2016+ CPU/GPU News

[jaxa]

9 hours ago, taffernicus said:

By the way arc b390 feels like a breath of fresh air from Intel. Finally a company that still care about consumer. AMD decided to sitting duck 

They had early bird laptops with Arc B390 that were under $1,500, but that pricing was basically a fluke. Real pricing is often above $2,000, comparable to products with AMD's faster Radeon 8050S/8060S graphics. Also well above laptops with faster RTX 5060 (mobile) dGPUs.

Arc B390 is relevant for its power efficiency, not performance/$.

Underneath the 10/12-core Xe models are a bunch of models with 4 Xe cores that are probably similar to AMD's Krackan Point (Ryzen AI 7 350, Radeon 860M, 8 CUs) which can be found in $400-500 laptops.

AMD did drag its feet this year with mobile refreshes, but don't be fooled into thinking Panther Lake is some laptop market savior.

[Bikerdude]

So the manufacture is taking advantage of the situation to try and scam consumers, colour me surprised.

Edited February 18 by Bikerdude

[SeriousToni]

Yeah I hate what currently happens there. Gladly you guys advised me last year to buy a new Gpu before this all was going down the sewers 

[Domark]

I was looking to update just recently and the prices have risen really steeply and inventories are depleted across the board. Could be years before this AI craze blows over, if it ever will.

[Bikerdude]

On 2/18/2026 at 10:31 PM, Domark said:

I was looking to update just recently and the prices have risen really steeply and inventories are depleted across the board. Could be years before this AI craze blows over, if it ever will.

The used market is an only option. In the Uk I have access to a company called CEX (www.cex.co.uk) that sells used PC hardware with a 5yr warranty. If you have something similar, this might be the way to go.

Edited March 4 by Bikerdude

[revelator]

Sadly most predictions say it will take years unless the bubble bursts in which case it will still be years while chipmakers scrape in the losses on the silicon they did not sell sigh...

Even worse it is starting to affect other products as well now such as smartphones tv's automobiles smart fridges gfx cards and a whole slew of other stuff like consoles. Many consumer oriented hardware bussinesses are predicted to go bankrupt.

[revelator]

Ssd's are also taking a big hit many models have seen 80 to 100% price hikes allready. And even mechanical harddrives are getting hard to find with western digital models allready being sold out.

So if you need an ssd ill get one now because next month chances are huge that even the smaller models will be unsellable.

[Bikerdude]

18 hours ago, revelator said:

So if you need an ssd ill get one now because next month chances are huge that even the smaller models will be unsellable.

Yep, already bought a 2TB sata SSD as a backup.

[jaxa]

https://videocardz.com/newz/microsoft-confirms-project-helix-codename-for-next-xbox-mentions-pc-game-support

It's virtually confirmed that the next Xbox will be an x86 gaming PC allowing other storefronts like Steam to work, and probably desktop Windows.

[datiswous]

Smartphones will also become more expensive. Every 2 years for most people.

[taffernicus]

https://www.techpowerup.com/347046/nvidia-reportedly-preparing-9-gb-geforce-rtx-5050-variant-with-gddr7

 

96-bit 3x3x3 GB ( 3 3GB mem chips) = 9 GB compensated with high-speed GDDR7 bandwidth. interesting.

something worth noticing :

https://www.tomshardware.com/pc-components/gpus/microsoft-adds-shader-execution-reordering-ser-in-latest-directx-sdk-for-more-efficient-ray-tracing-intel-arc-b-series-gpus-show-90-percent-performance-uplift

 

Edited March 7 by taffernicus

[taffernicus]

Is there really no significant cost difference between a 2GB module with a configuration of 4 pcs 2GB for 8GB and a 3GB module with a configuration of 3+3+3? I don't know anything about this when it comes to GPUs

The funny thing here is that the RTX 5050 mobile actually uses GDDR7 first, compared to the desktop version

It reminds me of the weird VRAM configurations of the past, such as GPU with 1280 MB, 768 MB of VRAM, and there were even some with 384 MB

* forgot that GTX 2xx era had 896 mb GPU...

 

Edited March 7 by taffernicus

[jaxa]

On 3/7/2026 at 10:04 AM, taffernicus said:

Is there really no significant cost difference between a 2GB module with a configuration of 4 pcs 2GB for 8GB and a 3GB module with a configuration of 3+3+3? I don't know anything about this when it comes to GPUs

The funny thing here is that the RTX 5050 mobile actually uses GDDR7 first, compared to the desktop version

You can find GDDR6 spot prices on a site like DRAMexchange, but they hide other info past a login. I don't know where to find reliable latest pricing for GDDR7. 

It's plausible that the 3 GB modules cost less than 50% more to make than 2 GB, which has been the pattern with memory for decades. Denser memory results in lower costs. But they are in relatively high demand because the 3 GB modules allow you to hit memory configurations like 96 GB on a 512-bit bus for workstation/AI cards.

