Review Intel Arc B580 review: The new $249 GPU champion has arrived

[thestryker]

i cant wait to see the specs of the B770 .. if it drops at $400usd and we see specs of 4080 /7900xtx or around that we have a solid cheap card !!

I think at the very least the B770 will be 7900xt 4070ti lvls of performance ..

If i see 7900xtx or 4080 which may be super optimistic i may swap to intel from my current 7900xtx

Rumors point towards the biggest Battlemage being 32 Xe2 cores which is +60% over the B580. The best I've seen currently for B580 is around a 10% performance increase from overclocking. So if they are leveraging more of the performance potential I could see B770 maybe matching a 4070 Super, but really I suspect it would be closer to 4070 performance. At $350 I feel like that'd be a good deal with 16GB VRAM and even a $400 initial price wouldn't be that bad.

 

[Pierce2623]

Tech equivalent for me means primarily equivalent resources turned into performance, which in a GPU design is mostly the RAM in size and bandwidth and and the electric power to burn.

There the B580 and the RTX 4070 get to use the same means, 12 GB of 192-bit VRAM at ~500GByte/s and 200 Watts of power but wind up in a very different performance class.

Pretty near all tech reviews of the B580 chose to stick with the 'Intel recommended price equivalent' RTX 4060, which delivers somewhat less performance but uses a 128-bit bus instead of 192-bit, 277 GByte/s instead of 500 GByte/s bandwidth, 8 instead of 12 GB of RAM and 115 Watts of power instead of 200, so siginificantly less resources for results that punch much higher than the linear equivalent of the resources given.

I made a mistake there, since the only one who compared the B580 against an RTX 4070 was Phoronix (and the first review to come out at all), I looked at that Phoronix overall results, where its 200 vs 96, which qualifies as "double" but failed to notice that he compared it against an RTX 4070 Super, which gets 220 Watts instead of 200 and a slightly better bin of the same chip. But even at 180 vs 96 it wouldn't be off by far.

Too many cards, too easy to miss, I'm sorry for that mistake!

Yet it doesn't change the overall picture much, in terms of technology the B580 proves how woefully behind Intel is compared to NVidia, which sounds like "envy" in Spanish for good reasons.

Intel may try to save face by offering it at that price, but current listings in Europe are at €325 which isn't €250.

The RTX 4070 is €550, the RTX 4070 Super is €650 and you can choose which one you want to compare it to.

The first may fall slightly short of "double" the second hits it per Phoronix, but perhaps Linux isn't the same story as Windows, even if it's the same games.

I consider throwing "dishonest" at me rather harsh and would argue that it is Intel who is trying very hard to tilt the scales.

And unless that price stays at half of what NVidia charges for a technically similar card, converting the same Wattage into significantly less performance is an uneasy decision to make. And that's not counting software aspects.

Of course, if it's good enough, it's good enough, at least you can't fall into the trap of buying more than you need within the Battlemage family currently.

But from a tech angle, Battlemage remains just shockingly bad, very much a Bulldozer vs a Core. And that means Intel stays far away from being a serious contender, which I'd like it to be.

But does that matter? Especially if energy efficiency isn't that high on your list?

It mostly depends on if Intel can make enough money from Battlemage to continue and perhaps become more efficient, too.

The price of the external resources, VRAM, PCBs etc. and power is the same for all, the ASIC the main differentiator. B580 and RTX 4070 don't differ vastly in transistor count, 19B vs. effectively 25B for the 4070 bin, but even if Intel could get theirs for half the price that NVidia has to pay, that doesn't mean they can maintain the advertised price gap if NVidia decides to squeeze a bit. Nor does half price seem very likely, not after Pat got in a spat with TMSC's founder.

If NVidia were to lower prices ever so slightly, Intel would turn from blue to red numbers while NVidia is still raking it in and we all know who can afford what.

I bought an A770 perhaps six months after launch, because I wanted to replace a GTX 1060 in a 24x7 box that is crammed with storage and needs something compact and low-power. That A770 mostly just failed to work with my dual DP KVM or DP generally: only HDMI was ok, so I sent it back after benchmarking and got the compact and efficient PNY RTX 4070 instead, which just works fine, but unfortunately at twice the money.

Perhaps a year later I bought an A770m in a i7-12700H NUC when that NUC became so cheap the A700m with 16GB VRAM and 500GB/s bandwidth was basically included for free (€700 for the Serpent Canyon). At that price it just wasn't a risk, more of a freebie.

It turned out ok for gaming, and it's still used in the family.

This summer I got a Lenovo LOQ ARP9 laptop for €750 that includes an RTX 4060m as well as a Rembrand-R 8-core Zen 3. This laptop tends to use half the Wattage at the wall the NUC uses, yet runs even Unreal 5 games like ARK Survival Ascend enjoyable enough with software magic, where the Serpent Canyon simply fails to exceed single digit FPS, perhaps only because ASA doesn't support XeSS, but who knows?

