The RTX PRO 6000 is one of the best advertising for AMD’s Ryzen 9800X3D (learn our evaluate) that we use in our GPU check benches, principally as a result of we’re seeing 5% to 14% scaling in gaming benchmarks versus the RTX 5090 (learn our evaluate). We additionally tore it down and located essentially the most insanely dense PCB we’ve ever seen with a reversion away from liquid steel.
However this doesn’t have 96GB of VRAM simply because that’s the place all of it went after they misplaced it on the RTX 5080 (learn our evaluate), 5070 Ti (learn our evaluate), 5070 (learn our evaluate), 5060 Ti 8GB (learn our evaluate), and 5060 (learn our evaluate)– it’s additionally there as a result of it’s higher for… AI.
Immediately, we’re benchmarking a few of these in a spherical of experimental benchmarks. We haven’t accomplished machine studying and LLM checks earlier than and we don’t have a full methodology outlined but, however we’ve begun experimenting.
NVIDIA RTX 6000 Overview
This card will not be for gaming. That a lot is clear. We’re nonetheless testing it in video games, and that’s principally for our personal functions: We wish to understand how a lot room the 9800X3D has in it to scale for GPUs, and luckily, it appears to be like like there’s nonetheless a lot for an additional era of opinions. The 9800X3D retains up with the RTX PRO 6000 even when at 1080p in some conditions.
The largest distinction between the PRO 6000 and the 5090 after the 96GB vs. 32GB reminiscence capability is the GPU itself: The PRO 6000 has 24,064 CUDA cores to the 21,760 of the 5090, almost an 11% improve. That’s a giant distinction.
There are a number of variations of the RTX 6000. The one we’re particularly known as the Blackwell Workstation Version, however NVIDIA additionally has so-called “server editions” of the playing cards, and people are meant to enter servers and racks the place they’re getting drive fed air from the entrance of the chassis slightly than utilizing an FE type cooler, which is the one we’re at present. So, the one we’re is extra just like what we might see in form of the patron to prosumer workstation class.
NVIDIA RTX 6000 Specs, Structure Fundamentals, and Worth
Moving into the specs:
The RTX PRO 6000 Blackwell GPU is constructed on the GB202 Blackwell die. The total GB202 die has help for as much as 24,576 CUDA cores, 768 TMUs, 192 ROPs which may even be current, and can also be used for the RTX 5090.
The PRO 6000 variation on this GPU has 24,064 CUDA cores, so it’s down by 512 and isn’t an ideal die, with 752 TMUs (down from 768) and nonetheless 192 ROPs, they usually’d higher all be there for the value. The cardboard can also be operating 188 RT cores and helps PCIe 5.0 x16. Marketed clocks are 1590 MHz base and 2617 MHz enhance. The actual purpose individuals purchase this card is its reminiscence, although, at 96GB of GDDR7.
The closest shopper class card to match in opposition to is the RTX 5090. It’s constructed on the identical die, however with 176 ROPs, 680 TMUs, 170 RT cores, 21,760 CUDA cores, 32GB of GDDR7 reminiscence, a 512-bit bus width, and a far cheaper price at $2,000 to, realistically, $3,000 slightly than $8,000 to $11,000.
For some notable variations: The 5090 solely has a 3rd of the reminiscence capability of the PRO card, and the RTX PRO 6000’s die is nearly 98% current whereas the RTX 5090’s die is nearer to 88.5% current, with the rest both being fused off or faulty fallout and fused off.
You possibly can additionally technically purchase 4 5090s for the value of 1 RTX PRO 6000, so the use case actually does come all the way down to having as a lot reminiscence as potential hooked up to a single GPU and PCIe slot.
NVIDIA RTX 6000 Tear-Down
The RTX Professional 6000 we examined is a 2-slot card with 2 giant followers on the entrance. The PCB is within the middle of the cardboard and you’ll see the place the fins begin to get taller. This means the place the PCB plate ends.
The cardboard’s PCIe slot is linked through a custom-made pinout adapter.
If you happen to take a look at the bottom of the cardboard the place the followers are, you’ll discover a slight melancholy down in the direction of the center of the fins. That is supposed to assist with stress drop. The design is full flow-through, which pushes air straight by way of the cardboard slightly than right into a PCB.
