NVIDIA's RTX 4090 was a massively powerful GPU with massive power consumption to match, something that's been making headlines since before launch with or without combusting adapter cables. Now the RTX 4080 is here and NVIDIA claims that it's going to double the performance of the previous generation which would put it dangerously close to the RTX 4090's performance level at what they say will be a much lower power draw. And yet NVIDIA's claims were far higher than the power draw we actually see which is the kind of lie that I don't mind hearing. The question now is can this thing live up to those performance claims? Or at $1,200 will the 40 series' inflated price tags make it a "skip it" generation like the RTX 20 series cards?
There's only one RTX 4080 and here it is
The RTX 4080 16 gigger, I guess I don't really need to call it that anymore, the 4080 has roughly 60% of the CUDA cores compared to the bigger 4090, with slightly lower clocks to go with them. 
This is coupled with a 256-bit instead of a 384-bit memory bus which means effective bandwidth is cut to just two thirds. Given that the starting price is roughly two thirds of the 4090, the math is mathing. Although the rated total graphics power is a bit higher than you'd expect given the spec cuts, still it's the same 320 watt TGP as the outgoing RTX 3080.
So if we're gaining performance like NVIDIA says we are, we're not paying for it there, at least in theory. Let's test it in practice.
We grab the RTX 4080 along with its predecessor and what we consider to be the closest competing GPUs in terms of price and performance and slap them into our 13th gen core bench to give them the best chance to stretch their legs.
4K Gaming Benchmarks
Starting with non-Ray traced 4K, the RTX 4080 pulls a respectable 25% lead over the RTX 3090 Ti in "Forza Horizon 5" and nearly reaches 80% of the RTX 4090's performance which puts it on solid footing out of the gate.
"F1 22" sees the 6950 XT pull in close thanks to the monstrous traditional rendering performance that card has, and the 3090 Ti closes the gap substantially thanks to its greater memory bandwidth. 
The 3090 Ti's performance carries over to "Cyberpunk" somewhat but the other cards fall behind NVIDIA's 40 series cards. 
"Hitman 3" sees the RTX 4090 pull significantly ahead of the 4080, again, likely thanks to its memory bandwidth as the 3090 Ti and 6950 XT both trade blows in this title. 
"Modern Warfare 2" pushes the 4080 back again compared to the 4090, but it's still commanding that roughly 50% lead over its predecessor and sits comfortably ahead of the 3090 Ti and 6950 XT. 
Older titles like "Red Dead Redemption 2" and "Shadow of the Tomb Raider" both display similar results across the board with the RTX 4080 clearly topping the last generation's best cards but trailing the 4090 by about 30%.
1440p Gaming Benchmarks
1440p results are unsurprisingly much tighter thanks to the lighter load at places on the GPU although not in the way you might expect. While the 4080 closes in on the 4090 the other GPU's remain where they were relative to the 4K results, suggesting that it's only the RTX 4090 that's CPU-bound in these tests. 
There are cases like "Cyberpunk" and "Modern Warfare 2" where the extra horsepower still makes a big difference, but for users of 1440p displays and especially in older titles like "Red Dead Redemption 2" and "Shadow of the Tomb Raider", the RTX 4080 may be more than enough to achieve close to 4090 level performance while costing substantially less money and drawing substantially less power which does also cost money. We'll get to the power draw, but first, Ray Tracing.
4K Ray Tracing Benchmarks
This is AMD's Achilles Heel and it shows even in "Shadow of the Tomb Raider" where the 6950 XT is only roughly equal to the lower tier RTX 3080. Even in this older title, NVIDIA's new RT cores appear to perform slightly faster than the 3090 Ti's giving the 4080 a broader lead than it had in traditional rendering. 
"F1 22", by comparison, had the 4080 lose relative performance against the 4090 versus when Ray Tracing was turned off. But, it pulls ahead substantially from the RTX 3080 making for frame rates that are playable at 4K rather than cinematic. 
"Cyberpunk" unfortunately remains cinematic on all of our cards at 4K with Ray Tracing enabled though it's here that we really do see the RTX 4080 pull over double the performance of its predecessor and AMD's 6950 XT. 
And while DLSS 3.0 is still early days it appears that the two cards handle its frame generation roughly the same way, which is unsurprising but we had to test it to be sure. 
For the same reasons as 1440p rendering performance, the RTX 4080 seems to benefit more from DLSS overall than the 4090 does, thanks to the latter becoming more CPU bound. And this is something we should see more of as the rest of the lineup continues to launch.
Productivity Benchmarks
Moving on to productivity, the RTX 4080 is nearly double the performance of the 3080 in blender rendering thanks both to the additional and upgraded RT cores, and it seems to come within three quarters of the speed of the 4090.
"DaVinci Resolve" favors the 4090 by about a minute with the 4080 coming in just a little slower than the 3090 Ti. However, it's addition of AV1 encoding means that any 40 series GPU I mean any of them, is going to make your PC substantially faster at this kind of rendering compared to any of the other GPU's we've tested here. 