I'd say I try my best to give Intel a fair chance, in CPUs and GPUs. And if they fail it typically means I have to spend more, so why would I want to tilt scales?

If the B580 really drops to €250, I might get one to test, even keep if, it gets the job done. But if I'll be able to get another RTX 4070 for not that much more, used or whatever, I'd pick that, because when the price for Nvidia is right, the value is better. And with all those boxes in the home, efficiency is money, too.

Unless the b580 and 4070 occupy the same price bracket, everything you wrote is meaningless. I can’t believe people are claiming it’s a bad thing that Intel is selling is more advanced node silicon at a given price. It shows that they are willing to cut into their profit margins if it allows them to launch a compelling product.

 

[Pierce2623]

Rumors point towards the biggest Battlemage being 32 Xe2 cores which is +60% over the B580. The best I've seen currently for B580 is around a 10% performance increase from overclocking. So if they are leveraging more of the performance potential I could see B770 maybe matching a 4070 Super, but really I suspect it would be closer to 4070 performance. At $350 I feel like that'd be a good deal with 16GB VRAM and even a $400 initial price wouldn't be that bad.

Plus we’ve also seen that Intel is giving more AI and ray tracing performance than Nvidia with a given level of rasterization. A $400 4070 tier card with the same general characteristics as the b580 would be an awesome deal.

 

[jlake3]

Unless the b580 and 4070 occupy the same price bracket, everything you wrote is meaningless. I can’t believe people are claiming it’s a bad thing that Intel is selling is more advanced node silicon at a given price. It shows that they are willing to cut into their profit margins if it allows them to launch a compelling product.

It’s worrying because there’s talk that they’re not making any profit margins on these, and it’s not actually a sustainable business model to sell bigger chips for less money when you and your rivals are both using the same foundry.

The launch inventory isn’t quite a paper launch… but it’s not looking much better. And if the economics are bad, restocks may be more of a trickle than a market-shifting flood.

 

[JarredWaltonGPU]

It’s worrying because there’s talk that they’re not making any profit margins on these, and it’s not actually a sustainable business model to sell bigger chips for less money when you and your rivals are both using the same foundry.

The launch inventory isn’t quite a paper launch… but it’s not looking much better. And if the economics are bad, restocks may be more of a trickle than a market-shifting flood.

I really feel Intel rushed this out a bit, in part because it has a good idea of what's coming from AMD and Nvidia and it needed to get Battlemage out the door first. And the economics really aren't looking good.

• Intel: $250 B580 graphics card, probably a $150~$175 BOM cost, launched in Dec 2024
• AMD: $270 7600 graphics card, probably a $100~$120 BOM cost, launched in May 2023
• AMD #2: $330 7600 XT, probably a $120~$140 BOM, launched in Jan 2024
• Nvidia: $300 4060, probably a $120~$140 BOM, launched in June 2023

So, Intel is at least a year behind AMD's launch, and a year and a half behind Nvidia's launch, with a product that's ~10% faster at 1440p but basically tied at 1080p, for $50 less. I think more driver tuning gets it to perhaps 20% faster. That's a better value now, but long-term it may not age well.

A lot will hinge on what AMD and Nvidia do with the next gen. I'm hoping RX 7600 ends up the anomaly and that AMD actually does something useful on the budget-mainstream end of the spectrum. And by "useful" I mean it needs at least 12GB of VRAM, and it should be at least 30% faster than the 7600. And if AMD does that, for $250~$300? It's going to be better than the Arc B580.

From Nvidia, I really and truly hope it has digested the fact that 8GB graphics cards are basically dead now. If it wanted to do a $200 RTX 5050, sure, I'd bite. But $250 and above absolutely must have a minimum of 12GB. It was pathetic when RTX 4060 got cut to 8GB, and doubly so for the RTX 4060 Ti to get 8GB (and the 16GB card only partially addressed the shortcomings). Whether it does 3GB GDDR6 or GDDR7 on a 128-bit interface, or actually puts a 192-bit interface back on the lower spec cards, I don't care. But the 4060 keeps looking worse and worse as time goes by, and the 4060 Ti 8GB is a joke.

But as @thestryker notes, with the amount of money Nvidia is making on AI, it's hard to imagine the company will want to do more than the minimum effort for budget to mainstream offerings (meaning $250 to $600 prices). At the same time, AI in the consumer space is also picking up, and AI needs more VRAM capacity. 8GB also isn't enough. So again, that suggests even Nvidia should want a wider interface. Because that's how you get more VRAM capacity, and you can double it with clamshell memory for the professional and AI market. So, please put a minimum of 12GB on all future dedicated GPUs.

