GPU Performance Hierarchy and the Best Graphics Cards, November 2022

With graphics cards finally back in stock, I've put together this comprehensive GPU performance hierarchy of the past two (now starting on three) generations of graphics cards. These are the best graphics cards around, but it's also important to know what you're getting. I've benchmarked all of these cards, at multiple resolutions and settings. Here's how they stack up across a test suite of 15 games.

I've also listed current GPU prices, which have improved substantially over the past few months. Most AMD graphics cards are now selling at or below their official MSRPs, and Nvidia's fastest GPUs (RTX 3080 Ti, RTX 3090, and RTX 3090 Ti) are also well below their MSRPs. The RTX 4090 and RTX 4080 have also launched, alongside the various Intel Arc GPUs. Here's the current list, updated regularly as new GPUs arrive.

GPU Benchmarks and Performance Hierarchy

Note: Click the column headings to sort the table!

Graphics Card	Overall Performance	4K Ultra	1440p Ultra	1080p Ultra	1080p Medium	Online Price (MSRP)	Specifications
GeForce RTX 4090	182.5	128.7	167.2	179.8	225.6	$2,177 ($1,599)	AD102, 16384 shaders, 2520 MHz, 82.6 TFLOPS, 24GB GDDR6X@21Gbps, 1008GB/s, 450W
GeForce RTX 4080	162.9	100.7	148.2	167.9	217.0	$1,249 ($1,199)	AD103, 9728 shaders, 2505 MHz, 48.7 TFLOPS, 16GB GDDR6X@22.4Gbps, 617GB/s, 320W
GeForce RTX 3090 Ti	144.7	83.0	128.2	153.5	209.7	$1,449 ($1,999)	GA102, 10752 shaders, 1860 MHz, 40.0 TFLOPS, 24GB GDDR6X@21Gbps, 1008GB/s, 450W
Radeon RX 6950 XT	137.7	74.6	124.4	151.3	210.9	$799 ($1,099)	Navi 21, 5120 shaders, 2310 MHz, 23.7 TFLOPS, 16GB GDDR6@18Gbps, 576GB/s, 335W
GeForce RTX 3090	136.0	75.1	119.2	145.5	205.2	$1,298 ($1,499)	GA102, 10496 shaders, 1695 MHz, 35.6 TFLOPS, 24GB GDDR6X@19.5Gbps, 936GB/s, 350W
GeForce RTX 3080 Ti	132.4	72.6	115.7	141.9	201.3	$1,149 ($1,199)	GA102, 10240 shaders, 1665 MHz, 34.1 TFLOPS, 12GB GDDR6X@19Gbps, 912GB/s, 350W
GeForce RTX 3080 12GB	129.0	70.1	112.5	138.4	198.3	$899 (N/A)	GA102, 8960 shaders, 1845 MHz, 33.1 TFLOPS, 12GB GDDR6X@19Gbps, 912GB/s, 350W
Radeon RX 6900 XT	127.6	66.9	113.4	142.2	203.2	$669 ($999)	Navi 21, 5120 shaders, 2250 MHz, 23.0 TFLOPS, 16GB GDDR6@16Gbps, 512GB/s, 300W
GeForce RTX 3080	123.9	65.9	106.5	132.8	197.5	$732 ($699)	GA102, 8704 shaders, 1710 MHz, 29.8 TFLOPS, 10GB GDDR6X@19Gbps, 760GB/s, 320W
Radeon RX 6800 XT	120.8	62.1	107.2	135.1	196.5	$597 ($649)	Navi 21, 4608 shaders, 2250 MHz, 20.7 TFLOPS, 16GB GDDR6@16Gbps, 512GB/s, 300W
Radeon RX 6800	108.0	53.5	93.8	120.6	187.6	$499 ($579)	Navi 21, 3840 shaders, 2105 MHz, 16.2 TFLOPS, 16GB GDDR6@16Gbps, 512GB/s, 250W
GeForce RTX 3070 Ti	106.0	50.6	91.0	117.4	182.0	$549 ($599)	GA104, 6144 shaders, 1770 MHz, 21.7 TFLOPS, 8GB GDDR6X@19Gbps, 608GB/s, 290W
GeForce RTX 3070	100.2	46.6	85.3	112.2	176.0	$529 ($499)	GA104, 5888 shaders, 1725 MHz, 20.3 TFLOPS, 8GB GDDR6@14Gbps, 448GB/s, 220W
GeForce RTX 2080 Ti	97.9	47.9	82.7	107.6	169.2	$997 ($999)	TU102, 4352 shaders, 1545 MHz, 13.4 TFLOPS, 11GB GDDR6@14Gbps, 616GB/s, 250W
Radeon RX 6750 XT	95.7	44.3	80.5	109.4	180.3	$420 ($549)	Navi 22, 2560 shaders, 2600 MHz, 13.3 TFLOPS, 12GB GDDR6@18Gbps, 432GB/s, 250W
GeForce RTX 3060 Ti	91.0	41.1	76.5	102.3	166.0	$389 ($399)	GA104, 4864 shaders, 1665 MHz, 16.2 TFLOPS, 8GB GDDR6@14Gbps, 448GB/s, 200W
Radeon RX 6700 XT	90.2	41.6	75.5	103.4	171.3	$379 ($489)	Navi 22, 2560 shaders, 2581 MHz, 13.2 TFLOPS, 12GB GDDR6@16Gbps, 384GB/s, 230W
