GPUs & CPUs & Enthusiast hardware: Questions, Discussion and fanboy slap-fights - Nvidia & AMD & Intel - Separe but Equal. Intel rides in the back of the bus.

[The Ugly One]

Eh. The sad thing is that my motherboard is PCIe 3.0. I'd be missing out on performance if I throw something like a newer 9060 in here.

A modern GPU shouldn't be demanding lots more data from the CPU in the same game. The #1 thing that clobbers the PCIe bus is when you don't have enough VRAM, so textures & meshes are constantly getting swapped in and out of main RAM. But whatever you get should have more VRAM than your old card. This is why it's so weird that Intel cards have such a problem with it.

 

[Betonhaus]

 

[JoseRaulChupacabra]

Tech demos are always fun, but we'll see what that's about when it gets here.

I do wonder just how much space it'll save when they get serious about it. Procedurals can look good, but only when you massively complexify the method of generation. With how plain that thing looks, you really will save a lot of space. Also, the very best texture work I've seen were still made with real textures.

 

[The Mass Shooter Ron Soye]

NVIDIA GeForce RTX 5070 SUPER reportedly features 6400 CUDA cores and 18GB memory

Standard 5070 uses GB205 with 6144 of 6400 cores, so you get ~4% more along with the memory correction. TDP increases from 250 to 275W.

NVIDIA also planning GeForce RTX 5070 Ti SUPER with 24GB GDDR7 memory

This is in addition to the long-known 5080 Super 24GB. TDP jumps from 300W to 350W, but with no increase in cores.

Intel Nova Lake performance leak claims 10% single and 60% multi-threaded uplift



">1.1x higher ST" increase is normal, but this seems like a conservative claim for the multi-threaded increase, considering there are double the P-cores and E-cores. These 52 cores could be sandbagged by power limits, a change of pace for Intel.

VideoCardz believes that "Leadership Gaming Performance" implies that Intel has an answer to 3D V-Cache, allowing them to legitimately make that claim.

Intel Nova Lake CPUs may finally bring a 3D V-Cache rival to desktop gaming

Another leaker says 8P+16E+4LPE and 8P+12E+4LPE models will come with "bLLC", Intel's 3D V-Cache competitor. Restricting it to just two single-CPU-tile models is probably a good idea, since it will ensure better supply and less scheduling problems. If Intel beats AMD with this, maybe AMD will feel pressured to charge less than $500 for its single chiplet X3D best seller.

GPD Ryzen MAX 395 “Strix Halo” handheld prototype offers 3x faster graphics than HX 370’s

Finally, there's an alleged Strix Halo gaming handheld prototype. I always thought it was possible, especially after that Strix Halo tablet was announced. But they should probably use the 8-core instead of the 16-core.

 

[The Ugly One]

View attachment 7572334

Can't wait for gamers to start complaining that trees are using "fake meshes" because devs are "too lazy to do proper optimization."

 

[The Mass Shooter Ron Soye]

GeForce RTX 5050 becomes third 8GB desktop Blackwell GPU to launch without reviews
heehee

NVIDIA confirms end of support for GeForce GTX 700, 900 and 10 series after 580 drivers

Samsung Foundry Would No Longer Compete For The “Node Supremacy” Race With TSMC; Announces a Huge Delay With 1.4nm Node

 

[WULULULULU]

GeForce RTX 5050 becomes third 8GB desktop Blackwell GPU to launch without reviews
heehee

NVIDIA confirms end of support for GeForce GTX 700, 900 and 10 series after 580 drivers

Samsung Foundry Would No Longer Compete For The “Node Supremacy” Race With TSMC; Announces a Huge Delay With 1.4nm Node

GG 1080TI users. You've served long enough.

 

[The Ugly One]

GeForce RTX 5050 becomes third 8GB desktop Blackwell GPU to launch without reviews
heehee

NVIDIA confirms end of support for GeForce GTX 700, 900 and 10 series after 580 drivers

Samsung Foundry Would No Longer Compete For The “Node Supremacy” Race With TSMC; Announces a Huge Delay With 1.4nm Node

Since the entire tech industry is projecting infinite technological growth out forever, I wonder what the ultimate ramifications will be. First Intel, now Samsung...it's just a matter of time before TSMC says, "Nah, next node is costing too much to build. You guys are gonna have to wait a few years."

 

[The Mass Shooter Ron Soye]

Since the entire tech industry is projecting infinite technological growth out forever, I wonder what the ultimate ramifications will be. First Intel, now Samsung...it's just a matter of time before TSMC says, "Nah, next node is costing too much to build. You guys are gonna have to wait a few years."

I think Intel is still pretending 18A is going to be amazing. They should break the good or bad news later this year. Then they have a 14A on the roadmap to deal with.

There is Japan's Rapidus attempting to make a fashionably late entrance to 2nm nodes (2027).

Samsung keeps failing to make process nodes with acceptable yields or Exynos chips worth a damn. At least now that the mask is off, they can try to fix it.

TSMC doesn't make its customers wait that long. TSMC has become good at making slow and steady progress while inserting half nodes, quarter nodes, whatever, every year for Apple and others to enjoy. And they have already delayed stuff to keep things on track. TSMC's backside power delivery implementation was delayed from the N2 nodes to A16, while Intel decided to keep it as an optional addition to 18A.

 

[Mister Fister]

10 years is a good run.

 

[Slav Power]

GeForce RTX 5050 becomes third 8GB desktop Blackwell GPU to launch without reviews
heehee

NVIDIA confirms end of support for GeForce GTX 700, 900 and 10 series after 580 drivers

Samsung Foundry Would No Longer Compete For The “Node Supremacy” Race With TSMC; Announces a Huge Delay With 1.4nm Node

It's still impressive that Nvidia supported 700, 900 and 10 series for this long. AMD has dropped support for their equivalents years ago.