 

[JarredWaltonGPU]

Rumors point towards the biggest Battlemage being 32 Xe2 cores which is +60% over the B580. The best I've seen currently for B580 is around a 10% performance increase from overclocking. So if they are leveraging more of the performance potential I could see B770 maybe matching a 4070 Super, but really I suspect it would be closer to 4070 performance. At $350 I feel like that'd be a good deal with 16GB VRAM and even a $400 initial price wouldn't be that bad.

This. Exactly. 32 Xe-cores at best means 60% more performance than the B580. And I'm 99% sure it will be a 256-bit interface, which means 16GB (or 32GB, but that's definitely not going to happen outside of a potential "professional AI" part). So, 60% more compute but only 40% more bandwidth (if it uses 20 Gbps GDDR6), means 50% more performance at best.

According to my GPU benchmarks, 10% faster than 4060 plus another 50% would put 1440p ultra performance at 101 fps. That's right at the level of the RTX 4070 and RX 6800 XT, a bit below RTX 3080 Ti and RX 6900 XT, and well below the 7900 GRE and 4070 Super. The RTX 4080 and 7900 XTX are a solid 30% faster than the 4070. There's no chance Intel gets there with a 32 Xe-core part.

If Intel doubled the Xe-core count of B580 so that it has 40 Xe-cores and a 384-bit interface, that could match the 4080 and 7900 XTX. But that would also require a price of $600+ for sure, and I don't think many people would be ready to pay that for the third string GPU company.

 

[ilukey77]

If Intel doubled the Xe-core count of B580 so that it has 40 Xe-cores and a 384-bit interface, that could match the 4080 and 7900 XTX. But that would also require a price of $600+ for sure, and I don't think many people would be ready to pay that for the third string GPU company.

$600 is still cheaper than current 4080 prices ,,

People will pay if the price is cheaper !!

Scan forums youtube videos almost every time i see videos on GPU's people are asking where is my $300usd 4080 like inflation isnt a thing or the global pandemic didnt show everyone that the prices have gone CRAZY but for every 1 million people complaining about prices there is 1 million more ready to pay the overpriced cost !! ( hence why the prices are crazy )

But there is ALOT of customers out there who are sick of Nvidia's overpriced rubbish and to some degree AMD's prices too !!

So you offer cheaper quality performance people will buy it ..

If intel can sustain a loss to a degree in their GPU market and if they keep producing quality products with real generational gains ( unlike there CPU's ) keep developing their drivers ( 1st gen ARC showed yes they had driver issues but over time they fixed and improved them )

They just need to start getting that market share off the other 2 ..

Picture this you want to buy a GPU Nvidia are selling high end or overpriced rubbish AMD are scared of high end so aiming at mid to low tier but charging near the same as Nvidia's junky mid to low tier ..

Enter Intel with a close to performance cheaper mid to low tier alternative!

Do that for a few years and people will be thinking Intel when they go to upgrade their card the next gen !!

I was quite impressed with my ARC 770 16gb card for a 1st gen newcomer to the market !!

I was equally impressed with Intel's willingness to be transparent and constantly work on upgrades and improve their drivers ..

Ive been spouting about going a 5090 next year but im not sure if its worth my money when it going to be over $4000aud near to $5000aud maybe im better to wait on my 7900xtx and see what intel do with celestial !!

 

[King_V]

... AMD prices their products too close but also because their software was/is such garbage that it chased off people like me who swore they'd never go nVidia...

You're trying to hedge your bets here. So, either you're saying their software IS garbage, in which case you're engaging in disinformation, or you're saying it WAS garbage, but you're trying to convince people to make decisions as if it still is garbage.

In either case, you're pushing a false narrative.

 

[Pierce2623]

$600 is still cheaper than current 4080 prices ,,

People will pay if the price is cheaper !!

Scan forums youtube videos almost every time i see videos on GPU's people are asking where is my $300usd 4080 like inflation isnt a thing or the global pandemic didnt show everyone that the prices have gone CRAZY but for every 1 million people complaining about prices there is 1 million more ready to pay the overpriced cost !! ( hence why the prices are crazy )

But there is ALOT of customers out there who are sick of Nvidia's overpriced rubbish and to some degree AMD's prices too !!

So you offer cheaper quality performance people will buy it ..

If intel can sustain a loss to a degree in their GPU market and if they keep producing quality products with real generational gains ( unlike there CPU's ) keep developing their drivers ( 1st gen ARC showed yes they had driver issues but over time they fixed and improved them )

They just need to start getting that market share off the other 2 ..

Picture this you want to buy a GPU Nvidia are selling high end or overpriced rubbish AMD are scared of high end so aiming at mid to low tier but charging near the same as Nvidia's junky mid to low tier ..

Enter Intel with a close to performance cheaper mid to low tier alternative!

Do that for a few years and people will be thinking Intel when they go to upgrade their card the next gen !!