GeForce RTX 2080 Super	83.4	37.2	70.9	94.4	152.5	$599 ($699)	TU104, 3072 shaders, 1815 MHz, 11.2 TFLOPS, 8GB GDDR6@15.5Gbps, 496GB/s, 250W
Radeon RX 6700 10GB	81.3	35.1	68.0	95.2	158.0	$329 (N/A)	Navi 22, 2304 shaders, 2450 MHz, 11.3 TFLOPS, 10GB GDDR6@16Gbps, 320GB/s, 175W
Intel Arc A770 16GB	80.9	39.4	68.4	89.6	139.0	$349 ($349)	ACM-G10, 4096 shaders, 2100 MHz, 17.2 TFLOPS, 16GB GDDR5@17.5Gbps, 560GB/s, 225W
GeForce RTX 2080	79.9	35.0	68.0	91.2	147.2	$699 ($699)	TU104, 2944 shaders, 1710 MHz, 10.1 TFLOPS, 8GB GDDR6@14Gbps, 448GB/s, 215W
Radeon RX 6650 XT	73.7	32.2	60.4	86.0	147.7	$299 ($399)	Navi 23, 2048 shaders, 2635 MHz, 10.8 TFLOPS, 8GB GDDR6@18Gbps, 280GB/s, 180W
GeForce RTX 2070 Super	73.7	31.8	62.5	84.7	137.0	$499 ($499)	TU104, 2560 shaders, 1770 MHz, 9.1 TFLOPS, 8GB GDDR6@14Gbps, 448GB/s, 215W
Radeon RX 6600 XT	72.0	31.4	58.4	84.0	145.6	$279 ($399)	Navi 23, 2048 shaders, 2589 MHz, 10.6 TFLOPS, 8GB GDDR6@16Gbps, 256GB/s, 160W
Intel Arc A750	71.4	33.8	59.9	79.0	128.6	$289 ($289)	ACM-G10, 3584 shaders, 2050 MHz, 14.7 TFLOPS, 8GB GDDR5@16Gbps, 512GB/s, 225W
GeForce RTX 3060	69.8	31.6	57.6	78.3	131.0	$299 ($329)	GA106, 3584 shaders, 1777 MHz, 12.7 TFLOPS, 12GB GDDR6@15Gbps, 360GB/s, 170W
GeForce RTX 2070	65.5	28.2	55.6	75.3	122.1	$499 ($399)	TU106, 2304 shaders, 1620 MHz, 7.5 TFLOPS, 8GB GDDR6@14Gbps, 448GB/s, 175W
GeForce RTX 2060 Super	62.2	26.4	52.5	72.0	116.8	$331 ($399)	TU106, 2176 shaders, 1650 MHz, 7.2 TFLOPS, 8GB GDDR6@14Gbps, 448GB/s, 175W
Radeon RX 5700 XT	61.7	29.3	53.3	73.7	125.8	$298 ($399)	Navi 10, 2560 shaders, 1905 MHz, 9.8 TFLOPS, 8GB GDDR6@14Gbps, 448GB/s, 225W
Radeon RX 6600	60.8	25.9	49.0	71.6	125.4	$209 ($329)	Navi 23, 1792 shaders, 2491 MHz, 8.9 TFLOPS, 8GB GDDR6@14Gbps, 224GB/s, 132W
Radeon RX 5700	54.4	25.7	47.2	64.8	111.3	$298 ($349)	Navi 10, 2304 shaders, 1725 MHz, 7.9 TFLOPS, 8GB GDDR6@14Gbps, 448GB/s, 180W
GeForce RTX 2060	51.9	20.6	42.1	60.8	106.2	$229 ($299)	TU106, 1920 shaders, 1680 MHz, 6.5 TFLOPS, 6GB GDDR6@14Gbps, 336GB/s, 160W
GeForce RTX 3050	50.6	22.3	41.1	57.1	98.1	$249 ($249)	GA106, 2560 shaders, 1777 MHz, 9.1 TFLOPS, 8GB GDDR6@14Gbps, 224GB/s, 130W
Radeon RX 5600 XT	48.5	22.5	42.0	58.1	100.6	$279 ($279)	Navi 10, 2304 shaders, 1750 MHz, 8.1 TFLOPS, 8GB GDDR6@14Gbps, 336GB/s, 160W
GeForce GTX 1660 Super	36.5	15.4	31.5	44.4	82.8	$195 ($229)	TU116, 1408 shaders, 1785 MHz, 5.0 TFLOPS, 6GB GDDR6@14Gbps, 336GB/s, 125W
Radeon RX 5500 XT 8GB	33.3	15.0	28.5	39.8	72.6	$229 ($199)	Navi 14, 1408 shaders, 1845 MHz, 5.2 TFLOPS, 8GB GDDR6@14Gbps, 224GB/s, 130W
GeForce GTX 1660	32.9	13.7	28.5	39.9	75.1	$199 ($219)	TU116, 1408 shaders, 1785 MHz, 5.0 TFLOPS, 6GB GDDR5@8Gbps, 192GB/s, 120W
Radeon RX 5500 XT 4GB	27.9	11.5	23.3	33.5	66.9	$149 ($179)	Navi 14, 1408 shaders, 1845 MHz, 5.2 TFLOPS, 4GB GDDR6@14Gbps, 224GB/s, 130W
GeForce GTX 1650 Super	26.5	10.5	21.0	33.2	67.9	$209 ($169)	TU116, 1280 shaders, 1725 MHz, 4.4 TFLOPS, 4GB GDDR6@12Gbps, 192GB/s, 100W
Radeon RX 6500 XT	25.9	9.3	18.8	31.9	68.5	$159 ($199)	Navi 24, 1024 shaders, 2815 MHz, 5.8 TFLOPS, 4GB GDDR6@18Gbps, 144GB/s, 107W
Intel Arc A380	25.8	9.4	21.2	31.0	58.7	$139 ($139)	ACM-G11, 1024 shaders, 2000 MHz, 4.1 TFLOPS, 6GB GDDR5@15.5Gbps, 186GB/s, 75W
GeForce GTX 1650	20.0	7.4	15.8	26.6	51.1	$156 ($159)	TU117, 896 shaders, 1665 MHz, 3.0 TFLOPS, 4GB GDDR5@8Gbps, 128GB/s, 75W
Radeon RX 6400	20.0	7.2	14.5	24.5	52.7	$139 ($159)	Navi 24, 768 shaders, 2321 MHz, 3.6 TFLOPS, 4GB GDDR6@16Gbps, 128GB/s, 53W