RIP 1060 and 1080Ti, you were, and forever will be the GOATs.

 

[Lone Wandering Courier]

RIP my GTX 1060, and RIP my old GTX 960 I still keep incase I would need a backup GPU.

 

[The Ugly One]

TSMC doesn't make its customers wait that long. TSMC has become good at making slow and steady progress while inserting half nodes, quarter nodes, whatever, every year for Apple and others to enjoy. And they have already delayed stuff to keep things on track. TSMC's backside power delivery implementation was delayed from the N2 nodes to A16, while Intel decided to keep it as an optional addition to 18A.

At some point, the next process shrink will be so resource-intensive that there won't be a new market for TSMC to grow into in order to finance it. Collapse of Complex Civilizations stuff.

 

[Slav Power]

RIP my GTX 1060, and RIP my old GTX 960 I still keep incase I would need a backup GPU.

As long as the silicon lives, the GPU lives. My 1060 is still alive and well, albeit in the leftover build that right now has Windows 7 installed on it so it has an even older driver running it. Served me well despite all the abuse I've put it through, so it's about time it gets a well deserved break.

 

[Betonhaus]

and then do we start exploring carbon nanotube circuits or other stuff?

 

[The Mass Shooter Ron Soye]

At some point, the next process shrink will be so resource-intensive that there won't be a new market for TSMC to grow into in order to finance it. Collapse of Complex Civilizations stuff.

I think the "resource" used is EUV lithography, and it would be "intensive" if they had to do octuple-patterning or something. But they are going to move to high-NA EUV... to buy them a couple of years, at only somewhat elevated costs (for example, $400m for a high-NA EUV machine instead of $200m).

If the AI bubble doesn't pop and people continue to take the idea of dumping trillions into new fabs seriously, then there won't be any problem. If it pops, maybe TSMC will have trouble hiking the wafer costs from $30,000 (N2) to $100,000 or something. Maybe they can slow walk progress until another bubble forms.

and then do we start exploring carbon nanotube circuits or other stuff?

It's already being explored, just not being commercialized. I think variations on GAAFETs can be used until the mid-2030s, so another 10 years. Then they might have to take CNTs or other exotic post-silicon technologies more seriously. Whatever it is will need to be in service of 3D chips.

Three-Way Race To 3D-ICs

---

NVIDIA GeForce RTX 5090 D V2 Coming To China, Cut-Down VRAM, To Be Fully Compliant With US Regulations

NVIDIA & AMD’s 8 GB GPUs Face Brutal Market Rejection As Higher VRAM Counterparts Manage To Dominate The Consumer Markets - Mindfactory (Germany) data

AMD FSR 4 Confirmed to Be Coming to PS5 Pro in 2026; Big Chunks of RDNA 5 Are Being Designed Through Engineering Partnership with Sony

 

[Gerelatha]

Honestly, if I can get a Ryzen 9 7900X for a decent price, I'll probably go that route. It just depends on what's available by the time I'm ready to make my new build and if I've made enough payments to shit to where I can save up a little bit for maybe a 9070.

7900X is a shit chip, it’s identical to a 7600X for gaming and the 265KF beats it out for less in productivity

 

[The Ugly One]

I think the "resource" used is EUV lithography

A lot of money gets spent chasing down ever-smaller particles in the clean room.

 

[The Mass Shooter Ron Soye]

I think Intel is still pretending 18A is going to be amazing. They should break the good or bad news later this year.

@The Ugly One
Exclusive: Intel's new CEO explores big shift in chip manufacturing business (archive)
Intel Might Scrap Latest 18A Technology For External Customers In Major Setback, Says Report (archive)

According to Reuters and "people familiar", 18A is cancelled for most external customers, but will still be used by Intel for Intel. They will try to make 14A their grand comeback instead. It could be great (for us) if it means more capacity thrown at Panther Lake, and especially cheap Wildcat Lake products.

AMD Ryzen AI 5 330 powered by “Krackan Point 2” silicon features 4 Zen5 cores and Radeon 820M graphics (archive)

The Phoenix2 die aka "Little Phoenix" was the cheaper Zen 4 APU. Krackan was the cheaper Zen 5 APU with only 4+4 cores and 8 CUs (vs. Strix Point's 4+8 cores and 16 CUs). Now it appears there's an even smaller Krackan 2 die with 2+4 cores and 4 CUs, which is then disabled into the confirmed quad-core with 1+3 cores and 2 CUs. The existence of a smaller Zen 5 APU die had been alluded to but not detailed.

The Krackan 1 prices suck some ass. It's two APUs: the Ryzen AI 7 350 (4+4, 8 CUs) in around $550-1200 laptops, surprisingly paired with RTX 5070/Ti mobile at the high end. Then Ryzen AI 5 340 (3+3, 4 CUs) for $520+, paired with the RTX 5050 mobile in an HP Victus model.

They will have to make a Ryzen AI 5 335 if they want to use that full die. "AI" is in the name of everything because they can, and because even this smaller die retains the full XDNA2 NPU.

G.Skill demonstrates DDR5 CAMM2 memory at 10000 MT/s on ASUS Z890 motherboard


There's discussion about this enabling even bulkier CPU coolers.

 

[Slav Power]

G.Skill demonstrates DDR5 CAMM2 memory at 10000 MT/s on ASUS Z890 motherboard


There's discussion about this enabling even bulkier CPU coolers.

I'm especially interested in LPCAMM2 becoming a standard in laptops, then they can be thin and serviceable, and if they stay thicker then they'll be even more serviceable once all the real estate is freed up for WLAN cards and whatnot. As long as both become a ubiquitous open standard like DIMM/SODIMM then it's the right step forward.