I was quite impressed with my ARC 770 16gb card for a 1st gen newcomer to the market !!

I was equally impressed with Intel's willingness to be transparent and constantly work on upgrades and improve their drivers ..

Ive been spouting about going a 5090 next year but im not sure if its worth my money when it going to be over $4000aud near to $5000aud maybe im better to wait on my 7900xtx and see what intel do with celestial !!

5090s are going to be sucked up by AI startups at inflated prices and gamers will have to like pay nearly double the price of a 4090.

 

[palladin9479]

From Nvidia, I really and truly hope it has digested the fact that 8GB graphics cards are basically dead now. If it wanted to do a $200 RTX 5050, sure, I'd bite. But $250 and above absolutely must have a minimum of 12GB. It was pathetic when RTX 4060 got cut to 8GB, and doubly so for the RTX 4060 Ti to get 8GB (and the 16GB card only partially addressed the shortcomings). Whether it does 3GB GDDR6 or GDDR7 on a 128-bit interface, or actually puts a 192-bit interface back on the lower spec cards, I don't care. But the 4060 keeps looking worse and worse as time goes by, and the 4060 Ti 8GB is a joke.

It's not the memory size, it's the memory bandwidth. You even demonstrated as much.

https://www.tomshardware.com/reviews/nvidia-geforce-rtx-4060-ti-16gb-review

Going from 8 to 16GB on an even stronger card didn't move the needle at all at 1080p

And slightly moves the needle at 2160p

And since 1440p is the target for these cards

What cripples the 4060 series is not having "only" 8GB of VRAM, as demonstrated above having 1 Peta Byte of VRAM wouldn't of changed anything. Having only 272 / 288 GB/s of memory bandwidth is what cripple's them vs other cards. The 4070 has "only" 12GB of VRAM, 25% less then the 4060 TI 16GB but 504 GB/s of memory bandwidth. The Intel B580 has 456 GB/s using a six chip 192-bit interface. If anything I would think that the memory bus might actually be overkill for the GPU compute capability, essentially the inverse of the 4060 situation.

Currently 128 bit memory interface means four 32-bit channels, currently GDDR6 is largely made in 8 or 16Gb sizes, we really don't want to use 8s. GDDR7 is also looking to be in 8/16Gb sizes though at 25~33% higher speeds, so I expect we'll still see a plague of cheap 4 chip 128-bit bus's.

 

[thestryker]

It's not the memory size, it's the memory bandwidth.

This is just wrong. Memory bandwidth only matters to the extent that the GPU can utilize it. Here's some obvious evidence of 8GB VRAM simply not being enough for all titles:

https://www.techpowerup.com/review/intel-arc-b580/22.html

 

[palladin9479]

This is just wrong. Memory bandwidth only matters to the extent that the GPU can utilize it. Here's some obvious evidence of 8GB VRAM simply not being enough for all titles:

https://www.techpowerup.com/review/intel-arc-b580/22.html

I don't use TPU, Toms or GN.

Both did a dive into it when the 4060 ti 16GB came out, with the exact same result. It didn't change anything except in a few outliers.

I've previously done a big writeup and the short version is that you only need enough VRAM to hold all assets in the current scene. That's it, anything else isn't even being referenced.

But here is the catch, modern game engines do not remove old objects until you run out of room, or they do a resource flush. A 16GB card can have 4-6GB of stuff used now and 8GB of stuff used 30 min ago. You will see 12-14GB of utilization, but most is junk.

Only games with extremely large scenes full of lots of different textures are going to need more then 8GB. They do exist but currently the HW reviewers don't know how to test for this.

 

[YSCCC]

$600 is still cheaper than current 4080 prices ,,

People will pay if the price is cheaper !!

Basically this. For those who can squeeze money out, most of them pull the trigger on the 80/90 class and the 7900XTX in AMD camp, basically none or very few will go down a bit to the GRE for less ram or to the 4070 class.

But for most consumers, PC is a second thought for gaming, which not necessarily be clicking everything on at ultra, for those who allocate their disposable income, or their pocket money/xmas gift, there is a hard price cap to divide between various components, so for those who have price cap at or below the 4070, if the BMG card is the best option under the price bracket, they will just use that instead.

I have had the same constrain at my Sandy Bridge era build, those days I opt for the 1060 6GB and it literally served till 3070Ti in my Alder lake build which subsequently only swarped out to a 14900k and is servicing today. Yea C/P is a thing but if the C/P value is so bad at the start it isn't a bit condieration anymore

 

[thestryker]

I don't use TPU, Toms or GN.

Both did a dive into it when the 4060 ti 16GB came out, with the exact same result. It didn't change anything except in a few outliers.

I've previously done a big writeup and the short version is that you only need enough VRAM to hold all assets in the current scene. That's it, anything else isn't even being referenced.