If you're looking at those numbers and wondering where they come from, they're a composite score generated by using the geometric mean of testing conducted at 1080p, 1440p, and 4K. Ray tracing hardware factors into the scores as well, which explains why Nvidia's RTX GPUs often outclass their AMD counterparts, and the same mostly goes for Intel Arc GPUs.

Prices are not a factor in the above rankings, but I've linked to Amazon in the first column — not all GPUs are available there, or in stock. Obviously, previous generation parts tend to be far less desirable, as prices have not dropped nearly enough relative to the performance on tap. Note that I do get Amazon affiliate compensation if you use my links.

With the overall performance rankings out of the way, I've also selected my top four picks for the enthusiast, high-end, mid-range, and budget markets. Here's where price becomes a much bigger consideration.

The Best Enthusiast Graphics Card

GeForce RTX 4090 24GB ($1,599)

Nvidia's GeForce RTX 4090 is the new ruler of the graphics card world, particularly for 4K gaming and especially if you want to play games with ray tracing effects. It's expensive and also sold out right now, but if you can nab one... damn, is it fast! You also get 24GB of memory, DLSS support — including the new DLSS 3 Frame Generation technology to further boost performance, especially in CPU limited scenarios — and improved efficiency. This is without question the top enthusiast graphics card right now.

The Best High-End Graphics Card

Radeon RX 6750 XT 12GB ($420)

Picking the best high-end GPU is a bit more difficult. The RTX 3080, 3070, and 3060 Ti are all worth a thought, but overall I give the edge to the Radeon RX 6750 XT. It has more memory, it performs nearly as well as the Nvidia cards — better if you're not worried about ray tracing effects — and it's reasonably priced starting at $449. You can also opt for the RX 6700 XT, which is only a few percent slower and costs about $50 less (for however long it remains in stock).

The Best Mainstream Graphics Card

Radeon RX 6600 8GB ($229)

There's little question about this one, as the RX 6600 delivers excellent performance and only costs $229 (sometimes less). Such prices were practically unheard of during the past two years for a GPU of this quality, but with the massive drop off in GPU mining profitability, there appears to be a growing surplus of AMD's Navi 23 parts. Our overall performance ranking puts it behind the RTX 3060, but that card costs about $100 more. Online prices are lower than even the RTX 3050, a card that's easily eclipsed by the RX 6600. This is arguably the best graphics card value right now. As a wild card alternative, you could also try to pick up an Intel Arc A750, which delivers superior performance and features... as long as the drivers don't bite you in the ass. Which they probably will.

The Best Budget Graphics Card

Radeon RX 6500 XT 4GB ($159)

This is a tough one, as the Radeon RX 6500 XT has a lot of limitations. It doesn't have much memory, it's limited to a 4-lane PCIe interface, it only has two display outputs, and it doesn't even have video encoding hardware. The problem is that the next step up basically puts us right back at the $250 midrange price point. Given you can find the RX 6500 XT for just $159, and it competes with the GTX 1650 Super that costs $60 more, there just aren't many other worthwhile options for under $200. Intel's Arc A380 would be an option, but it's continually sold out — which probably has more to do with a lack of supply than high demand.

Performance Charts

If you want a slightly different view of the performance rankings, here are the four charts — this time without trying to smush everything into a single composite score. So instead of looking at average and 1% lows "combined," you're getting the pure, unadulterated performance data.

Looking Forward to More Ada and RDNA 3

Nvidia's next generation Ada Lovelace architecture (RTX 40-series) and AMD's next generation RDNA 3 architecture (RX 7000-series) are slated to arrive before the end of the year, with the RTX 4080 and 4090 now launched and the AMD 7900 series coming December 13. While I could argue that you should just wait for the new parts to arrive, there's little reason to expect great prices — the RTX 4090 and now RTX 4080 sold out within hours at most and are now mostly sitting at scalper prices. Again. But I don't expect the high prices to remain the norm for very long, as there's no longer profitable GPU mining to support outrageous prices.

If you already have a decent graphics card, there's no rush to upgrade today. Also keep in mind that the wafer costs on the upcoming architectures are going to be significantly higher, and prices were paid in advance, many months ago — when GPU prices were still far higher. 4nm/5nm class GPUs will simply be inherently more expensive to make, and AMD and Nvidia will want to try and keep higher profit margins.

Beyond the rest of 2022, I do expect to see RTX 4070 Ti and lower tier Ada Lovelace GPUs start showing up in early 2023. AMD will also follow up the RX 7900 series with various other 7000-series parts, with probably at least two 7800-class cards in early 2023. The mainstream and budget cards won't come for a while, though.

Crossy Road Medieval Character Unlocks

I'm going to interrupt my normally scheduled GPU talk to go off on a tangent: Crossy Road. It's a stupid game. I've played it waaaaaay too much! Recently, the Medieval update came out with a new map, plus 14 new characters. I've already unlocked every other character, except for the Piggy Bank that requires you to pay $2.99 — I'm cheap, and also addicted. LOL

So, here are the 14 medieval characters, brief notes on each, and what you need to do to unlock them.

1. Gorgeous Prince: Prize Machine. Laughs knowingly and appears quite full of himself.
2. Princess: Prize Machine. Sometimes giggles, sometimes sniffs arrogantly, she's apparently the female counterpart to the Gorgeous Prince.
3. Blacksmith: Prize Machine. Anvils, anvils, everywhere! Lots of the regular objects on the medieval level are replaced with anvils, which the blacksmith can whack. It doesn't do anything other than make a sound.
4. Peasant: Prize Machine. Laughs like Beavis and Butthead, or some evildoer. If you stand still, he'll throw rotten fruit in whatever direction he's facing.
5. Falconer: Prize Machine. Not much to say, except the bird squawks and when you die it flies away.
6. Monk: Prize Machine. Carries a bible, which he lifts into the air when you stand still. He's the only character that can find Robin Hood.
7. Healer: Prize Machine. Carries around a book and a backpack, but otherwise nothing special that I could see.
8. Noblewoman: Prize Machine. Hums to herself and people bow to her when she passes.
9. Nobleman: Prize Machine. Mumbles importantly to himself and people bow to him when he passes.





10. The Wart: Prize Machine. A nod to Disney's A Sword in the Stone, The Wart is the only character that can draw Excalibur and unlock Merlin. Otherwise, he's pretty nondescript.
11. Black Knight: Requires 100 tokens, obtained by playing the Pecking Order daily challenge. He's an homage to Monty Python and the Quest for the Holy Grail. If you stand still for a bit, he loses his arms first and then his legs. Keep moving, or get to a log or a lily pad and he can rest.
12. Lady of the Lake: Jump in the sparkling river. There's an audio cue when you see the river that's different looking than the others. When playing as the Lady, all rivers are now sparkling. She carries a sword (presumably Excalibur), and will raise it in the air if you stand still for a bit.
13. Merlin: Play as The Wart and find an open field with the sword in the stone. Approach it and it shakes and rises up, and you unlock Merlin. Note that you can find the sword in the stone with other characters but the sword remains stuck. Merlin carries a wizard's staff, and if you stand still he'll raise it up and cause nearby objects to float.

14. Robin Hood: Play as the Monk (Friar Tuck?) and you can sometimes find him standing to the side. I found him near a river, but I don't know if that's always the case. Robin Hood will often encounter his Merry Men near jousting areas, and they will dance a jig with him.

Congratulations! Assume you have enough time and patience to spin the stupid Prize Machine a bunch of times (I was at 83,000 coins when the Medieval Update arrived), you can now unlock all of the new characters.

