NVIDIA RTX 5090 Geekbench Leak: OpenCL and Vulkan Tests Reveal True Performance Uplifts

[GGforever]

The RTX 50-series fever continues to rage on, with independent reviews for the RTX 5080 and RTX 5090 dropping towards the end of this month. That does not stop benchmarks from leaking out, unsurprisingly, and a recent lineup of Geekbench listings have revealed the raw performance uplifts that can be expected from NVIDIA's next generation GeForce flagship. A sizeable chunk of the tech community was certainly rather disappointed with NVIDIA's reliance on AI-powered frame generation for much of the claimed improvements in gaming. Now, it appears we can finally figure out how much raw improvement NVIDIA was able to squeeze out with consumer Blackwell, and the numbers, for the most part, appear decent enough.

Starting off with the OpenCL tests, the highest score that we have seen so far from the RTX 5090 puts it around 367,000 points, which marks an acceptable jump from the RTX 4090, which manages around 317,000 points according to Geekbench's official average data. Of course, there are a plethora of cards that may easily exceed the average scores, which must be kept in mind. That said, we are not aware of the details of the RTX 5090 that was tested, so pitting it against average scores does seem fair. Moving to Vulkan, the performance uplift is much more satisfying, with the RTX 5090 managing a minimum of 331,000 points and a maximum of around 360,000 points, compared to the RTX 4090's 262,000 - a sizeable 37% improvement at the highest end. Once again, we are comparing the best results posted so far against last year's averages, so expect slightly more modest gains in the real world. Once more reviews start appearing after the embargo lifts, the improvement figures should become much more reliable.



View at TechPowerUp Main Site | Source

 

[Daven]

I'm still guessing that the 5090 will be 20-30% faster in pure rasterization benchmarks over the 4090 (these leaked benchmarks seem to back that up somewhat). That's reasonable for a 25% increase in price and a 30% increase in power on the same process node. Of course, it's way out of my budget range but I'm sure some wealthy game enthusiasts will enjoy the best possible performance money can buy.

 

[windwhirl]

a sizeable 37% improvement at the highest end.

Not sizable in terms of actual IPC improvement if the 5090 comes with 21760 CUDA cores. The 4090 had 16384, so the 5090 would have roughly 33% more CUDA cores, so that'd be like 5% faster IPC, roughly (would have to account for clock differences and such so that's a rough number). That, and the higher TDP, 575W vs 450W, roughly 28% higher.

So, with regards to Vulkan, the card is a bit more efficient, has a bit more IPC, but the overwhelming majority of the improvement comes from increased CUDA core counts and power consumption.

 

[docnorth]

It has ~20% more transistors, so 20% performance increase where the faster VRAM doesn’t matter means zero improvement. With the faster RAM coming into play, a 25-30% ( and sometimes higher) uplift is expected.

 

[AnotherReader]

Not sizable in terms of actual IPC improvement if the 5090 comes with 21760 CUDA cores. The 4090 had 16384, so the 5090 would have roughly 33% more CUDA cores, so that'd be like 5% faster IPC, roughly (would have to account for clock differences and such so that's a rough number). That, and the higher TDP, 575W vs 450W, roughly 28% higher.

So, the card is a bit more efficient, has a bit more IPC, but the overwhelming majority of the improvement comes from increased CUDA core counts and power consumption.

I'm not sure about higher IPC; even the best OpenCL score is barely 16% more than the 4090 despite the 5090 having 33% more SMXs. In any case, the IPC comparison will have to wait for the 5070 which has almost the same number of SMXs as the 4070.

 

[AnarchoPrimitiv]

I'm still guessing that the 5090 will be 20-30% faster in pure rasterization benchmarks over the 4090 (these leaked benchmarks seem to back that up somewhat). That's reasonable for a 25% increase in price and a 30% increase in power on the same process node. Of course, it's way out of my budget range but I'm sure some wealthy game enthusiasts will enjoy the best possible performance money can buy.

Is it "objectively" reasonable? Or subjectively reasonable now that we've been conditioned to expect so much less with each release?

 

[windwhirl]

I'm not sure about higher IPC;

Made a small edit in my post to clarify it's about Vulkan. But yeah, OpenCL is actually worse. Not sure if it's just driver deficiencies or just Nvidia didn't care about OpenCL. I understand CUDA itself is far more popular? So maybe from Nvidia's POV OpenCL isn't super relevant so they just don't prioritize optimizing for it?

 

[theouto]

So my theory of the higher performance coming in because of the bigger chip was spot on, and thus the price on the 90 class card is higher because the chip is that much bigger, and there are no real improvements.
It's a throwaway generation really, we really are nearing stagnation.