But here is the catch, modern game engines do not remove old objects until you run out of room, or they do a resource flush. A 16GB card can have 4-6GB of stuff used now and 8GB of stuff used 30 min ago. You will see 12-14GB of utilization, but most is junk.

Only games with extremely large scenes full of lots of different textures are going to need more then 8GB. They do exist but currently the HW reviewers don't know how to test for this.

Here's a simple fact: There are games that need more than 8GB of VRAM. Some you can see like the TPU chart where the FPS difference is obvious, and others do things like unload textures. Theoreticals don't matter when there are real world examples. Just because 8GB VRAM shouldn't be a problem doesn't mean that it isn't a problem.

 

[lchien52]

am i missing something? Title says $249 GPU but all the pricing says $390... that's not even close. or is that the pre-release high demand price gouging at work?

 

[abufrejoval]

I really feel Intel rushed this out a bit, in part because it has a good idea of what's coming from AMD and Nvidia and it needed to get Battlemage out the door first. And the economics really aren't looking good.

• Intel: $250 B580 graphics card, probably a $150~$175 BOM cost, launched in Dec 2024
• AMD: $270 7600 graphics card, probably a $100~$120 BOM cost, launched in May 2023
• AMD #2: $330 7600 XT, probably a $120~$140 BOM, launched in Jan 2024
• Nvidia: $300 4060, probably a $120~$140 BOM, launched in June 2023

So, Intel is at least a year behind AMD's launch, and a year and a half behind Nvidia's launch, with a product that's ~10% faster at 1440p but basically tied at 1080p, for $50 less. I think more driver tuning gets it to perhaps 20% faster. That's a better value now, but long-term it may not age well.

"rush" sounds odd, when they are at least two years late to the game and still not in a similar performance category.

I'd say they took the only shot they got, because they won't look better in Spring. Nvidia will be able to drive them below sustainable easily, perhaps even AMD, because Intel simply doesn't deliver enough performance for what they consume in resources.

I'd wish you'd add the RTX 4070 to that BOM vs. retail list, because that's how NVidia would gauge their competition: similar resources, much better performance therefore able to charge a vast premium.

A lot will hinge on what AMD and Nvidia do with the next gen. I'm hoping RX 7600 ends up the anomaly and that AMD actually does something useful on the budget-mainstream end of the spectrum. And by "useful" I mean it needs at least 12GB of VRAM, and it should be at least 30% faster than the 7600. And if AMD does that, for $250~$300? It's going to be better than the Arc B580.

From Nvidia, I really and truly hope it has digested the fact that 8GB graphics cards are basically dead now. If it wanted to do a $200 RTX 5050, sure, I'd bite. But $250 and above absolutely must have a minimum of 12GB. It was pathetic when RTX 4060 got cut to 8GB, and doubly so for the RTX 4060 Ti to get 8GB (and the 16GB card only partially addressed the shortcomings). Whether it does 3GB GDDR6 or GDDR7 on a 128-bit interface, or actually puts a 192-bit interface back on the lower spec cards, I don't care. But the 4060 keeps looking worse and worse as time goes by, and the 4060 Ti 8GB is a joke.

I'd disagree ever so slightly on the 12GB minimum, because it depends on what you pay and what you play and your resolution.

With that €750 Lenovo LOQ its RTX 4060m at 1920x1080 is quite ok using DLSS even for the most demanding game in my library, which is ARK Survival Ascend, an early Unreal 5 title. And it works very well with the vast majority of my titles, where I am a few years behind the latest and greatest (Tomb Raider variants, FarCry 3-6). But it's also doing great on FS2024 and Hogwarts Legacy.

The 4060 is obviously not great at 3 or 4k, but that's just not the market it's designed for and the reason Intel's choice of putting the B580 and the RTX 4060 in the same slot "tilts the scales" and isn't a technical fit: it takes a 4070 for fairly competent 3k, while even 4k often isn't looking too bad with upscaling.

I'm recycling some hardware from my shelves these days and I've just tried a GTX 980ti with 6GB of 330GB/s bandwidth on a Kaby Lake i7-7700k on Win 11 24H2 with my game library and again at 1920x1080 that's still a fairly competent combination you might pick up for next to nothing somewhere 2nd hand, especially with FS2024, which doesn't really doesn't use more than a single CPU core, which it needs to be fast (same story for two decades).

Likewise I've tested an RTX 2080ti with 11GB of 660GB/s VRAM and 250 Watts of max power: I'd hazard it might be the better value than the B580 if you can get it cheaper. Somehow DLSS works well enough even with the 1st generation tensors now and it's much better than I recalled ...if you don't try to go beyond 3k.

While the original BOM of the 2080ti was vastly higher, I find it telling that at 250 vs 200 Watts, it's delivering very near B580 results: it makes purported efficiency gains of Battlemage sound somewhat hollow.