What's the Medieval level like? The trains are replaced by jousters, there are roads with carriages and people riding horses, and you get the usual rivers and such. All told, it's a fun little updates for people who, like me, spend too much time playing stupid games while running benchmarks.

AMD Radeon RX 6700 10GB (Non-XT) Review and Specifications, Featuring Sapphire

We've got something new for today: My first graphics card review for the site. It's mostly because this is a card that hasn't been reviewed much, the AMD Radeon RX 6700 non-XT 10GB model. This GPU came later in the life cycle, and it sits roughly between the RX 6650 XT and RX 6700 XT — in price as well as performance. It's sort of a last gasp for RDNA 2, right as RDNA 3 cards are about to launch, though the RX 7900 XTX and RX 7900 XT will be priced in an entirely different category. You can already find the FPS results in our GPU performance hierarchy, but let's dig into things a bit deeper for this review.

AMD Radeon RX 6700 10GB Specifications
Architecture	Navi 22
Process Technology	TSMC N7
Transistors (Billion)	17.2
Die size (mm^2)	336
Compute Units	36
GPU Cores (Shaders)	2304
Ray Accelerators	36
Boost Clock (MHz)	2450
VRAM Speed (Gbps)	16
VRAM (GB)	10
VRAM Bus Width	160
Infinity Cache	80
Render Outputs	64
Texture Mapping Units	144
FP32 TFLOPS (Single-Precision)	11.3
FP16 TFLOPS (Half-Precision)	22.6
Bandwidth (GB/s)	320
Total Board Power (Watts)	175
Launch Date	March 18, 2021
Launch Price	$369 (unofficial)

For quite some time, the Radeon RX 6700 XT was the only desktop card to use AMD's Navi 22 GPU. Later, the RX 6750 XT joined the fun, and both are fully enabled variants of the chip. Given there are always defective die from a wafer, and most of those can be harvested by disabling the affected functional units (CUs, SMs, memory controllers, etc.), we long expected there would be an RX 6700 non-XT card. But it didn't come, and it didn't come, and then it suddenly stealth-launched from a couple of AMD's AIC (add-in card) partners: Sapphire and XFX. Maybe there are others making the 6700 10GB, but in the US at least these are the only ones available.

There's no official pricing on the RX 6700 10GB, sort of like there's no official price for the RTX 3080 12GB or some of the other custom GPUs. It's been a few months since they first appeared, and our recollection is they cost over $400 at launch, but pricing has dropped in the ensuing months. Now, the Sapphire Pulse card we're reviewing goes for $349 while the XFX Speedster model costs $329.

That's probably fair, as Sapphire tends to build a better product — quieter fans, same or better performance. Still, the Pulse is a baseline model with no frills like RGB lighting to get in the way. Also note that it has just a single 8-pin power connector, a tried and true option that shouldn't melt and will provide more than enough electricity for the 175W part. The card has three DisplayPort 1.4a outputs and a single HDMI 2.1 port, typical of most modern graphics cards.

Compared to the RX 6700 XT, you get four fewer CUs (Compute Units) and a 160-bit memory interface. Along with one of the memory channels, you also lose 16MB of Infinity Cache. Memory bandwidth is still reasonable at 320 GB/s, and the extra 2GB might help performance in some games. Speaking of which...

There's not much point in spending a lot of time analyzing the performance. As expected, it lands about midway between the RX 6650 XT and the RX 6700 XT. That makes sense considering it has 11–25  percent more memory bandwidth than the RX 6600/6650 XT and about 5–7 percent more compute available. It also has more Infinity Cache, which means it might make better use of the compute, and the real-world clocks tend to be higher than the official specs.

It's interesting to see how this latecomer fares against other similarly priced graphics cards as well. Overall, it trades blows with the Intel Arc A770 16GB, winning at 1080p medium, tying at 1080p ultra, and then falling behind at 1440p and 4K. It also wins against the RTX 3060, sometimes by a lot, and at times it looks more like an RTX 3060 Ti competitor.

But raytracing proves to be the usual Achilles' Heel for AMD's RDNA 2 architecture, and in our more demanding DirectX Raytracing tests the RX 6700 10GB often falls behind the 3060, and in Minecraft it even trails the lackluster RTX 3050. Buy it for the rasterization performance, in other words, and not for its raytracing prowess.

Also, given the $20 savings, if you're looking at the RX 6700 10GB you might want to opt for the XFX variant. Sapphire's card is fine, but every penny counts in these trying times. If you can afford the extra $20, you could probably also afford another $50 and start getting into cards that are overall more capable, like the RX 6750 XT.

Keep an eye out for sales this holiday season, as we're expecting a lot of GPUs to get cleared out to make way for the incoming RX 7000-series and RTX 40-series graphics cards. Those will start at the extreme and high-end pricing segments, but come spring 2023, we'll probably see something closer to the mainstream market — RX 7700 maybe, as a trimmed down take on the RX 7700 XT? Make it so, AMD, make it so, and this time skip the one year delay between those card launches.

Holy Nerdvana Verdict: 81/100

A capable GPU that would have been a lot more attractive had it arrived shortly after the RX 6700 XT. Pricing and performance are still decent, raytracing proves a weak spot for AMD as usual, but we can't help but look longingly toward the near future when we'll see next-generation mainstream graphics cards deliver significantly more performance for a hopefully similar price.

AMD Radeon RX 7900 XT Specifications

The AMD Radeon RX 7900 XT is the slightly smaller sibling of the just-announced Radeon RX 7900 XTX. Using the same RDNA 3 architecture and GPU chiplets, but with one fewer MCD (Memory Cache Die), it will presumably land just a bit lower down on our GPU performance hierarchy when it arrives on December 13.

We have to admit that the naming certainly represents a bit of an annoyance. The only difference between the RX 7900 XTX and the 7900 XT is one letter, and yet the latter will definitely have less performance. For all the people that were up in arms over the RTX 4080 16GB and RTX 4080 12GB shenanigans, this doesn't really feel any better. Regardless, these are the names AMD has chosen and we'll have to see how they match up. Here's the rundown of the RX 7900 XT specs.