The opposite aspect of the cardboard appears to be like similar to a 5090.
The highest of the RTX Professional 6000 retains the angle on the cardboard’s vented slats. This tasks the air out at an angle that’s up and away from the cardboard.
Just like the 5090, the cardboard makes use of a 12VHPWR cable.
Beginning disassembly, we eliminated the cardboard’s again plate. Then we eliminated a single screw, which allowed us to drag off the cardboard’s exterior body.
From right here, we eliminated some screws across the middle of the cardboard to take away the middle covers. Doing this reminded us of how masterful the mechanics of the FE design is. It’s assembled rather well, it’s simple to take aside, and there’s no mechanisms to try to stop disassembly.
Subsequent, we unscrewed the clamp that holds the PCIe slot to the cardboard.
Right here’s a detailed take a look at the eliminated PCIe adapter.
From right here, we pried a plate off of the cardboard, exposing the PCB.
This gave us our first take a look at the place the reminiscence is linked. We counted 16 reminiscence modules on the again with the clay sort of thermal pads.
The PCB is loopy dense, which is an engineering problem.
We then proceeded to take away ribbon cables and screws to the leaf spring to entry the GPU.
Pulling off the PCB, we had been anticipating to see liquid steel, however we noticed thermal paste as a substitute. We actually weren’t anticipating that. We additionally noticed a large GPU substrate.
Upon eradicating the PCB, we observed that 1 pad went over the sting a bit.
Analyzing the paste software sample, it appeared just like the heaviest imprint was proper within the middle with lighter stress on the outer edges, which is ok.
Wanting on the thermal pads, they had been making clear contact, although we did see some pads that had been smashed over the perimeters a bit. That is fantastic, nevertheless. We counted 32 complete reminiscence places, which signifies that they’re 3GB reminiscence modules as that totals 96GB, the capability of the cardboard.
Taking a more in-depth look, we observed some areas of poor contact alongside one fringe of the inductors, the place we observed that half of the inductors are literally coated by a thermal pad. One among them solely had a few third of the inductor coated. That’s not good as they really want to get the pad placement appropriate on GPUs that value 8-11 thousand {dollars}.
The board additionally has a number of little tiny parts, resulting in a really dense PCB.
Lastly, we cleaned off the thermal paste off of the GPU, exposing the GB 202-870-A1 SKU. This can be a 5090 die, only a fuller model of it.
NVIDIA RTX 6000 Thermal Benchmarks
Thermal testing is up now. The RTX 5090 makes use of liquid steel with a 2-slot FE cooler that, genuinely, is likely one of the most spectacular GPU coolers we’ve ever labored on. The prototype 4090 cooler we disassembled was extra spectacular in mechanical and cooling capabilities, however it was big and impractical to fabricate. The two-slot FE has accomplished nicely typically, with the exception being reminiscence cooling. We discovered that reminiscence ran scorching on our 5090 FE. That’s extra of a priority with the PRO 6000 as nicely because it has a lot reminiscence.
Right here’s a thermal check throughout a set render workload. The 5090 ran at round 72 levels Celsius for the GPU core temperature on this check, which is general glorious contemplating its 2-slot design.
The VRAM for the 5090 ran at round 90 levels Celsius, which was hotter than we’re comfy with. As soon as accounting for set up right into a case and potential long-term implications of pad dryout and dirt, that’s excessive. The PRO 6000 ran its GPU core at a considerably hotter 82 levels Celsius on this check, so about 10 levels hotter than the 5090. The GPU reminiscence was about the identical, measuring 88 levels through software program.
Fan Response
None of which means something with out fan speeds.
Utilizing solely the auto settings, so following no matter VBIOS has programmed for the goal GPU temperature, the RTX 5090 ran at 1550-1600 RPM for the common fan velocity. The PRO 6000 ran at 1700. That’s slightly quicker with a a lot hotter core, however related reminiscence thermals.
Acoustics
Acoustics are subsequent. For this testing, we’re utilizing our hemi-anechoic chamber that we closely invested in for bettering our check high quality. Our subsequent main enchancment will probably be upgrading the microphone, which is presently our bottleneck. We may hit a 6-8 dBA noise ground with a greater mic, however we’re presently at 13.6 to fifteen.0 dBA for the ground. That’s nonetheless nice although.
