Graphics Cards - Page 131

NVIDIA's next-gen Ada Lovelace GPU architecture will be fighting AMD's next-gen RDNA 3 GPU architecture later this year, with both GPUs rumored to provide some rather exciting upgrades in performance over the current generation GPUs.

The new NVIDIA AD102 GPU should be the chip that powers the new flagship GeForce RTX 4090 Ti and GeForce RTX 4090 graphics card -- or, according to some rumors the RTX 5090 Ti and RTX 5090 -- with leaker "kopite7kimi" tweeting that AD102 could have over 100 TFLOPs of compute performance.

100 TFLOPs of power from AD102 would make it 2.5x faster than the RTX 3090 Ti, which spits out 40 TFLOPs and 2.8x better than the RTX 3090 with its 35.5 TFLOPs of FP32 compute performance. Remember: pure TFLOPs doesn't mean = gaming performance, but kopite7kimi does say "Lisa and Jensen's competition will give us a 100 TFLOPs gaming war in a few months" with a key takeaway from that: 100 TFLOPs + gaming war.

The GPU rumor game has changed considerably over the last couple of years, in a pandemic-fueled Twitter and YouTuber frenzy that has seen the hype, specs, hype, rumors, and hype reach new bounds. Navi 31 is in the middle of it, with some fresh new leaks that water down AMD's next-gen MCD-based GPU design.

Industry leaker "Greymon55" has tweeted out a fresh rumor, as well as deleting his previous tweet referring to AMD's next-gen Navi 31 GPU having ~92 TFLOPs of compute performance. Greymon55 has tweeted just now, with a simple:

What does that mean? GPU cores on AMD's next-gen RDNA 3-based Navi 3X GPUs, in order: Navi 31, Navi 32, and Navi 33 with 12288, 8192, and 4096 GPU cores respectively.

ASUS has just launched its ROG Strix GeForce RTX 3090 EVANGELION Edition, a new custom GeForce RTX 3090 -- non-Ti -- themed on anime-mecha story EVANGELION.

The new ASUS ROG Strix GeForce RTX 3090 EVANGELION Edition graphics card uses the ROG Strix OC design and is the now fourth variant of the card. ASUS is going with a wicked-looking green/purple design, with black/red throughout -- themed like EVANGELION -- with fantastic results.

You're not going to be able to buy the card anywhere, with CoolPC store the first to take preorders. CoolPC says that they will only have 15 of the new ASUS ROG Strix GeForce RTX 3090 EVANGELION Edition graphics cards for sale, with the Taiwanese store selling them for 63,990 TWD which works out to around $2173 in the US.

EVGA must be having some serious fun in its OC lab, with overclocking legend Vince "KINGPIN" Lucido, playing around with not just one... but two of its crazy custom GeForce RTX 3090 Ti KINGPIN graphics cards.

EVGA is using a very custom design with its GeForce RTX 3090 Ti KINGPIN, with a TGP rated at 516W and a 1200W power limit.

With KINGPIN using the two cards each with 2 x 16-pin PCIe 5.0 power connectors mean for 4 x 16-pin PCIe power connectors into 12 x 8-pin PCIe power adapters. The 1200W+ per GPU means EVGA was using a separate PSU per GPU, something that's required at this (absolutely bonkers) stage of 2400W of power from two GPUs and all those power adapters.

NVIDIA first announced its next-gen H100 Hopper GPU at GTC 2022 (GPU Technology Conference) earlier this year, its next-gen GPU is built on the new 4nm process node at TSMC and rocking a huge 80 billion transistors. Well, that next-gen GPU is now for sale... and it'll cost ya.

Japanese HPC retailer GDEP Advance has the NVIDIA H100 GPU listed for sale, which costs a whopping 4,745,950 yen (or around $36,550) which includes taxes and shipping: the card on its own costs 4,313,000 yen (around $33,200). This is also for the PCIe-based model with passive cooling for server enclosures, so the price of other SKUs is unknown at this point.

As for the NVIDIA H100 GPU, it's based on the new Hopper GPU architecture with NVIDIA using TSMC's 4nm process technology and packing 80 billion transistor into 814mm2 of silicon. The PCIe-based version of the NVIDIA H100 rocks 80GB of HBM2e memory with 2TB/sec of memory bandwidth, rocking on the PCIe 5.0 x16 connector.