But as @thestryker notes, with the amount of money Nvidia is making on AI, it's hard to imagine the company will want to do more than the minimum effort for budget to mainstream offerings (meaning $250 to $600 prices). At the same time, AI in the consumer space is also picking up, and AI needs more VRAM capacity. 8GB also isn't enough. So again, that suggests even Nvidia should want a wider interface. Because that's how you get more VRAM capacity, and you can double it with clamshell memory for the professional and AI market. So, please put a minimum of 12GB on all future dedicated GPUs.

Even if NVidia is making less money with dGPUs than with HPC or AI, they are still making a tidy profit without giant new investments: why should they stop? And let a future competitor grow where it might eventually turn uncomfortable? They've been around the block often enough and at the receiving end of Chipzilla. They don't even need to sell at Intel or AMD prices, nor do they need to lower prices for long to stamp out a blue flame. Nvidia doing laptop APUs certainly shows their appetite to eat Intel market share and then some.

As to 8GB of VRAM I'd argue [from the very different consumer perspective] that there might be a niche depending on price. There is a lot of mobile-on-desktop hardware around these days that offers vastly more CPU performance than needed for games at 3k or below, but that has iGPUs still a little too weak. If you can get a dGPU at a matching price point, the result can beat a console in performance but gain you a PC at a similar price, pretty much like the LOQ ARP9 at €750, which is way cheaper than anything Strix Point currently, obviously more luggable than mobile, but very much a console in terms of gaming performance. If you have four kids with friends and want to play together, a lesser system is better than none.

And like at the time when I argued that the 16GB RTX 4060 might have a point when you urgently want to play with low precision LLMs on a low power and money budget, today I'd argue that 8GB RTX 4060 might have a point at budget 1920x1080 gaming. And creating that extra SKU evidently paid off for them, even if reviewers decried it as a mad ripoff.

Trying to have everything at once, computing, gaming, and AI is where you wind up paying nearly the same as having a separate machine for each use case. They have data analytics and hype to drain your pockets, they still can't entirely avoid surplus inventory, so sometimes you can get better value than they planned for.

 

[-Fran-]

I don't use TPU, Toms or GN.

Both did a dive into it when the 4060 ti 16GB came out, with the exact same result. It didn't change anything except in a few outliers.

I've previously done a big writeup and the short version is that you only need enough VRAM to hold all assets in the current scene. That's it, anything else isn't even being referenced.

But here is the catch, modern game engines do not remove old objects until you run out of room, or they do a resource flush. A 16GB card can have 4-6GB of stuff used now and 8GB of stuff used 30 min ago. You will see 12-14GB of utilization, but most is junk.

Only games with extremely large scenes full of lots of different textures are going to need more then 8GB. They do exist but currently the HW reviewers don't know how to test for this.

Here's a simple fact: There are games that need more than 8GB of VRAM. Some you can see like the TPU chart where the FPS difference is obvious, and others do things like unload textures. Theoreticals don't matter when there are real world examples. Just because 8GB VRAM shouldn't be a problem doesn't mean that it isn't a problem.

You are both right. It' neither or the other in isolation: it's both.

A 1050 with 32GB of VRAM would be as useless as a 4090 with 1GB of VRAM in the same vein as a 1050 with 2TB/s bandwidth is still useless alongside a 4090 with 1GB/s bandwidth.

I do like palladin's point of "how to test this". To which I'll have to say: it's already known how to test. Anyone with a 4090 looking at textures know the answer. Anyone looking at polygon count/complexity knows the answer. Having more assets in the VRAM just helps with less trips to RAM, and those trips depend on how fast your connection to it is. It's been demonstrated that cards with 8GB will offer a worse experience than those with higher capacities when loading textures and keeping things in memory for longer so you don't have to flush and re-aquire them later on. This is not me talking theoretically, but already proven by HUB and other media outlets I can't remember for now (I think Anand did one such piece a long way back).

In any case, you at least have to agree: more VRAM and bandwidth is not a bad thing, even when the GPU itself may seem like it can't "process" more than it does. Much like system RAM, having more is just allowing the GPU to age better.

Regards.

 

[abufrejoval]

You are both right. It' neither or the other in isolation: it's both.

A 1050 with 32GB of VRAM would be as useless as a 4090 with 1GB of VRAM in the same vein as a 1050 with 2TB/s bandwidth is still useless alongside a 4090 with 1GB/s bandwidth.

A GTX 1050 perhaps not, but I'd take an RTX 4060 with 32GB of VRAM for LLM work, if the price was below €1000.

Because a 4060 can do low precision LLMs using INT2-8, even FP8, so at 250GB/sec it's still at least 2-4x as fast as modern CPUs and DDR4/5 RAM because of bandwidth.