What's neat is that this alleged RTX 5050 9 GB would have a slight increase in bandwidth over the RTX 5050 8 GB desktop model from switching over to GDDR7. So it's in no way inferior despite having a reduced 96-bit bus, and could be a superior replacement in every way. The extra gigabyte doesn't sound like much, but it could make a difference when there are features that you can turn on like DLSS that use up only a couple hundred megabytes at most.

It would be very funny to see the RTX 5050 9 GB beating cards like the RTX 5060 8 GB in some situations because of the increased VRAM. If it's real, I hope the benchmarkers hunt down examples just for a laugh.

The reason RTX 5050 mobile got GDDR7 while the desktop card got GDDR6 is because GDDR7 is more power efficient. Every little bit counts for mobile. Entry-level gaming laptops are a bigger market than DIY discrete GPUs anyway, just as all laptops are around 2/3 of PC sales.

https://www.tomshardware.com/pc-components/gpus/nvidia-defends-rtx-5050-desktop-gddr6-decision-says-power-efficient-gddr7-is-a-better-choice-for-laptops

[taffernicus]

On 3/10/2026 at 9:28 AM, jaxa said:

You can find GDDR6 spot prices on a site like DRAMexchange, but they hide other info past a login. I don't know where to find reliable latest pricing for GDDR7. 

It's plausible that the 3 GB modules cost less than 50% more to make than 2 GB, which has been the pattern with memory for decades. Denser memory results in lower costs. But they are in relatively high demand because the 3 GB modules allow you to hit memory configurations like 96 GB on a 512-bit bus for workstation/AI cards.

What's neat is that this alleged RTX 5050 9 GB would have a slight increase in bandwidth over the RTX 5050 8 GB desktop model from switching over to GDDR7. So it's in no way inferior despite having a reduced 96-bit bus, and could be a superior replacement in every way. The extra gigabyte doesn't sound like much, but it could make a difference when there are features that you can turn on like DLSS that use up only a couple hundred megabytes at most.

It would be very funny to see the RTX 5050 9 GB beating cards like the RTX 5060 8 GB in some situations because of the increased VRAM. If it's real, I hope the benchmarkers hunt down examples just for a laugh.

The reason RTX 5050 mobile got GDDR7 while the desktop card got GDDR6 is because GDDR7 is more power efficient. Every little bit counts for mobile. Entry-level gaming laptops are a bigger market than DIY discrete GPUs anyway, just as all laptops are around 2/3 of PC sales.

https://www.tomshardware.com/pc-components/gpus/nvidia-defends-rtx-5050-desktop-gddr6-decision-says-power-efficient-gddr7-is-a-better-choice-for-laptops

i'm waiting for rtx 5050 9GB benchmark for real , At first, I got the impression that maybe the extra 1 GB was the answer for " 8 GB might not be enough anymore" haha

 

hmm 1 vram modules is 32 bit , isnt it? i want to grok the calculation

so 512 bit / 32 = 16 mem modules with a clamshell method?

 

Edited March 14 by taffernicus

[jaxa]

50 minutes ago, taffernicus said:

hmm 1 vram modules is 32 bit , isnt it? i want to grok the calculation

so 512 bit / 32 = 16 mem modules with a clamshell method?

RTX 5090 has 32 GB from 16x 2GB 32-bit memory modules, not using the clamshell method.

Clamshell takes each one down to 16-bit, basically forcing every two modules attached to each other to use a 32-bit controller, so you can fit 32 memory modules on 512-bit.

That's how you get RTX PRO 6000 Blackwell Workstation Edition with 96 GB. 32x 3GB 16-bit modules on the 512-bit bus.

[jaxa]

@revelator Here's some "news" for you, X99 (6-core i7-5820K) being tested in 2026:

 

[revelator]

haha nice though i see he had the same problem with the corsair ram modules i had.

they dont play well with the asus x99 so people who want to try with an x99 like that should use gskill modules.

also these modules need to be paired otherwise it is allmost impossible to get all ram sockets running (took me several hours where i had to to try a pair first then the next pair and the next and so on while it was sometimes needed to reverse mount a pair because otherwise it would not boot).

the cpu has enough power for a 3070 rtx without bottlenecking to badly when the cpu runs with a modest overclock.

a 3080 rtx would probably be to much for it though, and a 3080 ti or 3090 rtx would only work ok'ish in 4k resolutions.

the 2080 ti runs really well on it though which is impressive in itself as the board came out about the time of the geforce 980. 

[revelator]

in retrospect these hedt platforms were allways picky with ram due to the quad pumped setup.

even my old sandy hedt platform had problems with ram pairing so its pretty normal but as he also points out scares people away from trying them out even if they get them for a steal.

intel went away from the quad pumped ram platform because as it turned out it really did not give much extra speed and was to problematic for most people.