Update: Ahhh, the leaker has deleted his tweet... which said "92 TFLOPs" and now is saying that there are toned down specifications of Navi 31 that would radically reduce the FP32 compute power down to around 73 TFLOPs (a 20% reduction).

AMD's next-gen Navi 31 GPU has been shaping up to be a monster so far, but with every rumor the RDNA 3 architecture is looking better and better... to the point where it could topple NVIDIA and its next-gen Ada Lovelace GPU architecture.

In a new tweet, leaker "Greymon55" tweeted something rather simple: "~92T FP32" as a reply to his Navi 31 leaks from November 9, 2021. In that tweet, rumors had Navi 31 at 75T FP32 compute performance, but now the leaker has upped those specs by a considerable 28%.

A wild new GeForce RTX 3090 Ti graphics card has appeared, with Japanese pre-built PC-maker introducing the new Sycom GeForce RTX 3090 Ti G-Master Hyrdo graphics card.

The new Sycom GeForce RTX 3090 Ti G-Master Hyrdo graphics card has a triple-fan cooler on its triple-slot design, but also has a dual-fan AIO cooler. Yeah, not one -- but two coolers -- the first of its kind: rocking 5 cooling fans in total, I'm sure it's going to be running pretty chill.

Sycom is using an OEM triple-fan cooler joined by the higher-end Asetek Hybrid GFX 240mm AIO cooler, which is joined by dual 12cm Noctua NF-A12X12 ULN fans. So, you should expect low GPU and GDDR6X memory temperatures, as well as near-silent operation as the fans only make 12.1 dBA of noise. The company says that their RTX 3090 Ti is up to 15C cooler with the AIO cooler.

AMD has been featured considerably in new leaks from Moore's Law is Dead, with Tom providing some more details on AMD's next-gen Instinct MI300 accelerator.

The new AMD Instinct MI300 accelerator will be the next-gen datacenter-focused GPU, where we should see multiple versions of it hitting the market. In the new leaks, the fact is made clear: "the first thing that needs to be made clear is every source tells me (Tom @ MLID) there are many variants of MI300. The one I am detailing today is just one I'm familiar with".

With the most confidence in the leaks, we're being told to expect that each 6nm die has connections for two stacks of HBM3 memory, with 12+ metal layers in total. The 6nm dies were built to "work with multiple products" that can be stacked on top. The MI300 design itself has a bottom layer that should be occupied mostly with a huge interpose (around 2750mm2 which is insane). On top of the interposer is reportedly a "series of rectangular 6nm tiles that house I/O controllers, IP Blocks, "and likely cache" of around 320-360mm2 per 6nm tile according to Tom.

AMD's next-gen Navi 31 GPU isn't too far away now, scheduled to debut later this year in Q4 2022 -- and will be the first consumer-focused MCD (multi-chip die) GPU for gamers and enthusiasts.

The new Navi 31 GPU and its "MCD" die will be made on 6nm, with some serious performance beef against not just the Navi 21 that it will replace (the GPU that powers the Radeon RX 6900 XT and soon-to-be-released Radeon RX 6950 XT) but also NVIDIA's current, and possibly even next-gen GeForce RTX graphics cards.

AMD's new Navi 31 GPU will be a multi-chip die, similar to the chiplet technology used on the Zen CPU architectures by AMD. In his latest leaks, Tom @ Moore's Law is Dead is continuing what I thought was established -- RDNA 3 will deliver twice the performance over RDNA 2, as RDNA 2 was over double the performance of the original RDNA GPUs.

NVIDIA has provided some new optimizations for the most popular 3D applications in its new April Studio Driver suite, with optimizations for Unreal Engine 5, Cinema 4D, and Chaos Vantage.

The company goes a little further, highlighting that its second-fastest GPU -- the GeForce RTX 3090 -- is faster than AMD's fastest consumer CPU, the Ryzen Threadripper 3990X, in Maxon Cinema 4D 26 Redshift Renderer... by quite a lot.

The GeForce RTX 3090 screams through the render in just 6 minutes and 51 seconds, while the Zen 3-powered 64-core, 128-thread Ryzen Threadripper 3990X takes much, much longer: 65 minutes. That's a significant difference, but it's also a single piece of rendering software. Still, of course, NVIDIA is going to be screaming from the roof tops that its GPU is faster than its competitors... GPU, and CPU.