AMD Radeon RX 7900 XT Specifications
Architecture	Navi 31
Process Technology	TSMC N5 + N6
Transistors (Billion)	58
Die size (mm^2)	300 + 185
Compute Units	84
GPU Cores (Shaders)	10752
Ray Accelerators	84
Boost Clock (MHz)	2400
VRAM Speed (Gbps)	20
VRAM (GB)	20
VRAM Bus Width	320
Infinity Cache	80
Render Outputs	192
Texture Mapping Units	336
FP32 TFLOPS (Single-Precision)	51.6
FP16 TFLOPS (Half-Precision)	103.2
Bandwidth (GB/s)	800
Total Board Power (Watts)	300
Launch Date	December 13, 2022
Launch Price	$899

The core design looks to be the same, but there's one fewer active MCD and thus only 20GB of GDDR6 with a 320-bit memory interface, cutting bandwidth by 17%. The GCD is also likely a harvested die with only 84 of the potential 96 Compute Units (CUs) active, a 12.5% reduction. AMD gave a Game Clock figure of just 2000 MHz (compared to 2300 MHz on the 7900 XTX), but we're going to assume the actual boost clocks are much closer and put in 2400 MHz for now.

To be 100% clear, AMD hasn't given an exact number on the teraflops for the 7900 XT, but we're basing this on past experience with the RX 6800 and RX 6800 XT. Those cards had Game Clocks of 1815 MHz and 2015 MHz, with official boost clocks of 2105 MHz and 2250 MHz. In our actual testing, however, the reference RX 6800 averages around 2250 MHz while gaming compared to the RX 6800 XT at 2340 MHz. Time will tell precisely where the 7900 XT and XTX land, but we think actual clocks will be far closer than the revealed Game Clocks suggest.

Which brings us to the competition. The RX 7900 XT will start at $899, just $100 less than the RX 7900 XTX. That seems... far too close. On paper, the XTX has 19% more theoretical compute (using our numbers) and 20% more bandwidth, yet it only costs 11% more. If you're willing to spend $900 on the 7900 XT, we can't see any reason why you'd want to "save" $100 and not get the faster XTX.

Elsewhere, the 7900 XT has the same base price as the "unlaunched" RTX 4080 12GB. We wouldn't be surprised if Nvidia got word of the AMD card and realized just how badly the 4080 12GB would look in comparison and so it decided to go back to the drawing board. There's a good chance the RX 7900 XT will still be reasonably competitive with the upcoming RTX 4080, which based on Nvidia's own testing was about 25–30 percent faster than the 4080 12GB that's no longer a thing. (News flash if you didn't catch the sarcasm: The AD104 12GB GPU is absolutely a thing still, but it will probably end up as a slightly tweaked RTX 4070 card or maybe even an RTX 4070 Ti.)

Right now, the RX 7900 XT offers more VRAM and more theoretical compute performance than the RTX 4080, but of course it doesn't have Tensor cores and DLSS support, and it probably still has slower DXR performance, despite AMD's 50% improvement in ray tracing performance with RDNA 3. How will it stack up? We'll find out next month.

AMD Radeon RX 7900 XTX Specifications

The AMD Radeon RX 7900 XTX has officially been announced, the first in a salvo of RDNA 3 architecture GPUs that will compete with the best graphics cards. It's a radical new approach to GPU designs, using chiplets — in a similar fashion to how AMD uses chiplets on its Zen 3 and Zen 4 CPUs. Except here the use of chiplets has been tuned and tweaked to work best for graphics rather than CPUs. 

The specifications for the Radeon RX 7900 XTX are mostly known at this point, though there are a few missing pieces. For example, AMD provided details on its "Game Clock" but not on the Boost Clock, and it appears to have rounded things off to the nearest 100 MHz. Maybe. We do have a compute teraflops figure that seems to be based off the boost clock, however, which means we mostly have what we need.

AMD Radeon RX 7900 XTX Specifications
Architecture	Navi 31
Process Technology	TSMC N5 + N6
Transistors (Billion)	58
Die size (mm^2)	300 + 222
Compute Units	96
GPU Cores (Shaders)	12288
Ray Accelerators	96
Boost Clock (MHz)	2500
VRAM Speed (Gbps)	20
VRAM (GB)	24
VRAM Bus Width	384
Infinity Cache	96
Render Outputs	192
Texture Mapping Units	384
FP32 TFLOPS (Single-Precision)	61.4
FP16 TFLOPS (Half-Precision)	122.8
Bandwidth (GB/s)	960
Total Board Power (Watts)	355
Launch Date	December 13, 2022
Launch Price	$999

AMD says the RX 7900 XTX won't be competing with the RTX 4090 — that Nvidia's top Ada Lovelace part is effectively in a league of its own. Instead, it will be targeting the next step down, the upcoming RTX 4080, and it certainly looks like AMD will give Nvidia some much-needed competition. Total theoretical compute on the 7900 XTX is 61.4 teraflops, based on a tentative boost clock of 2500 MHz — the actual boost clock will probably be ±10 MHz, given past experience.

Where things get interesting is in some of the architectural changes. AMD now has double the FP32 performance per Compute Unit (CU), except it only has the same INT32 rate per CU. Put another way, AMD has followed Nvidia's example and now it has two blocks of 64 dual-issue Stream processors that are FP32 SIMD registers, with half as many INT32 SIMD registers. What does that mean for performance? Probably AMD's real-world performance per theoretical teraflops will now track closer to Nvidia's real-world performance per theoretical teraflops. We don't know that for sure yet, as we haven't been able to test the cards, but the architectural changes do make that a distinct possibility.

AMD's GPU chiplets are super interesting and should help reduce the overall price of the graphics cards quite a bit. The Graphics Compute Die (GCD) uses TSMC's "5nm" N5 process node, which should be relatively similar to the 4N node Nvidia uses for Ada Lovelace in terms of transistor density and power characteristics — both are effectively 5nm-class, at least. The Memory Cache Die (MCD) meanwhile uses TSMC's slightly older N6 (7nm-class) process node, which probably costs less than half as much per wafer.

The result of the chiplet approach is that the big GCD is less than half the size of Nvidia's AD102 chip, and then the cache and memory interface logic — which doesn't scale well to smaller process nodes — gets put on a bunch of tiny MCDs. Furthermore, AMD can put up to six (functional) MCDs around the GCD, but as we'll see with the RX 7900 XT, it can also do fewer MCDs with a partially disabled GCD to improve overall yields and reduce costs.

Let's quickly cover the die sizes and chips per wafer estimates as well. The GCD measures 300mm^2, with dimensions of approximately 24.6x 12.2mm. Based on that size, AMD can get around 174 chips per N5 wafer — nearly double the number of AD102 chips Nvidia can get, or effectively half the cost. The MCDs meanwhile are 37mm^2, or approximately 7.5x4.9mm, which means AMD gets around 1,650 dies per wafer. There's no need to even worry about harvesting partial die to improve yields with a chip that small; it either works properly or it gets discarded (or turned into a "dummy die" for something like the 7900 XT).