The RTX 5090 ran at about 32.5 dBA, with the PRO 6000 on the similar noise degree; nevertheless, the noise ground was barely totally different between these two, so adjusting for that, the PRO 6000 can be about 1 dBA louder underneath the identical situations. That is hardly noticeable as a distinction, if noticeable in any respect for most individuals.
General, the cooler design stays good when taking its measurement into consideration. The 5090 had increased noise ranges within the increased frequency vary, significantly 8000 to 10000 Hz. It additionally had a peak round 175 Hz, with the PRO peaking at 194 Hz and following the same sample, simply adjusted proper.
Usually talking, this principally follows the identical development. Our frequency cutoff is 150 Hz, so the upper outcome on the left may be ignored.
NVIDIA RTX 6000 Gaming Benchmarks
Gaming benchmarks are subsequent.
Dragon’s Dogma 2 – 4K
Dragon’s Dogma 2 is first.
Examined at 4K, the RTX PRO 6000 landed at 140 FPS AVG, which has it forward of the RTX 5090 by 5.8% for common framerate. Lows will not be notably totally different. In different phrases, that’s about an additional $1,000 per 1% enchancment, or about $780 per 1 FPS improve over the 5090 when calculating by the distinction in value for these two.
The lead within the RTX PRO 6000 is about 42% over the 4090 (watch our evaluate), with the 5090 already forward of the 4090 by about 34% right here.
Dragon’s Dogma 2 – 1440p
At 1440p, the lead over the 5090 is expanded to six.3%, making certain we get our full gaming worth out of the cardboard. Low efficiency will not be totally different outdoors of the same old scaling together with the advance in common frametimes.
The RTX PRO 6000 hits 201 FPS AVG right here, main the 189 FPS of the 5090 and the 155 of the 4090. NVIDIA presently holds the complete high cluster of this chart, with AMD principally specializing in the fashionable mid-range and Intel targeted on the fashionable low-end.
Dragon’s Dogma 2 – 1080p
We subsequent examined 1080p, which is what everyone knows consumers of this card actually wish to use it for. So-called AI use circumstances are clearly secondary in comparison with enjoying video video games at 1080p.
The RTX PRO 6000 leads this extraordinarily essential and pivotal chart with a 226 FPS AVG, dropping to only a 5.3% lead. Actually, there’s something invaluable that comes out of this chart: It’s advertising for AMD’s Ryzen 7 9800X3D, which is someway maintaining with these GPUs sufficient that it might nonetheless produce distinguished outcomes between the $2,000-$3,000 5090 and the $8,000 to $11,000 PRO 6000. That’s spectacular. That is extra amusing to us because it proves the longevity of our bench {hardware} greater than anything.
FFXIV 4K
Remaining Fantasy 14 is up now, examined first at 4K.
The RTX Professional 6000 ran at 208 FPS AVG, main the 5090’s 184 FPS AVG outcome by about 13%. That’s a bigger acquire than we’ve seen in a few of the different video games, although clearly that is nonetheless a card meant for VRAM-intensive use circumstances and never gaming. That there’s nonetheless some energy left past the 5090 is what’s extra fascinating, alongside the truth that the 9800X3D is so succesful. Lows scale with the common and will not be meaningfully higher with the PRO 6000.
FFXIV 1440p
Examined at 1440p, we see a ten% enchancment in framerate in opposition to the 5090, up at 348 FPS AVG for the 6000. Once more, we’re principally excited to see the 9800X3D persevering with to scale and illustrating that there’s room left for an additional era of testing on these benches. Lows stay in-step with the common.
FFXIV 1080p
At 1080p, the PRO 6000 once more outperforms the 5090. This time, it’s lowered to about 4.3%. Let’s transfer on.
Starfield – 4K
Up subsequent, we’re testing a card nobody buys for gaming with a sport nobody performs. In Starfield at 4K, the PRO 6000 ran at 115 FPS AVG, main the 5090 by about 7%. Lows are principally inside error. The worth right here is someplace within the vary of $700 to $1,000 per 1% enchancment, relying on the value of the 5090.