AMD will be launching its RDNA 2 refresh GPUs on May 10, with ASRock preparing its new flagship Radeon RX 6950 XT OC Formula graphics card.

ASRock's new custom Radeon RX 6950 XT OC Formula graphics card has been spotted with the ECC (Eurasian Economic Commission) and will rock the tweaked Navi 21 XTXH GPU, and 16GB of GDDR6 memory that we should see clocked at 18Gbps. The current Radeon RX 6900 XT has the same 16GB of GDDR6, but it's clocked at 16Gbps.

GPU frequencies will reportedly hit 2475MHz boost, 10% higher than the reference GPU clocks of AMD's upcoming Radeon RX 6950 XT reference specifications. ASRock will be tweaking the hell out of the card like usual, with a fully custom PCB that has 21-phase VRM and 3 x 8-pin PCIe power connectors.

NVIDIA is reportedly testing a new board with its new AD102 GPU, powered by the new Ada Lovelace GPU architecture, and this one is a freaking monster.

The new AD102-powered test board rumor is coming from leaker "kopite7kimi" who said that "there is another full-fat AD102 SKU with 900W TGP, 48GB 24Gbps GDDR6X, 2x16pin and higher frequency". So we have an AD102 GPU, 48GB of GDDR6X @ 24Gbps, a very hefty dual 16-pin power setup and an insane 900W of power.

This sounds like a next-gen Ada Lovelace-powered TITAN RTX, especially with 48GB of GDDR6X memory. Even the newest flagship GeForce RTX 3090 Ti rocks 24GB of GDDR6X, same with the card it replaced: the RTX 3090 also had 24GB of GDDR6X. But dual 16-pin power connectors, 900W of power, and 48GB of VRAM? Yeah, that doesn't sound like a GeForce... at least not yet, maybe in 2024+.

NVIDIA has released its new GeForce Game Ready 512.59 WHQL drivers, which are Game Ready for Dune: Spice Wars, but as usual -- there's much more here.

The new NVIDIA GeForce Game Ready 512.59 WHQL drivers offer supports for the latest updates to Chernobylite, which now offers ray tracing. JX3 Online has NVIDIA DLSS support now, which is supported inside of the new GeForce Game Ready 512.59 WHQL drivers, while the new drivers also offer support for the early access launch of Vampire: The Masquerade - Bloodhunt.

You can download the new NVIDIA GeForce Game Ready 512.59 WHQL drivers here (785MB download).

NVIDIA's new Ada Lovelace GPU architecture will be here later this year, with a fresh rumor from leaker "kopite7kimi" that it will be using PCIe 4.0 and not PCIe 5.0 standard... but what about the process node? 4nm... not 5nm, according to the very latest leaks.

In some of the earlier March leaks, Tom @ Moore's Law is Dead said that the highest-end AD102 GPU would have "60-80% higher performance than Ampere" and was "initially designed as a 400-500W card". We're hearing now that it will be using up to 600W in reference form (more with custom RTX 4090 models).

But as for the process node, NVIDIA is reportedly tapping the custom 4N process node for its new Ada Lovelace GPUs... not the N5 process like previous rumors. Tom actually didn't state the process node, but we knew that NVIDIA was running away from Samsung and back into the (very expensive) arms of TSMC (Taiwan Semiconductor Manufacturing Company).

NVIDIA's next-gen GeForce RTX 4090 graphics card will be using the AD102 GPU based on the new Ada Lovelace GPU architecture, but now we are hearing it will NOT be based on PCIe 5.0... but rather PCIe 4.0 instead.

The new NVIDIA Hopper GPU architecture that was revealed in March 2022 has support for the new PCIe 5.0 standard, but the new consumer-focused Ada Lovelace GPUs including the new GeForce RTX 4090, will be using the current PCIe 4.0 standard.

AMD and Intel's new platforms launching this year -- AMD with its "Raphael" Zen 4-powered Ryzen 7000 series CPUs, and Intel with its 13th Gen Core "Rocket Lake" CPUs -- will be supporting the next-gen PCIe 5.0 standard. Intel already has DDR5 + PCIe 5.0 support on its new flagship Z690-based motherboards, remember.