Total combined die size for the 7900 XTX is 522mm^2, so still smaller than Nvidia's biggest Ada chip, with a cost that's roughly equivalent of making a monolithic ~370mm^2 N5 chip. But if AMD had gone the monolithic route, Navi 31 would still be in the ~500mm^2 range, so the net cost savings for AMD looks to be around 25–30 percent.

To quickly cover the speeds and feeds, AMD says it improved overall utilization of the shader resources by increasing the various cache and buffer sizes. It's not clear exactly how much faster the new RDNA 3 CU is versus the RDNA 2 CU, but there's a lot more FP32 available — 160% more than the RX 6950 XT. The Ray Accelerators are also 50% faster, which should help to narrow the performance gap in ray tracing games. Clock speeds aren't too much higher than on RDNA 2, but power requirements haven't really changed much. Where the RX 6950 XT was a 335W TBP, RX 7900 XTX is only 355W — easily handled with slightly tweaked cooling designs.

AMD didn't go the 16-pin 12VHPWR route either, which means we can use the tried and true dual 8-pin power connectors. Partner cards will almost certainly have triple 8-pin connectors on the higher end models, and TBP will probably go up into the 400~450W range on factory overclocked cards, but we'll have to wait and see what that does for actual performance. We suspect AMD targeted optimal efficiency for its own cards, rather than following in the footsteps of Nvidia's power-hungry RTX 4090.

Nvidia GeForce RTX 4090 Specifications

The Nvidia GeForce RTX 4090 currently reigns as the king of graphics cards — to see how performance stacks up, check our extensive list of GPU benchmarks and the best graphics cards. It uses Nvidia's latest Ada Lovelace graphics architecture, with the largest of the Ada chips: AD102. The combination of lots of cores and high clock speeds results in incredible levels of performance — the reference Founders Edition has a 2,520 MHz boost clock, and often runs at over 2.7 GHz in testing.

Note that the RTX 4090 doesn't use a fully enabled AD102 chip, as it has 128 of the potential 144 SM (Streaming Multiprocessor) blocks enabled, along with 72MB of the potential 96MB of L2 cache. We'll probably see a GeForce RTX 4090 Ti in the future, as there's certainly room for it.

Nvidia GeForce RTX 4090 Specifications
Architecture	AD102
Process Technology	TSMC 4N
Transistors (Billion)	76.3
Die size (mm^2)	608.4
Streaming Multiprocessors	128
GPU Cores (Shaders)	16384
Tensor Cores	512
RT Cores	128
Boost Clock (MHz)	2520
VRAM Speed (Gbps)	21
VRAM (GB)	24
VRAM Bus Width	384
L2 Cache	72
Render Outputs	176
Texture Mapping Units	512
FP32 TFLOPS (Single-Precision)	82.6
FP16 TFLOPS with Sparsity (FP8)	661 (1321)
Bandwidth (GB/s)	1008
Total Board Power (Watts)	450
Launch Date	October 12, 2022
Launch Price	$1,599

Nvidia uses a custom TSMC 4N (4nm Nvidia) process node for the RTX 4090, with a die size of 608.4mm2 — a bit smaller than the previous generation GA102 used in the RTX 3090. The change in process allowed Nvidia to roughly triple the number of transistors, which also helps with the performance. Total raw computational power measures 82.6 teraflops for 32-bit floating-point, over 661 teraflops for FP16, and 1.3 petaflops for AI workloads that can leverage the FP8 support on the Tensor cores.

Again, we strongly suspect Nvidia isn't putting all of its cards on the table just yet, if you'll pardon the pun. There will almost certainly be a higher tier 40-series at some point, though given the brouhaha of melting 16-pin 12VHPWR power connectors and adapters, perhaps Nvidia will take a bit more time to work out the kinks.

If you're not quite willing to go whole hog and splurge on the RTX 4090, the RTX 4080 should arrive on November 16, 2022. That's a big drop in specs and potential performance, possibly the largest gap we've seen from Nvidia in recent years. On paper, the 4090 has 69% (!) more compute performance than the 4090, 41% more memory bandwidth, and 50% more memory capacity. Considering it only (/sarcasm) costs 33% more than the 4080, this is arguably the better value.

Those waiting for a 40-series card that won't absolutely break the bank will need to hold out hope for the RTX 4070. We expect those will arrive in January most likely. Considering the canceled RTX 4080 12GB, the 4070 will likely slot into a similar space, but hopefully it won't have the same $899 price point as the unlaunched card.

Nvidia GeForce RTX 4080 Specifications

Let's take a moment this Thanksgiving season to express our gratitude over the fact that Nvidia caved into pressure from the PC enthusiast community and canceled the RTX 4080 12GB, which means the RTX 4080 16GB is now the only RTX 4080-class GPU coming down the pipeline — we'll add it to our list of GPU benchmarks and the best graphics cards as soon as it's available.

There will of course be future RTX 4080 Ti / Super / Whatever cards at some point — probably next year — but those will presumably be clearly labeled as such. Really, it's a case of people being pissed about Nvidia trying to charge $899 for a card with only 12GB VRAM on a 192-bit memory interface.

While the RTX 4080 takes the reigns from the previous generation RTX 3080, it's also worth pointing out that it's inheriting the pricing of the RTX 3080 Ti rather than the far more attractively priced RTX 3080 (10GB). The RTX 4080 (formerly 16GB) meanwhile still has the following specs.

Nvidia GeForce RTX 4080 Specifications
Architecture	AD103
Process Technology	TSMC 4N
Transistors (Billion)	45.9
Die size (mm^2)	378.6
Streaming Multiprocessors	76
GPU Cores (Shaders)	9728
Tensor Cores	304
RT Cores	76
Boost Clock (MHz)	2505
VRAM Speed (Gbps)	22.4
VRAM (GB)	16
VRAM Bus Width	256
L2 Cache	64
Render Outputs	112
Texture Mapping Units	304
FP32 TFLOPS (Single-Precision)	48.7
FP16 TFLOPS with Sparsity (FP8)	390 (780)
Bandwidth (GB/s)	717
Total Board Power (Watts)	320
Launch Date	November 16, 2022?
Launch Price	$1,199

As far as we know, the RTX 4080 has the same specification as the formerly announced 16GB card. Nvidia hasn't updated the product page to say otherwise, so unless that happens the RTX 4080 will be as above. Unlike on most previous architectures, the top card at launch and the penultimate card from Nvidia are using two different GPU dies, AD102 for the RTX 4090 and AD103 for the RTX 4080. Neither GPU is fully enabled, though the RTX 4080 comes pretty close.