A 4090 with 1GB indeed is indeed a much harder sale; I'd spend a few € to extra for 48GB instead of the 24 mine has, but not quite what they actually charge for the professional variant.

But coming back to the GTX 1050 with 32GB of VRAM: I've heard about people using GPUs for databases, which seems completely insane, when you consider just how light databases tend to be on compute.

Until you realize that in places like high-frequency trading latency is so important, people put databases into GPU memory soley for the bandwidth and the maximum time it takes for a full scan.

 

[-Fran-]

A GTX 1050 perhaps not, but I'd take an RTX 4060 with 32GB of VRAM for LLM work, if the price was below €1000.

Because a 4060 can do low precision LLMs using INT2-8, even FP8, so at 250GB/sec it's still at least 2-4x as fast as modern CPUs and DDR4/5 RAM because of bandwidth.

A 4090 with 1GB indeed is indeed a much harder sale; I'd spend a few € to extra for 48GB instead of the 24 mine has, but not quite what they actually charge for the professional variant.

But coming back to the GTX 1050 with 32GB of VRAM: I've heard about people using GPUs for databases, which seems completely insane, when you consider just how light databases tend to be on compute.

Until you realize that in places like high-frequency trading latency is so important, people put databases into GPU memory soley for the bandwidth and the maximum time it takes for a full scan.

I didn't mention it explicitly, but the GPU must support the RAM nuggets in their chip as well, since you can't really come up with memory capacities which aren't "supported". You can indeed mod a 4060ti to 32GB if the memory nuggets exits (4GB ones?), but they don't, so there's that as well. Both 4060's are tied to a 128bit BUS.

Also: I used extremes to illustrate the point in the context of gaming, so sorry to say, but your tangent is kind of moot. AMD and nVidia know that as well, and they have solutions (kind of) for those specific niche use cases.

Regards.

 

[palladin9479]

Here's a simple fact: There are games that need more than 8GB of VRAM. Some you can see like the TPU chart where the FPS difference is obvious, and others do things like unload textures. Theoreticals don't matter when there are real world examples. Just because 8GB VRAM shouldn't be a problem doesn't mean that it isn't a problem.

Here is another fact, people who buy the low tier cards do not play on Ultra.

I understand why we use Ultra as a baseline when doing benchmarks to get ideas of relative performance, but it's not reasonable for end users. People who buy 4060 (should of been called a 4050) and 4060ti or other lower tier cards are going to play on high / custom for the better experience.

Having only 8GB is not the problem for the 4060, you could give it 1PB of VRAM and you'd get the same result. This has been demonstrated when they released the 16GB 4060 TI.

 

[palladin9479]

I didn't mention it explicitly, but the GPU must support the RAM nuggets in their chip as well, since you can't really come up with memory capacities which aren't "supported". You can indeed mod a 4060ti to 32GB if the memory nuggets exits (4GB ones?), but they don't, so there's that as well. Both 4060's are tied to a 128bit BUS.

Also: I used extremes to illustrate the point in the context of gaming, so sorry to say, but your tangent is kind of moot. AMD and nVidia know that as well, and they have solutions (kind of) for those specific niche use cases.

Regards.

So GPU VRAM is built kinda differently then system VRAM even though they use the same general concepts.

Every GPU chip has a memory bus, which is really just a cluster of memory management units. The industry "standard" size for DRAM is 32-bits per chip. For a GPU like the 4060 with 128-bit memory bus, that is four 32-bit channels each going to a single 16Gb chip. Four of them together is 64Gb which converts to 8GB. If we were to add two more 32-bit interfaces to the GPU we get six channels for 192-bit bus with 96Gb (12GB) of VRAM. Add two more channels for eight and now we are at 16GB of VRAM and 256-bit bus. Now all DRAM has the ability to be daisy chained, you can have more then one chip per channel and this is what they did with the 16GB 4060Ti. The memory clock speed might have to go down a little bit to compensate for the longer signal path to the second chip though. For GDDR6 this is called clamshell mode, and it's more that each chip gets half the bus.

That is why VRAM can not be arbitrary, it's always sized in discrete chunks of number of memory channels * density of chips. Currently GDDR6 is made in 8Gb (1GB) and 16Gb (2GB) varieties, GPU makers are using the higher density node with device manufactures using the lower density one. GDDR7 was going to be the same but Samsung kinda broke the mold with a 24Gb (3GB) chip.

Samsung changes the GPU market with new 3GB GDDR7 chips - OC3D

Samsung is ready to change the GPU market with their new 3GB GDDR7 memory modules. Great news for PC gamers and AI enthusiasts.

overclock3d.net

I'm hoping next generation cards use these new chip densities, but nVidia is not very consumer orientated right now so it might be up to AMD and Intel.