Starfield – 1440p
At 1440p, the RTX PRO 6000 ran at 154 FPS AVG, main the 147-148 FPS results of the 5090 marginally. The advance over the 4090 is 17%, with the 5090 having beforehand led the 4090 by 12% on this check, as we’re turning into sure elsewhere.
The 1% lows for the 6000 are noteworthy right here: At 89 FPS for the averaged 1% low, the PRO 6000 outperforms the common framerate of the 5070 and almost matches the 4070 Ti with simply the PRO’s 1% numbers alone.
Resident Evil 4 – 4K
Resident Evil at 4K is up subsequent. On this one, the RTX PRO 6000 ran at 219 FPS AVG, main the 207 FPS AVG results of the 5090 by 6%. That’s in keeping with a lot of the different checks to this point. The lows, once more, will not be meaningfully totally different. The 4090’s 151 FPS AVG permits the 6000 a lead of 45%, or the 5090 a lead of 36.7%.
Resident Evil 4 – 1440p
At 1440p, the Blackwell workstation card ran at 371 FPS AVG, pushing the framerate measurably increased than the roughly 350 FPS AVG of the 5090 beforehand. We’re a few 6.5% enchancment, aligning with prior outcomes once more, and but once more displaying simply how good the 9800X3D is.
Black Fantasy: Wukong – 4K
Black Fantasy: Wukong is up now, examined first at 4K and representing one in all our heaviest non-RT workloads. The RTX PRO 6000 didn’t break the 100 FPS barrier, however it was the primary card to exceed 90 — not that this significantly issues to anybody when the 5090 was already within the mid 80s. The 6000 outperforms our 5090 FE outcome by 7.4%, above what we’ve seen on common to date by about 1 share level. 92 FPS AVG is an efficient displaying on this benchmark when at 4K. The subsequent card within the stack is the 4090 at 67 FPS AVG, then the 5080 at 58 FPS AVG.
Black Fantasy: Wukong – 1440p
At 1440p, the PRO 6000 ran at 138 FPS AVG and led the 5090 marginally, however it was technically nonetheless a measurable enchancment. Whole uplift is 6.1%, aligning with Resident Evil 4 beforehand. As soon as once more, the subsequent card is the 4090. After that is the 5080, which isn’t that totally different from the 4080 Tremendous (learn our evaluate), which isn’t that totally different from the 4080 (watch our evaluate), which isn’t that totally different from the 5070 Ti.
Black Fantasy: Wukong – 1080p
Black Fantasy: Wukong at 1080p is subsequent. Selfishly, 1080p is fascinating as soon as once more for probing at if there’s room within the CPU for extra GPU. Seems, there’s: With the 9800X3D, the RTX PRO 6000 ran at 168 FPS AVG, bettering on the 5090 by about 5%. That’s lower than we noticed at 1440p, which was lower than we noticed at 4K, so the cardboard is gaining floor as decision will increase.
Dying Gentle 2 – 4K
Dying Gentle 2 is up subsequent. At 4K, the PRO 6000 ran at 144 FPS AVG, which is a noteworthy enchancment on the 126 FPS AVG of the 5090 FE. That’s a shocking 13.8%, which to this point matches just one different sport we’ve examined — and that was Remaining Fantasy. Many of the different video games are nearer to 5-7%.
The 4090 is subsequent at 91 FPS AVG, that means the PRO 6000 is 59% forward of the 4090.
Dying Gentle 2 – 1440p
At 1440p, there’s nonetheless scaling: The PRO 6000 is now at 246 FPS AVG, main the 5090 by virtually 13%. That’s a drop from what we noticed at 4K, however nonetheless greater than most different video games.
Cyberpunk 2077: Phantom Liberty – 4K
Cyberpunk 2077: Phantom Liberty is one in all our subsequent most intensive video games after Black Fantasy: Wukong. At 4K, the RTX PRO 6000 breaks by way of 100 FPS and hits 108, main the 5090 by 13.5%. That is really a reasonably large lead and is one other one of many video games that has a bigger common enchancment. That’s nonetheless not value $8,000 or $11,000 or no matter it’s they’re promoting these for, however it’s fascinating to see any distinction in any respect given the deal with so-called AI workloads.