With RTX 4080, you get 76 SMs out of a potential 80 SMs for the full AD103. We could see a fully enabled GPU in the future, but more likely is that it will not be a desktop GeForce card — it could be a mobile chip, or for the professional Nvidia GPU offerings (probably upgraded to 32GB of memory). That means 9,728 CUDA cores and 304 Tensor cores, alongside 76 RT cores.

Theoretical performance tips the scales at 48.7 teraflops FP32, 390 teraflops of FP16, and 780 teraflops of FP8 for AI workloads (with sparsity enabled). That's about 20% faster than the RTX 3090 Ti, though you're getting less memory and memory bandwidth. The 64MB of L2 cache should make up for the drop in memory bandwidth in most workloads, particularly gaming.

Speaking of which, this will be the first and so far only graphics card to officially use GDDR6X memory clocked higher than 21 Gbps. Micron also makes a 24 Gbps chip, but Nvidia apparently decided to step back to 22.4 Gbps for the RTX 4080. There are already RTX 4090 cards that apparently have 24 Gbps chips (marked as 21 Gbps, but overclocking proves they're not), and based on how those cards perform, the 24 Gbps chips run cooler and use less power, which is how Nvidia manages to keep the power down to a relatively tame 320W.

The launch date is rumored to be November 16, just a few weeks away and about a month after the RTX 4090 launch. Nvidia likely hopes to beat AMD's RDNA 3 to market, though we'll know more about the AMD offerings shortly (on November 3).

Nvidia GeForce RTX 4070 Specifications

The Nvidia GeForce RTX 4070 launched in Aprile 2023, with reduced specs compared to the higher tier models. Nvidia increased the generational pricing by $100 this round, and performance basically matches the previous generation RTX 3080 10GB card.

Nvidia GeForce RTX 4070 Specifications
Architecture	AD104
Process Technology	TSMC 4N
Transistors (Billion)	32.0
Die size (mm^2)	294.5
Streaming Multiprocessors	46
GPU Cores (Shaders)	5888
Tensor Cores	184
RT Cores	46
Boost Clock (MHz)	2475
VRAM Speed (Gbps)	21
VRAM (GB)	12
VRAM Bus Width	192
L2 Cache	36
Render Outputs	64
Texture Mapping Units	184
FP32 TFLOPS (Single-Precision)	29.1
FP16 TFLOPS with Sparsity (FP8)	233 (466)
Bandwidth (GB/s)	504
Total Board Power (Watts)	200
Launch Date	April 2023
Launch Price	$599 ($549)

The simple solution is to not launch a $600 RTX 4070 until most of the RTX 3080/3090 inventory has been sold. It's the holiday shopping spree season now, and obviously Nvidia thinks it can get away with this tactic. Delaying the RTX 4070 a couple of months — and building up a larger inventory of such GPUs in the meanwhile — won't really hurt, especially if Nvidia can successfully clear out all of those Ampere GPUs.

And it probably can! Technologically savvy people who read this blog might not be duped into buying an overpriced RTX 30-series card at this point in time, but there are tons of less knowledgeable gamers that just want a new high-end PC or graphics card for Christmas, Hanukah, or whatever. And if Nvidia and its partners can't sell all of those Ampere GPUs directly to gamers, rest assured there are large OEMs like Dell (Alienware), HP (Omen), Lenovo (Legion), etc. who will buy loads of cards at a discount and foist them off on the type of people who can't be bothered to build their own PCs.

Looking at the potential RTX 4070 specifications, it could very well end up with the same amount of VRAM and theoretical compute performance as the previous generation RTX 3080 Ti. Of course it has just a bit more than half the memory bandwidth, but the 48MB of L2 cache should make up for that. You also get all the Ada Lovelace architectural upgrade, like support for DLSS3.

While we wait the arrival of the RTX 4070, anyone who wants more performance and are willing to pay for it can step up to the RTX 4080 or RTX 4090, Just the way Lord Jensen would like it.

Nvidia GeForce RTX 3090 Ti Specifications

Last generation's ultimate GPU, the GeForce RTX 3090 Ti only launched in late March of 2022 — to see how performance stacks up, check our extensive list of GPU benchmarks and the best graphics cards. It was also a product of the cryptocurrency era, which meant GPU supply still hadn't caught up with demand and Nvidia felt it could get away with an exorbitant launch price of $1,999.

A few months later, demand plummeted and Nvidia dropped the 3090 Ti Founders Edition price to $1,099 at Best Buy, which in turn pissed off some of its add-in card (AIC) partners like EVGA, who ultimately decided to stop making graphics cards altogether — and who probably has a shload of RTX 30-series parts still in stock that it's trying to offload as quickly as possible. But don't feel bad for EVGA or the other AIC partners, because they made out like bandits throughout all of 2020 and 2021!

As far as the RTX 3090 Ti specifications, this is the fully enabled GA102 chip, the top of the Ampere architecture. It's interesting to compare this chip with the RTX 4090 specifications, to see how much things have changed. Note that the die size on GA102 is slightly larger than the new AD102, and yet AD102 packs nearly three times as many transistors. That's all thanks to the move from Samsung 8N ("8nm Nvidia") that was really just a tweaked version of Samsung's 10nm process, over to TSMC 4N ("4nm Nvidia") that's a tweaked variant of TSMC's 5nm N5 process. Obviously, the latter offers significantly higher transistor density.

Nvidia GeForce RTX 3090 Ti Specifications
Architecture	GA102
Process Technology	Samsung 8N
Transistors (Billion)	28.3
Die size (mm^2)	628.4
Streaming Multiprocessors	84
GPU Cores (Shaders)	10752
Tensor Cores	336
RT Cores	84
Boost Clock (MHz)	1860
VRAM Speed (Gbps)	21
VRAM (GB)	24
VRAM Bus Width	384
L2 Cache	6
Render Outputs	112
Texture Mapping Units	336
FP32 TFLOPS (Single-Precision)	40.0
FP16 TFLOPS (Sparsity)	160 (320)
Bandwidth (GB/s)	1008
Total Board Power (Watts)	450
Launch Date	March 29, 2022
Launch Price	$1,999

The maxed out Ampere chip in the RTX 3090 Ti has 84 SMs and 10,752 CUDA cores, along with 336 Tensor cores. Combined with clocks, the 3090 Ti offers up to 40 teraflops of FP32, or 320 teraflops of FP16. Unlike the new Ada Lovelace Tensor cores, there's no FP8 support, which means theoretical compute for AI workloads is about one-fourth of what the RTX 4090 can offer — not accounting for other architectural differences.

