliances, aims to shatter that perception. The company's upcoming Alafia Aivas SuperWorkstation packs serious hardware muscle, demonstrating that all-in-one systems can match the performance of their more modular counterparts. At the heart of the Aivas SuperWorkstation lies a 128-core Ampere Altra processor, running at 3.0 GHz clock speed. This CPU is complemented by not one but three NVIDIA L4 GPUs for compute, and a single NVIDIA RTX 4000 Ada GPU for video output, delivering a combined 28,416 CUDA cores for accelerated parallel computing tasks. The system doesn't skimp on other components, either. It features a 4K touch display with up to 360 nits of brightness, an extensive 2 TB of DDR4 RAM, and storage options up to an 8 TB solid-state drive. This combination of cutting-edge CPU, GPU, memory, and storage is squarely aimed at the demands of medical imaging and AI development workloads.

The all-in-one form factor packs this incredible hardware into a sleek, purposefully designed clinical research appliance. While initially targeting software developers, Alafia AI hopes that institutions that can optimize their applications for the Arm architecture can eventually deploy the Aivas SuperWorkstation for production medical imaging workloads. The company is aiming for application integration in Q3 2024 and full ecosystem device integration by Q4 2024. With this powerful new offering, Alafia AI is challenging long-held assumptions about the performance limitations of all-in-one systems. The Aivas SuperWorkstation demonstrates that the right hardware choices can transform these compact form factors into true powerhouse workstations. Especially with a combined total output of three NVIDIA L4 compute units, alongside RTX 4000 Ada graphics card, the AIO is more powerful than some of the high-end desktop workstations.



View at TechPowerUp Main Site | Source

 

[wolf]

I wonder how they got a total of 28,416 CUDA cores from two cards, I can't find a single card with half that (14208 cores), and to top if off they look like SFF models, where the fastest one the A4000 Ada has 6144 cores per card. Even the A5000 Ada has 12800 cores, x2 is 15600, and two A6000 Ada's with 18176 each would be 36352. Doesn't seem to match Ampere cards either....

hmmmmmmmmmmm

 

[Crackong]

It doesn't look like a capable chassis to house all those hardware with acceptable cooling -> noise as a AIO PC which sits in front of a user

 

[alphaLONE]

It's weird, the ones in the demo unit look like SFF RTX cards like the A2000 or A4000 SFF but those core counts don't map to any GPU out there with 14208 cuda cores... the RTX 5880 comes closest but with 14080 cores it's a bit short. Also those cards are too big to fit in a half height enclosure.

Regardless of that I think it's funny to advertise CUDA cores, FLOPS and TOPS would be more interesting especially since a CUDA core can't do much by itself. An SM would be more of a core imo but I can imagine that 110 is less impressive than 28k

Edit: sniped by wolf haha

 

[Onasi]

@wolf
It’s actually even more bizzare. I am looking at their spec sheet and it’s actually markedly different from the article. They don't mention core count there, but the GPU is listed as a (single) RTX Ada with 20 gigs of ECC GDDR6 (such SKU doesn’t exist to my knowledge) (I was wrong, see below) and the AI accelerators are listed as 3 Ada Teslas, but again, no three Ada Tesla SKUs will add up to 28416 cores. They mention 72 gigs of memory, so my assumption is that there are 3 L4 Teslas, but that doesn’t get us the core count from the news. So I have no idea where the number comes from.

Edit: Actually, no, I’ve cracked the code. The 28,416 comes from adding Teslas AND the GPU. It’s three L4 plus a RTX 4000 Ada. That adds up.
That means that the article is poorly worded to the point of being almost blatantly wrong though.

 

[Daven]

@wolf
It’s actually even more bizzare. I am looking at their spec sheet and it’s actually markedly different from the article. They don't mention core count there, but the GPU is listed as a (single) RTX Ada with 20 gigs of ECC GDDR6 (such SKU doesn’t exist to my knowledge) (I was wrong, see below) and the AI accelerators are listed as 3 Ada Teslas, but again, no three Ada Tesla SKUs will add up to 28416 cores. They mention 72 gigs of memory, so my assumption is that there are 3 L4 Teslas, but that doesn’t get us the core count from the news. So I have no idea where the number comes from.

Edit: Actually, no, I’ve cracked the code. The 28,416 comes from adding Teslas AND the GPU. It’s three L4 plus a RTX 4000 Ada. That adds up.
That means that the article is poorly worded to the point of being almost blatantly wrong though.

So the RTX 4000 serves as a graphics card (plus compute) to internally connect the display and the L4s are just accelerators? If true, the article needs clarifying.

 

[AleksandarK]

So the RTX 4000 serves as a graphics card (plus compute) to internally connect the display and the L4s are just accelerators? If true, the article needs clarifying.

Source (Tom’s) notes that there are two GPUs, so we have to wait for updates to see.

 

[Onasi]

So the RTX 4000 serves as a graphics card (plus compute) to internally connect the display and the L4s are just accelerators? If true, the article needs clarifying.

Seems exactly what’s happening here, yeah. The article also claims the workstation is powered by “two high-end Nvidia RTX professional graphics cards” which is just straight up wrong since, depending on how you count, it’s either ONE graphics card and three accelerators (Teslas are not graphics cards strictly speaking) or four GPUs total.

Edit: @AleksandarK I have literally provided a link to company’s own spec sheet. There is no update needed, Tom’s just wrong in this case.

 

[AleksandarK]

Seems exactly what’s happening here, yeah. The article also claims the workstation is powered by “two high-end Nvidia RTX professional graphics cards” which is just straight up wrong since, depending on how you count, it’s either ONE graphics card and three accelerators (Teslas are not graphics cards strictly speaking) or four GPUs total.

Edit: @AleksandarK I have literally provided a link to company’s own spec sheet. There is no update needed, Tom’s just wrong in this case.

Awesome I didn’t see the spec sheet link. Thanks!

 

[Daven]

Awesome I didn’t see the spec sheet link. Thanks!

Thanks for the article changes, Aleksandark. One last point, you might want to remove the Tom's link and just link to the spec sheet as the source. So far Tom's hasn't updated their article.

 

[Vayra86]

I wonder how they got a total of 28,416 CUDA cores from two cards, I can't find a single card with half that (14208 cores), and to top if off they look like SFF models, where the fastest one the A4000 Ada has 6144 cores per card. Even the A5000 Ada has 12800 cores, x2 is 15600, and two A6000 Ada's with 18176 each would be 36352. Doesn't seem to match Ampere cards either....

hmmmmmmmmmmm

They placed a small mirror between the two GPUs, now they can raytrace double core count
There's probably also a tiny bottle of snake oil somewhere in there.

 

[mechtech]

pffff cuda specs

look at those heart specs 70 bps!!!!!!!!!!!!

like a machine gun!!

 

[Wirko]

As a customer, I would probably want to know a thing or two about upgradability. Physical size of GPUs and the PSU? Available power? Cooling? Fans? Yes, this workstation will likely serve a single purpose but demands will be higher in 2029, and GPU upgrades may be justified by then.

pffff cuda specs

look at those heart specs 70 bps!!!!!!!!!!!!

It's bits per second, can't download from brain faster with the primitive tech we have.

 

[thesmokingman]

Fake till you actually make it.