[revelator]

for those who want to try with such a setup here is what you get.

the asus x99 deluxe has a frak ton of sata 3.1 ports a few sata 6.0 ports an nvme slot (vertical yeah rather funny), a good ammount of usb 3.0 and a few usb 3.1 ports (with the expansion card that came with the board) plus a few usb 2 ports.

the x99 deluxe also sports gigabit wifi bluetooth and no less than 2 gigabit lan connctions.

the board also has many options for overclocking even ln2 (liquid nitrogen cooling). which imo is a bit overkill but you can drive it as high as 6ghz if you dont mind having to squat liquid nitro in an open setup  (dont spill any on your hands!!!).

there was even a later model which included the missing bling with aura and loads of lighting the deluxe 2 ewww.

its a good overclocking board when all is setup correctly (which as explained takes a good deal of patience and some knowledge). it runs win 11 without any problems if you use rufus and tell it to skip the hardware check when creating a bootable usb. 

heres a shot at one running 5.8 ghz https://valid.x86.fr/1zcvfi with the same cpu as mine.

[jaxa]

33 minutes ago, revelator said:

intel went away from the quad pumped ram platform because as it turned out it really did not give much extra speed and was to problematic for most people.

X99 was HEDT, and Intel has continued to have "accessible" quad-channel offerings. Cascade Lake in 2019, then Sapphire Rapids-WS / Xeon W-2400 (e.g. 16-core w5-2465X w/ quad-channel DDR5-4800 ECC support at $1389 MSRP).

I see one w5-2465X for $745, and another for $999 on ebay. Not exactly a healthy market, but eh.

Edited April 9 by jaxa

[revelator]

oh i seem to have heard they would skip quad channel sometime after x299 my bad if they are still around.

yeah hedt boards are expensive as hell but ouch those prices...

[revelator]

hmm seems nova lake might return with a hedt model, intel actually seems to have ditched it for some time as the last hedt model was from 2019. i suspect it will cost a minor yacht though.

[revelator]

https://www.ebay.at/itm/377093018524?_skw=x299+motherboard&itmmeta=01KNQW3ZKN029NFZB3GJC6KBQS&hash=item57cc7ed39c:g:qtUAAeSwx~hp1BaV&itmprp=enc%3AAQALAAAA8GfYFPkwiKCW4ZNSs2u11xBBKfmsE8vStmcdoFo7BPw5OaLzzuyhrN3HEj6pnC30uMYvCYIzrnX5eSye4Z2b%2BzdVm8iDgUHxisy3QC1ety%2FS1NOINDsldBt7qvNTop4cCT9Uf0gAUcCaCB1vbjTnr3DxeD6v8Szd6GyfqrlorGEzD6%2FFxN8diQz%2BC6NStdwnrTp8szEqu%2Bc06%2BKb0LhwNC36iqvr4g5LOyt2ABNM%2BW4L5%2BshEgjdsZygiTGJzcPha3SYxvTimM8yuohHX4FZvoPsnHh7jjzcN8SmEMotA2IdfypnOugZwhQU0T1h2JZKCw%3D%3D|tkp%3ABk9SR4b6j_ytZw

hmm 170 euro not to bad though its an older platform by now it was an absolute monster performer with a core i9 10900x.

should also support win 11.

[jaxa]

13 minutes ago, revelator said:

hmm seems nova lake might return with a hedt model, intel actually seems to have ditched it for some time as the last hedt model was from 2019. i suspect it will cost a minor yacht though.

The Nova Lake-S flagship is practically HEDT without being true HEDT. 52 cores (16 P-cores, 32 E-cores, 4 LPE-cores), with the Ps and Es split onto two compute tiles (a bit like Ryzen) and the LPEs separately on the SoC tile or somewhere.

Furthermore, the rumor mill is saying there will be a dual-bLLC option, that is Intel's answer to AMD's 3D V-Cache and specifically chips like the recently announced 9950X3D2 ($900 MSRP). Basically an enlarged compute tile (not 3D stacked cache) with 144 MiB of L3 cache for 8P + 16E instead of 36 MiB, and two of those, so 288 MiB total L3 cache. I would guess $1000-1200 for the price. Zen 6 X3D will match the L3 cache with 48 MiB on standard CCD, 144 MiB on X3D CCD, and 288 MiB for a 24-core Zen 6 successor to the 9950X3D2.

And these would all be dual-channel parts. Intel's will probably have official DDR5-8000 support, but 10,000+ MT/s CUDIMMs could be used. It adds a clock driver to push stable speeds up.

6 minutes ago, revelator said:

hmm 170 euro not to bad though its an older platform by now it was an absolute monster performer with a core i9 10900x.

should also support win 11.

Supporting Windows 11 is barely a feature.

Edited April 9 by jaxa

[revelator]

the weird thing with the cpu's for x299 was that pretty much all of them were refreshes of the old skylake processor just with upped core ammounts. they generally perform barely ok at stock speeds but when overclocked things get interresting.

sadly none of them could beat the ryzen 3900XT at the time not even when overclocked...

one key factor here was that they were crippled by having a hard limit of a 65 watt power draw which in most cases caused them to be unable to reach max turbo speed.

in games they delivered though so not all bad but also a bit odd.