 

[palladin9479]

I do like palladin's point of "how to test this". To which I'll have to say: it's already known how to test. Anyone with a 4090 looking at textures know the answer. Anyone looking at polygon count/complexity knows the answer. Having more assets in the VRAM just helps with less trips to RAM, and those trips depend on how fast your connection to it is. It's been demonstrated that cards with 8GB will offer a worse experience than those with higher capacities when loading textures and keeping things in memory for longer so you don't have to flush and re-aquire them later on. This is not me talking theoretically, but already proven by HUB and other media outlets I can't remember for now (I think Anand did one such piece a long way back).

That's not accurate, the value is only how much VRAM has been allocated not what is currently in use. As I've stated modern engines do not evict old resources until they absolutely have to. Actually it's not really the game engine but WDDM that manages this. You have a lot more graphics memory then you do VRAM, run dxdiag and check Display 1 if you want to see how much you have available. In my case I have 32GB of system memory and a 3080 hydro (12GB) card, WDDM will report me as having 27.7GB of available graphics memory, meaning there can be a total of 28GB of graphics resources loaded at once.

Now that region is obviously split into two, 11.8GB display memory and 15.9 of system memory and it's WDDM's job to manage which of the two places graphics objects are located. I'll use a simple five room game level that is using dynamic loading (since nobody likes loading screens) to illustrate.

Game starts - loads 1 GB of common assets.
Player enters Room 1, 4GB of graphics assets are loaded.
Player enters Room 2, 2GB of new graphics assets are loaded that didn't exist in Room 1.
Player enters Room 3, 2GB of new graphics assets are loaded that didn't exist in either of the previous rooms.
Player enters Room 4, 2GB of new graphics assets are loaded.

How much total memory is "allocated" vs "needed" in this scenario? GPU-Z and other tools will show 11~12GB of total VRAM "in use" or "allocated", because there is indeed 11~12GB of total graphics resources loaded by WDDM. How much is "needed" at any one time is only 4~6 GB. Players inside Room 4 do not need the assets from Room 2 that aren't present in Room 4 and when the player is moving between rooms the engine is asking WDDM to pre-load the new assets, which is then moves into graphics VRAM. WDDM won't unload pre-existing assets unless it runs out of VRAM because they may be referenced again. Think of unused VRAM as a rudimentary graphics cache.

Now that we understand that, what is the resource that consumes the most VRAM, textures. This is where graphics pre-sets become very important. Ultra setting is almost always this unreasonably large texture sizes, as in the textures are larger then the display resolution. On a large 4K display with a high powered card these could create slightly better detail then textures half their size, but on a 1080~1440p display with a low powered card they are utterly useless. Reducing just the texture size alone will dramatically reduce the VRAM needed at any one moment in time.

More then 8GB of VRAM is only necessary on the larger more expensive cards using higher resolutions and ultra texture sizes. Nobody buying a 4060 / 4060Ti is going to be playing on a high resolution display with "ultra" settings.

 

[Gururu]

Likewise I've tested an RTX 2080ti with 11GB of 660GB/s VRAM and 250 Watts of max power: I'd hazard it might be the better value than the B580 if you can get it cheaper. Somehow DLSS works well enough even with the 1st generation tensors now and it's much better than I recalled ...if you don't try to go beyond 3k.

While the original BOM of the 2080ti was vastly higher, I find it telling that at 250 vs 200 Watts, it's delivering very near B580 results: it makes purported efficiency gains of Battlemage sound somewhat hollow.

Well, 2080 ti running for $500-800 dollars right now, so I think your argument is hollow. I am not sure I would want to pick up a used burned out data miner for cheaper either.

 

[thestryker]

Here is another fact, people who buy the low tier cards do not play on Ultra.

No that's actually not a fact that's just more nonsense from you rather than actually admitting you're wrong. The example I gave has playable frame rates that are much better with sufficient VRAM in the case of nvidia and goes from unplayable to good on AMD. If the reason people have to lower settings is due to insufficient VRAM capacity that's a problem which shouldn't happen.

I understand why we use Ultra as a baseline when doing benchmarks to get ideas of relative performance, but it's not reasonable for end users. People who buy 4060 (should of been called a 4050) and 4060ti or other lower tier cards are going to play on high / custom for the better experience.

Having only 8GB is not the problem for the 4060, you could give it 1PB of VRAM and you'd get the same result. This has been demonstrated when they released the 16GB 4060 TI.

So you're just ignoring the evidence given because you "know you're right" which is just absurd. It'd be really nice if you just stopped posting lies so I could stop responding to make sure people don't buy into your nonsense.

Here's more evidence that 8GB is a problem:

View: https://youtu.be/ecvuRvR8Uls?si=zFjGvTukpHE97SYe

 

[tommtajlor]

Thx for the review! I find it odd that the A580, the predecessor is not in the charts. Then i checked tpu: in price/perf actually there is 3% regressuon compared to A580. Pfff. Thats not good.