Cyberpunk 2077: Phantom Liberty – 1440p
At 1440p, the PRO 6000 ran at almost 200 FPS AVG, about 9.4% forward of the 5090. This stays one of many bigger gaps for 1440p, though once more, it’s clearly not one thing anybody ought to remotely think about primarily for gaming. The 4090 is down at 135 FPS AVG, that means the PRO 6000 leads it by 48%.
Cyberpunk 2077: Phantom Liberty – 1080p
We’re leaving 1080p in as a result of it’s the one place we see a bottleneck. Lastly, we run into what looks as if a CPU restrict. The PRO 6000 and 5090 are at about the identical degree of efficiency, as is the 4090. All the pieces right here is restricted. Meaning our 1440p outcomes had been no less than considerably restricted as nicely, with minimally the higher sure bouncing off of limits.
NVIDIA RTX 6000 Ray Tracing Benchmarks
Ray tracing benchmarks are subsequent.
Ray Tracing – Black Fantasy: Wukong 4K
We’ll begin with the heaviest one, which is Black Fantasy: Wukong at 4K with upscaling.
On this check, the RTX PRO 6000 ran at 92 FPS AVG, which was barely any totally different from the 88 FPS noticed on the 5090. We’re underneath 5% of uplift right here. It’ll be fascinating to see if RT workloads don’t produce as a lot change as rasterization.
Ray Tracing – Black Fantasy: Wukong 1440p
At 1440p upscaled, the RTX PRO 6000 ran at 131 FPS AVG, main the 5090 by simply 3%. There’s simply not a lot change on this check usually.
We’ll skip 1080p given the dearth of enchancment.
Ray Tracing – Cyberpunk 2077 (4K, RT Medium)
Cyberpunk at 4K with RT Medium is subsequent. On this check, the RTX PRO 6000 ran at 65 FPS AVG, main the 5090 by 11%. That’s roughly in-line with what we noticed in Cyberpunk rasterized. That is with none upscaling, which can be serving to the PRO card’s relative acquire.
And eventually, in contrast to what we noticed within the 5060 and 5070 class playing cards with important VRAM limitations harming low efficiency, the PRO doesn’t have any VRAM points. That is clearly the answer. The answer is to purchase $8K-$11K video playing cards.
As earlier than, word that some playing cards with low VRAM, just like the 5070 and 4070 Ti, have averages that look higher than the fact (although the averages are additionally not nice).
Ray Tracing – Cyberpunk 2077 (4K, RT Extremely)
4K with RT Extremely is subsequent. We virtually by no means publish this check because it’s so intensive that it turns into considerably ineffective, however it’s fascinating right here.
The RTX PRO 6000 ran at 56.6 FPS AVG, main the 5090 by 6.4%. That’s a major drop from what we noticed at RT Medium, indicating that the rise in depth for the RT workload is minimizing the good thing about different features of the PRO 6000’s enhancements. It’s getting overrun by the RT workload depth. The lead over the 4090 is about 44% right here for the 6000, with the 5090 main the 4090 by 35%.
Be aware that the decrease portion of this chart is unreliable for common framerate for the reason that playing cards are all unplayable and stuttering. This will generally be resulting from exceeding VRAM limitations, similar to on the RTX 5070. The common appears to be like much better than the fact, although the common can also be unplayable.
Ray Tracing – Cyberpunk 2077 (1080p, RT Medium)
At 1080p with RT Medium, the PRO 6000 ran at 182 FPS AVG with lows paced proportionally. That has it forward of the 5090 by about 12 FPS AVG, or about 7%.
Ray Tracing – Dragon’s Dogma 2 4K
Dragon’s Dogma 2 with ray tracing and with out upscaling is subsequent. Examined at 4K native, the RTX PRO 6000 at 121 FPS AVG leads the 5090 FE by about 7% in common framerate. That is beginning to match a sample, so we’ll let this be our final sport check. Let’s transfer on to one thing else.
RTX 6000 AI Benchmark Charts:
Now, we’ll get into a few of our first ever LLM and machine studying benchmarks. As a disclaimer, we all know sufficient to get some charts collectively to run some benchmarks the place we really feel just like the controls of the check surroundings are good, however we don’t but know sufficient to have a full image of capabilities outdoors of the check suite that we’re utilizing presently.