MATLAB GPUBench is NVIDIA only but we can see that depending on the algorithm the RTX 4080 can come within spitting distance of the 4090 in single precision or be a good 40% behind in double precision. 
Either way, it's consistently well ahead of the RTX 3090 Ti and 3080, making it an excellent choice for GPU compute. 
It's less clear cut for Topaz AI, where our test suite finished sooner on the 4080 than the 3090 Ti and within 10 minutes of the 4090. But AMD actually pulled ahead here suggesting that Topaz is geared more towards raw GPU compute than AMD's AI accelerating tensor cores, or at least the AI model we used was.
Finally SPECviewperf puts the RTX 4080 in a good overall position, heavily outperforming both the RTX 3080 and 3090 Ti, although AMD pulls off some wins in CATIA, Creo, Energy, Medical, and especially Siemens NX, though the latter is likely because that version that's used in SPECviewperf doesn't support acceleration on NVIDIA yet.
Power Consumption and what IS TGP anyway?
Power consumption has been a hot topic since the earliest rumors of the RTX 40 series launch and the RTX 4090 proved our concerns were valid. 
However, the RTX 4080 is surprisingly tame. Yeah, during our testing it almost never reached its 320 watt rate of total graphics power, instead hovering around the 300 watt mark or usually lower.
We were so confused by this that we asked NVIDIA what was up and they responded by saying, "Oh, yeah actually we're changing what TGP means." Sure. Okay. So what did they say? 
Basically, TGP has popularly been thought of as a power target until recently. It's the power level that the GPU tried to adhere to when under load. Opportunistic boost meant that GPUs would often spike higher for short periods, but usually settled in around their TGP rating.
Now, TGP is being looked at more like a power limit, the maximum the card should hit under normal conditions, and NVIDIA's built themselves a buffer. Igor's lab recently published an article showing this in greater depth with the RTX 4090.
NVIDIA themselves claimed that the RTX 4080 draws roughly 251 watts on average while gaming, though it's closer to 300 watts at 4K depending on the game. And yeah, even when running a power virus we couldn't get the 4080 to go much higher than the rated 320 watts despite the GPU load remaining high and the core clocks remaining above two gigahertz.
Thermals and cooling
This leads to some interesting questions because each and every RTX 4080 we've tested has had a massively overbuilt cooler that resulted in the card barely ever getting warm. 
Like seriously, under full synthetic load We're looking at around the low 60 degrees at worst and in games it sits around in the 50s.
The fans never even began to approach 50%, let alone higher. 
Even the internal case temperature was the lowest of any GPU in our test suite with a delta T over ambient of between just 10 and 11 degrees.
Unlike the RTX 4090, this GPU would be well suited to a small form factor chassis. I'm wondering if NVIDIA made this adjustment to TGP late in development and nobody got the memo in time. The question now is why? Though NVIDIA didn't volunteer that information when we asked, when we overclocked the cards we noted that there wasn't a lot of headroom for the GPU core clocks.
Overclocking
Our cards could only push between 110 and 150 megahertz over stock, and while this is obviously going to vary from card to card, that's only about 1% which is much lower than the roughly 15% you might find back on the 10 series cards. This might be one reason why NVIDIA dialed back the GPU power a bit. In fact, even when we pushed the memory and the core clock as high as they would go we still didn't get substantially higher frame rates.
However, we did substantially increase power consumption lending further credence to the idea that NVIDIA saw no reason to push the GPU to its rated 320 watts under normal operation. What we've observed is its sweet spot.
Should you buy a 4080, or an older GPU?
With that in mind, what are you getting for $1,200 US? You're getting a GPU that's much faster than anything the previous generation can muster with dual NVENC encoders that support AV1, a thermally very tame design with boatloads of thermal headroom that could potentially be cut down to fit lower profile chassis, and official support for deep learning frame generation to boost perceived frame rates where it needs to.
However, unlike the RTX 4090, which is a Halo product with no price pair currently, the RTX 4080 has to compete with those last gen cards. And while the RTX 3090 Ti is close in its current retail price, it's not as close in performance or power consumption. You're paying an extra hundred dollars for the 4080 and getting, in my opinion, more than a hundred dollars in value.
The RX 6950 XT and RTX 3080, however? Those do represent a threat in terms of value if the extra features simply don't matter to you. Some of these can be purchased for as little as $800 brand new with used cards routinely going for less sometimes even $600. At half the price they definitely do not represent just half the value based on our testing. Bear in mind, the AMD card draws substantially more power.
Conclusion
So if that's a concern, go with the 3080 for now. We'll have all of those linked below. None of which is to say that you shouldn't buy the RTX 4080. It's a solid card and it's only just getting started. Just maybe wait for reviews of AMD's RDNA 3 GPUs to drop before pulling the trigger. Who knows? It'll probably even go "on sale" if AMD does a good enough job. Competition is beautiful.
Comments (1)