In order to try and differentiate the RTX 3090 Ti from the existing RTX 3090, Nvidia also bumped up the TBP (Total Board Power) by 100W, which allows the GPU to sustain substantially higher clocks compared to the 3090. It also allowed Nvidia and its partners a chance to test the waters for a 450W graphics card and pave the way for the 4090. Ironically, where the 12-pin and 16-pin adapters that came with RTX 3090 Ti cards all seemed to work fine, the new 4090 16-pin adapters appear to have used poor soldering that's prone to cracking and can cause a fire hazard. Oops.

The RTX 3090 Ti also uses new 2GB (16Gb) 21 Gbps GDDR6X memory, where the original RTX 3090 used 1GB (8Gb) 21 Gbps chips — this is the same 21 Gbps GDDR6X memory as the new RTX 4090, which means they have the same memory bandwidth. However, Ada cards have significantly larger L2 caches — 12X larger for the 4090 versus 3090 Ti — and that in turn means effective memory bandwidth is substantially improved. If the L2 cache hit rate on a 4090 for example is 60% at 4K, that means only 40% of memory accesses actually hit the GDDR6X memory, which means 2.5X more effective bandwidth. Nvidia has not detailed the actual L2 hit rates, which will vary by workload, but 50–70 percent seems likely.

At present, RTX 3090 Ti cards can be found starting at around $1,290 on Amazon. Considering the pending arrival of the GeForce RTX 4080, I can't help but think the RTX 3090 Ti represents a terrible value at this time — or actually any time since launch. At first it was far too expensive, but by the time Nvidia slashed prices and forced the AIC partners to do the same, Ada Lovelace was right around the corner and there was no reason to even consider buying a 3090 Ti. It's generally the fastest graphics card from the Nvidia Ampere and AMD RDNA 2 era — at least at 4K and in professional workloads — but it arrived too late, provided too little of a performance uplift (only about 10% faster than the 3090), and also required substantially more power.

If you're even thinking about buying an RTX 3090 Ti card, used or otherwise, you should instead look at the RTX 4090. Well, unless you can pick one up for under $700. And don't forget that RTX 4070 is also on the way.

Nvidia GeForce RTX 3090 Specifications

It's hard to believe that Nvidia's GeForce RTX 3090 has been around for over two years now, though it has now been displaced by the new RTX 4090. You can see where the RTX 3090 ranks among other cards in our GPU performance hierarchy.

Originally launched in September 2020 alongside the RTX 3080, people complained about the price — how it was a Titan-class price without the Titan-class features like improved professional application drivers and support. And then the cryptocurrency mining boom of 2020 through 2022 hit, and suddenly $1,500 for an RTX 3090 that could potentially earn (at its highest point) over $30 per day seemed like a steal.

Naturally, prices shot up to compensate, scalpers got super involved with graphics cards, and the rest is sort of history. Painful history, at least for gaming enthusiasts — miners loved the RTX 3090. This was the fastest consumer graphics card throughout 2021, packed with features to make 4K gaming truly viable. And yet most of these cards probably spent all of 2021 chipping away in the Ethereum mines. RIP, Ethereum mining! But let's look at the specifications.

Nvidia GeForce RTX 3090 Specifications
Architecture	GA102
Process Technology	Samsung 8N
Transistors (Billion)	28.3
Die size (mm^2)	628.4
Streaming Multiprocessors	82
GPU Cores (Shaders)	10496
Tensor Cores	328
RT Cores	82
Boost Clock (MHz)	1695
VRAM Speed (Gbps)	19.5
VRAM (GB)	24
VRAM Bus Width	384
L2 Cache	6
Render Outputs	112
Texture Mapping Units	328
FP32 TFLOPS (Single-Precision)	35.6
FP16 TFLOPS (Sparsity)	142 (285)
Bandwidth (GB/s)	936
Total Board Power (Watts)	350
Launch Date	September 24, 2020
Launch Price	$1,499

RTX 3090 uses the same GA102 chip as several other Nvidia GPUs — RTX 3080, RTX 3080 12GB, RTX 3080 Ti, and RTX 3090 Ti are all based off GA102. Back in 2020, there probably weren't that many nearly fully functional GA102 chips available, so supplies of the 3090 were quite limited and Nvidia pushed the narrative that these were "professional cards, not just for gamers." That's despite having GeForce branding right in the card name, so you can guess how well received such claims were.

With 82 of the available 84 SMs enabled, and the full 384-bit memory interface matched up with 24 1GB GDDR6X chips, this was a beast of a card. Power consumption on the reference model was rated at 350W, but third-party AIC vendors pushed things well into the 400W and higher range on factory overclocked cards.

One of the big issues with the RTX 3090, particularly for cryptocurrency miners, is that the GDDR6X chips ran hot. Fire up Ethereum mining and many 3090 cards, including the Founders Edition, would quickly reach 110C on the memory before starting to throttle. Dismantling the cards to replace the original thermal pads with better variants was common, and even in gaming workloads you could easily hit more than 100C on the memory.

The root of the problem was that, because Micron was only making 1GB (8Gb) capacity GDDR6X chips at the time, the RTX 3090 required half of the 24 chips to reside on the back of the PCB, while the other half were on the same side as the GPU. Take a look at that chunky Founders Edition up top: It has a chunky a triple-slot cooler to help keep all the chips cool... except that's only for the "front" of the PCB; the back of the PCB just has a metal cover with thermal pads and no active cooling. I can state from experience that the back of the card — with the RTX 3090 logo seen above, which usually faces upwards in most computer cases —can get extremely toasty under load!

Theoretical compute performance for the RTX 3090 tips the scales at 35.6 teraflops, and the Tensor cores can do up to 285 teraflops of FP16 with sparsity enabled. All of that number crunching prowess is put to good use in ray tracing games, and Nvidia's DLSS feature can leverage the Tensor cores to help with image upscaling that looks almost as good as native, at least in Quality mode. Almost all of the most demanding ray tracing games support DLSS, though, which actually brings 4K gaming into reach.