VASTARMOR entered the Radeon GPU market last year, with the Chinese brand entering the lower-end market at first... but now they're entering the high-end Radeon GPU market with a custom Radeon RX 690 XT graphics card.

The first graphics cards released included custom models for the Radeon RX 6600 XT and Radeon RX 6600, which were followed by the Radeon RX 6700 XT, Radeon RX 6500 XT, and the recent low-end Radeon RX 6400. VASTARMOR has introduced not one, but two new custom RX 6900 XT cards: the Alloy and Super Alloy cards.

VASTARMOR's new custom Radeon RX 6900 XT Alloy and Super Alloy GPUs both rock rather large triple-fan, triple-slot designs... but you've got just 2 x 8-pin PCIe power connectors. GPU clocks, we're looking at 1950/2135/2365 MHz for Base, Game, and Boost GPU clocks respectively (and that's for both the Alloy, and Super Alloy SKUs).

AMD has just launched its low-power RDNA 2 graphics card without much fanfare, with the introduction of the new Radeon RX 6400.

The new AMD Radeon RX 6400 is powered by the Navi 24 GPU, which is made on TSMC's new 6nm process node. AMD's new Navi 24 GPU has 768 Stream Processors, with a base GPU clock of 1923MHz and boost GPU clock of 2039MHz.

4GB of GDDR6 is used, spread out on a smaller 64-bit memory bus which spits out 112GB/sec of memory bandwidth. Now that the card is out in the wild, reviewers have got their hands-on it and thanks to Expreview and DIYZOL, we see the new Radeon RX 6400 sits nicely between the slightly higher-end Radeon RX 6500 XT and NVIDIA's GeForce GTX 1650 (at least in synthetic tests).

NVIDIA's next-gen GeForce RTX 4090 graphics card, which will be powered by the AD102 "Ada Lovelace" GPU, will have a TBP (typical board power) of 600W... and that's in reference form, not in custom AIB form where 600W+ could be in the rear-view mirror.

The rumors are coming from Tom @ Moore's Law is Dead, who said his sources has confirmed the 600W power consumption news. NVIDIA is going to do a "lot of weird things to not lose this generation" to AMD and its next-gen RDNA 3 architecture.

Tom's sources explained: "It seems the top 6000/8000 Lovelace Professional card could use 320-375W. Right now they are testing prototypes at the limits of an EPS 8-pin (375W), or they are even testing some prototypes with a 16-pin connector. It's still 2-slot, and still a blower cooler. So there are limited to what they can cool". The source added: "But this is an unprecedented increase in power for a standard class of professional card".

AMD delayed the launch of its RDNA 2 refresh GPUs from... well... right now... until May 10. But now it seems that the RDNA 2 refresh GPUs are not a refresh, but a replacement of the current Radeon RX 6000 series GPUs.

The news is coming from Hassan Mujtaba at Wccftech, who tweeted: "It seems like AIBs will stop making 6600/6700/6900 XT cards and focus on the new 6650/6750/6950 XT series instead. So RDNA 2 refresh is looking more like a replacement rather than a higher-end alternative to existing models".

There are a heap of AMD Radeon RX 6000 series GPUs in the wild right now, so maybe the delay will help AMD get rid of inventory before the new Radoen RX 6950 XT, Radeon RX 6750 XT, and Radeon RX 6650 XT graphics cards hit shelves across the world.

AMD's new Instinct MI210 accelerator is now out in the wild in Japan, with the PCIe version of the GPU that features 64GB of HBM2e memory costing over $15,000.

The new AMD Instinct MI210 accelerator packs the "Aldebaran" MCM (multi-chip module) CDNA 2 GPU, with 1 MCM die (compared to the 2 MCM dies on the Instinct MI250 and MI250X). AMD's new Instinct MI210 accelerator packs a single GCD with 104 CUs out of the possible 128 CUs on the Aldebaran GPU.

64GB of HBM2e memory is on the AMD Instinct MI210 accelerator, on a huge 4096-bit memory bus that pushes out 1.6TB/sec (1638.4GB/sec) of memory bandwidth. AMD is advertising 181 TFLOPs of peak FP16 compute performance, 45.3 TFLOPs of peak FP32 Matrix, and more.