Intel just CRUSHED Nvidia & AMD GPU pricing

This cute little guy is the Intel Arc

Pro B50. This is its bigger brother, the

Arc Pro B70. It just came out. Now, the

B50 didn't need any extra power cables.

It got all its power straight from the

PCIe bus. Nice, simple, civilized. The

B70 is a little bit less civilized. This

one needs power, but it also brings 32

GB of VRAM and comes in under $1,000.

And because local AI tends to turn into

a hardware addiction, I'm plugging in

four of them. That gives me 128 gigs of

total VRAM, which means I'm going to

need a computer that can actually power

all four, which is exactly the kind of

sentence Local AI makes you say out

loud. Now, I already did a video on the

B50 last year, and it was the most bang

for the buck GPU you can get in 2025.

So, now I want to know if the B70 could

do the same thing for Local AI in 2026.

Because just for comparison, a single

RTX5090

from Nvidia also has 32 GB of VRAM, but

that comes in in just under $4,000 at

this point. So, yeah, we're about to

find out. So, of course, to do this

comparison properly, I did what any

reasonable person would do. I bought

more GPUs. This is the Nvidia RTX Pro

4000 Black. Well, check out that price

tag right there. Not flexing, just

saying that this is retailed two times

higher than this card and it's 24 gigs

of memory, not 32. However, it is GDDR7,

so it's the newer VRAM. And it has 672

GB per second of memory bandwidth. It's

also very skinny, just like my wallet.

Wow, I can see you through the fan. the

comparison physically. The RTX Pro 4000

is a much skinnier card and it also

gives you the four display port ports.

So, that's a single card slot with

pretty impressive bandwidth and very

impressive price, too. By the way, I got

it at MicroEnter was $1699, not 1999.

Not sponsored by MicroEnter, but still a

lot of money. Next, I got AMD's Radeon

AI R9700.

And up until now, this was the only GPU

you can get

with 32 gigs of memory for about $1,300.

This is 32 gigs of GDDDR6 and 640 memory

bandwidth, so not as high as the Nvidia

one. So, they match on VRAM, but this

card is about $350 more per GPU. Now,

the B70 comes in at the lowest price.

However, it also has the lowest memory

bandwidth at 68. So yeah, this has been

pretty heavy on the wallet, which is why

I'm thankful to the sponsor of this

video. So these days, I'm always

flipping between models. GPT for

research, cloud for coding, nanobanana

for image generation, VO cling, and

runway for video, six tabs, six bills,

and counting. Enter chat LLM teams. One

dashboard houses every top LLM and route

Olympics, the right one. GPT Mini for

ultra fast answers, Claude Sonnet for

coding, Gemini Pro for massive context.

They recently added Gemini 3 and GPT 5.1

the moment they dropped. Create

professional presentations with graphs,

charts, and deep research detailed

content. Need human sounding copy?

Humanize rewrites text to defeat AI

detectors. Need visuals? Pick Frontier

or open- source models. Nanobanana

Midjourney Flux for images. Magnific

upscaling plus VO WAN and Sora for video

all builtin. You also get Avac's AI deep

agent to pretty much do anything. build

full stack apps, websites, reports with

just text prompts and deploy them on the

spot. They have Abaca's AI desktop,

which is the brand new coding editor and

assistant that lets you vibe code and

build productionready apps. And the

kicker, it's just $10 a month, less than

one premium model. Head over to

chatlm.abacus.ai

or click the link below to level up with

Chat LLM teams. I just finished testing

the B60 and depending on the software

stack that you're running, you get very

different results. Here's Cickle, which

is running Llama CPP. You can run Llama

CPP and that'll utilize either Sickle,

which is a Intel specific stack. It

gives you really good performance there.

By the way, this is the Quen 34B model

at Q4KM

quantization. We're getting 1,0 tokens

per second for prompt processing here.

And since these are professional level

cards, it's also a good idea to test

them with a higher concurrency. So, this

is concurrency of one, which kind of

simulates chatting with the thing,

right? But if you have a concurrency of

four, which is leaning towards more of a

agentic workflow or multiple users at

the same time, then we come down to 898

tokens per second here. And that just

shows that Llama CPP is not the best for

higher concurrency throughput. Now, for

token generation, we're getting 66

tokens per second here for C1. And for

C4, which is concurrency 4, we're

getting 83. So just a little bit higher

than your single. And of course, I did

Llama CPP for Vulcan, which is a

cross-platform approach. And Vulcan did

better in certain scenarios, like for

example, prompt processing, which is

kind of surprising. 1,162 tokens per

second there for Single. But Sickle did

do better for token generation, 66

tokens per second versus Vulcan at 44.

However, the best the best performance

we got was from VLM, of course. Look at

that huge difference right there. This

is meant to run on professional GPUs

like this with high concurrency and

throughput. And this is the 4bit AWQ

quantization for VLM. 8,118 tokens per

second here for concurrency of one token

generation 67. So not that much higher

for single concurrency for token

generation than cickle. But look at the

scaling when it comes down to

concurrency of four. 215 tokens per

second for token generation. All right,

that's just a review of the B60. Let's

see what happens when we do the B70

along with these other ones. So, I'm

going to kick things off with one B70

and this Nvidia 4000. I do like that one

slot feel. Ah, that fresh new GPU smell.

Nvidia SMI. We've got 145 watts

available on this thing and 24 GB. So, I

kicked off VLM cuz I'm not doing Llama

CPP at this point. VLM is the way to go

on these kinds of machines and these

kinds of GPUs. I'm running both of these

now and I'm pointing to Quen 34B. This

is the full BF-16 version. And the way

VLM likes to work is it takes up as much

memory as possible. So, it's going to

fill up all 24 GB on this board and all

32 GB on that board just because it

likes to have extra room for KV cache.

And I'm using a tool called Llama Beni

by Yugger over here. It's an open source

project and it's a really nice tool for

doing this kind of benchmarking. Don't

confuse this with just Llama Bench. This

is Llama Beni. All right, it's

different. So, what's different about it

specifically is in the read me, so you

can go read it. But just to give you a

brief overview is first of all, it's

going to work with any kind of server,

not just Llama CPP, including VLM. And

second, it allows me to prefill the

context so that we can test filled

contexts also, not just empty context.

And that is very useful. All right, here

we go. We're going to do a little race,

but it's not going to be instant because

I got it on two different windows. Boom.

And boom. And I

It's funny the sounds that these things

make cuz they all have coil wind. And

also the coil wind is different based on

what model you're running and what

concurrency you're running. And probably

other parameters will affect it too. So

if you have a really really keen

hearing, you'll be able to tell me

exactly what model is running, how many

concurrence requests were processing.

I'm just kidding. But maybe AI will be

able to tell the difference. And here we

go. We got our first results. 56 tokens

per second token generation on the B70

compared to 51 over here. So the B70

beats it out just by a little bit. Ooh,

that's hot. Um, also PROM processing