 

[NoneRain]

I'm still guessing that the 5090 will be 20-30% faster in pure rasterization benchmarks over the 4090 (these leaked benchmarks seem to back that up somewhat). That's reasonable for a 25% increase in price and a 30% increase in power on the same process node. Of course, it's way out of my budget range but I'm sure some wealthy game enthusiasts will enjoy the best possible performance money can buy.

Not in my book.
The price will keep increasing while there are people paying. 1:1 perf% x price% may be the norm, but just until people stop swallowing ( ͡° ͜ʖ ͡°)

 

[AnotherReader]

Made a small edit in my post to clarify it's about Vulkan. But yeah, OpenCL is actually worse. Not sure if it's just driver deficiencies or just Nvidia didn't care about OpenCL. I understand CUDA itself is far more popular? So maybe from Nvidia's POV OpenCL isn't super relevant so they just don't prioritize optimizing for it?

CUDA is more popular by far, but Nvidia has invested in OpenCL support as well. Vulkan might be a better point of comparison, but even there, gains range from 26% to 37%. The latter figure is probably an overclocked SKU so I suspect it's closer to 26% which is less than the 33% increase in SMX count.

 

[Baccala]

I never gave it a second thought until now but does anybody know if geekbench 'points' are actually tied to a more official unit of measure?

 

[windwhirl]

I never gave it a second thought until now but does anybody know if geekbench 'points' are actually tied to a more official unit of measure?

No. Each benchmark is only comparable to the same exact benchmark, and preferably trying to minimize setup discrepancies. You can't grab Geekbench and say something like "ten points in Geekbench is the same as 20 points in 3Dmark" or whatever.

 

[evernessince]

Not sizable in terms of actual IPC improvement if the 5090 comes with 21760 CUDA cores. The 4090 had 16384, so the 5090 would have roughly 33% more CUDA cores, so that'd be like 5% faster IPC, roughly (would have to account for clock differences and such so that's a rough number). That, and the higher TDP, 575W vs 450W, roughly 28% higher.

So, with regards to Vulkan, the card is a bit more efficient, has a bit more IPC, but the overwhelming majority of the improvement comes from increased CUDA core counts and power consumption.

I'm pretty sure it's a 0% IPC improvement. There are 32.8% more cores on the 5090 than the 4090 and the boost clock of the 5090 is 7.69% higher.

Combine the core count and frequency increases and you get a number higher than the actual performance increase.

This is definitely a tock generation and one of the most lackluster one's at that. It's Nvidia's equivalent to the R9 300 series. No IPC gains, no efficiency gains, no new marquee features (only updates to existing ones).

 

[Francoporto]

2 years and a half to get only +30% perf, but with consumption increased and price increased....

 

[tpuuser256]

I'm still guessing that the 5090 will be 20-30% faster in pure rasterization benchmarks over the 4090 (these leaked benchmarks seem to back that up somewhat). That's reasonable for a 25% increase in price and a 30% increase in power on the same process node. Of course, it's way out of my budget range but I'm sure some wealthy game enthusiasts will enjoy the best possible performance money can buy.

with 33% more cores and 75% memory bw, there is not way the improvement is less than 30%

 

[AnotherReader]

with 33% more cores and 75% memory bw, there is not way the improvement is less than 30%

Despite having 68% more SMXs and 41% more memory bandwidth than the 4080, the 4090 is only 28% faster than the 4080 at 4K. The best case is ray tracing in Cyberpunk 2077 where the increase in performance is 36% for hybrid and 41% for path tracing at 4K.

 

[OSdevr]

CUDA is more popular by far, but Nvidia has invested in OpenCL support as well. Vulkan might be a better point of comparison, but even there, gains range from 26% to 37%. The latter figure is probably an overclocked SKU so I suspect it's closer to 26% which is less than the 33% increase in SMX count.

Last I checked Nvidia's investment in OpenCL is the absolute bare minimum. You know how DirectX 12 has "feature levels?" Well OpenCL is worse. On paper Nvidia's 4000 series supports OpenCL 3.0 which came out in 2020. However OpenCL 3.0 only requires the complete OpenCL 1.2 functionality which came out in 2011! Everything more recent is optional, and AFAIK Nvidia's latest only implement the 1.2 feature set.

 

[mama]

No surprise. I expect reviews when they emerge will essentially say the same thing they said about the 4090: Stupid price to performance (don't buy it) but it is the strongest available card.

 

[JustBenching]

Isn't the 5090 a prime example of why we need AI / DLSS / MFG etc.? The 5090 has a lot more of everything (power, cores, bandwidth, vram) and yet the gains are average to bad. Realistically a hypothetical 5090 that is twice the size of the 4090 - at similar power - would be like what, 50% faster? Makes no sense for nvidia to pursue that, it's completely unsustainable

 

[Chrispy_]

The 5090 has 33% more resources than the 4090, so a ~37% increase in performance is yet more proof that nearly all of the improvements for the 50-series are software gimmicks to improve the end result, rather than more actual raw performance.