We have examined with LM Studio, 3DMark testing, and ML Perf. We form of settled on one set simply to experiment for now. And that is our disclosure that we’re experimenting with new testing, which signifies that the charts get the experimental chart label on high of them. And that is simply so everybody’s conscious that these are lower than our full confidence requirements of our regular benchmarks that we do on a regular basis as a result of we’re nonetheless studying right here.
LM Studio Testing
For our LM Studio testing, we recorded the response speeds from 8 totally different fashions that incrementally improve in measurement.
We’ll get to bigger fashions in a second.
Our first chart illustrates the AVG tokens/second outcomes from the three smaller fashions – starting from 8.5GB to round 19GB in measurement. We expect these first three are extra consultant of like-for-like checks in comparison with another fashions we’ll be going over at present – as a result of any of the GPUs on our chart can load these fashions with out being restricted by VRAM capability.
In DeepSeek Llama 8B Distil, the RTX PRO 6000 is functionally tied for first place with the RTX 5090 FE – each with response speeds of roughly 81 tokens/second, bettering on the 4090’s 62 tokens/second AVG by about 30%.
In 8-bit Phi-4, the PRO 6000 places far between its personal tokens/second AVG of 62 and the 5090’s 51 tokens/second AVG, displaying a 22% enchancment. In comparison with the 4090, the RTX PRO improves by about 43%.
In Qwen 2.5, the PRO 6000’s 44 tokens/second AVG sees enhancements of 25% over the 5090 and 37% over the 4090.
Textual content Technology
Shifting to our subsequent chart, we’ll actually begin to see efficiency gaps start to widen as VRAM limitations take impact.
The RTX 6000 tops the chart in all 5 fashions.
In InternLM, the RTX PRO 6000’s 50 tokens/second AVG improves upon the 5090’s 40 by 25% and upon the usual 24GB 4090 by 319%.
In Mistral Small 26GB, the 4090 and 5090 see a lot sharper efficiency decreases than the PRO 6000. The workstation GPU achieves 42.4 tokens/second AVG, or 147% improved from the 5090’s 17 and 560% better than the 4090’s AVG of 6.4.
As soon as we attain Gemma 3 27B, the workstation card actually begins to separate itself from the lowly $2,000 gaming playing cards. On this mannequin, the RTX PRO 6000 achieves a 29 tokens/second AVG, with the 5090 solely reaching 5 and the 4090 solely seeing 4 tokens/second, give or take some change. The comparisons stay largely the identical in QwQ 32B.
Within the closing Llama 3.3 70b Q4_K_S mannequin we examined, the RTX PRO 6000 achieved its best lead but. The workstation card sees enhancements of 928% over the 5090 and 1141% over the usual 4090.
A extra simplistic approach to interpret this knowledge is: At a sure level, the opposite playing cards run out of VRAM, whereas the workstation GPU doesn’t even use half of its VRAM within the largest Llama mannequin we examined. You possibly can broaden past this as nicely, after all, however that’s the extent of our experimentation for now.
NVIDIA RTX 6000 Conclusion
That is it for the RTX Professional 6000 check. It is a evaluate in a means however on the similar time, we might wish to do much more checks within the ML class to essentially have a fully-fledged evaluate. This piece has been an thrilling experiment for us.
The largest discovery for us was seeing no liquid steel whereas we had been doing the tear-down. That was most likely recognized, however it’s not one thing that we had appeared into. In order that was fascinating. We may see why NVIDIA did that for reliability causes. We may additionally see it for legal responsibility causes if the corporate is fearful about liquid steel leaking out onto costly servers. We form of doubt this was accomplished to save lots of on prices, although perhaps it was accomplished for legal responsibility prices, however it simply would not seem to be NVIDIA goes to try to save a buck on this sort of card.
We spoke to Wendell from Level1Techs and he famous that there is some obvious buggy habits with Blackwell proper now and in order that’s nonetheless getting up to date and maybe there’s efficiency optimizations that could possibly be accomplished. On the aspect that we’re acquainted with, which is the gaming facet, seeing 5 to 14% enchancment is usually helpful not since you ought to ever in any means think about this card for gaming, as a result of that may be an insane waste of cash if that is all you’d do with it, however extra as a result of it exhibits us that there is room left within the 9800X3D for scaling, which is fairly cool.