Isn't the 5090 a prime example of why we need AI / DLSS / MFG etc.? The 5090 has a lot more of everything (power, cores, bandwidth, vram) and yet the gains are average to bad. Realistically a hypothetical 5090 that is twice the size of the 4090 - at similar power - would be like what, 50% faster? Makes no sense for nvidia to pursue that, it's completely unsustainable

I interpret it differently.
The 5090 is an example of 33% more physical hardware generating a 37% more performance, so your hypothetical double 4090 would be more than double as fast.

I'm not quite sure how you interpret a 33% more hardware for 37% faster as "average to bad".

 

[JustBenching]

The 5090 has 33% more resources than the 4090, so a ~37% increase in performance is yet more proof that nearly all of the improvements for the 50-series are software gimmicks to improve the end result, rather than more actual raw performance.


I interpret it differently.
The 5090 is an example of 33% more physical hardware generating a 37% more performance, so your hypothetical double 4090 would be more than double as fast.

I'm not quite sure how you interpret a 33% more hardware for 37% faster as "average to bad".

More hardware, more power, more bandwidth and more vram.

 

[Chrispy_]

It has ~20% more transistors, so 20% performance increase where the faster VRAM doesn’t matter means zero improvement. With the faster RAM coming into play, a 25-30% ( and sometimes higher) uplift is expected.

Don't use transistor count, it incluedes tens of billions of transistors that have nothing to do with compute performance - things like the video engine, fixed-function hardware for features, display output, PCIe connectivity, communications controllers etc. These do not scale linearly with the power of the card. The 4060 has 19% the core count of a 4090 but 26% of the transistor budget, and that's despite cutbacks to fixed function hardware like the PCIe interface in both generation and lane count, reductions to the number of nvenc encoders, and the removal of mGPU support logic altogether.

More hardware, more power, more bandwidth and more vram.

more hardware scales linearly - more pipelines/cores means more operations per clock.

power doesn't, it's just a side effect of more hardware drawing current at once
bandwidth doesn't, it's just a side effect of needing to feed more hardware without starving it
vram doesn't, it adds zero performance and is simply required to hold the data. Not having enough means you simply cannot run those settings or that dataset for simulation/LLM/compute.

 

[chrcoluk]

So my theory of the higher performance coming in because of the bigger chip was spot on, and thus the price on the 90 class card is higher because the chip is that much bigger, and there are no real improvements.
It's a throwaway generation really, we really are nearing stagnation.

Their focus is on the AI stuff, which can see has been boosted on all the cards. For consumer side to give something sellable we of course get given the this circus multi frame gen.

I have always had the opinion, if there isnt anything meaningful to release, then dont release it. Its so annoying we have instead this "release to a schedule, so there is always something new out there".

The 5000 series by far the best thing about it is the slimmer coolers and angled connectors. Does that warrant a new generation by itself? Probably not, instead release a rev 2 model with these shroud improvements, but I expect they wanted to use the new chips for the AI market, which has really led to the change on the consumer side as well, as well as of course marketing reasons.

If software hadnt become so inefficient, we could be all using 1080ti's getting 300fps every game. Or rather playing at 60fps consuming 50w.

 

[Prima.Vera]

I'm still guessing that the 5090 will be 20-30% faster in pure rasterization benchmarks over the 4090 (these leaked benchmarks seem to back that up somewhat). That's reasonable for a 25% increase in price and a 30% increase in power on the same process node. Of course, it's way out of my budget range but I'm sure some wealthy game enthusiasts will enjoy the best possible performance money can buy.

There is absolutely NOTHING reasonable about this card, from all points of view, including price, performance via previous gen, or power consumption. Stop being so gullible and easily brainwashed by nVidia marketing and payed reviewers/influencers.

 

[Chrispy_]

There is absolutely NOTHING reasonable about this card, from all points of view, including price, performance via previous gen, or power consumption. Stop being so gullible and easily brainwashed by nVidia marketing and payed reviewers/influencers.

On top of that, distributors and AIB insiders are claiming that stock availability is so low for the 5090 that it might as well just be vaporware for gamers.

According to MLID, an Nvidia employee said that even Nvidia employees have limited stock in their own employee store.

I don't know if that's poor yields on such a gargantuan chip, or a result of the 5090 being too valuable to sell to gamers at "$2000" when the same silicon can be sold for $10000+ as an RTX 6000B workstation or server card instead. Remember, every 4090 that hit the market was because Nvidia had temporarily satisfied all the RTX 6000A customers willing to pay $8000+ for the exact same silicon.