Bounce to: Structure | Efficiency | Vitals | Conclusion

Sizzling on the heels of the Radeon RX 7900 XTX huge reveal final week, add-in board companions have been busy releasing their very own takes on AMD’s most interesting shopper graphics card. The Made By AMD (MBA) card is spectacular sufficient in its personal proper, however the stealthy black aesthetic lacks the wow issue related to prime aftermarket designs.

Seeking to set the file straight, Sapphire has launched the Radeon RX 7900 XTX Nitro+ Vapor-X beast of a card. Bettering ever-improving reference board designs is quick turning into troublesome for add-in companions, however, with a wealthy historical past of innovation and design inside Pulse, Nitro+, and Poisonous households, the Hong Kong-based firm has all the time remained a step forward.

Radeon RX 7900 XTX is initially supplied in Pulse and Nitro+ guises. Each exceed AMD’s 2,500MHz XTX increase pace by clocking in at 2,525MHz and a pair of,680MHz, respectively. It’s to the £1,200 Nitro+ Vapor-X we draw consideration.

There’s little incentive for Sapphire to ape the MBA board’s design. Going larger and taller with measurements of 320mm size, 135.75mm width and 71.6mm top – equating to a 3.5-slot card in regular parlance – Sapphire makes use of the identical Nitro+ design on XTX and XT variants.

– Commercial –

The hefty monster’s larger frequencies end in elevated energy ranges of as much as 420W for the whole card, in comparison with 355W attributed to MBA. We are saying as much as as a result of being a Nitro+ card means there’s a three-position swap which toggles frequencies. The default place allows switching of main and secondary modes by way of Sapphire’s TriXX software program. The subsequent place is the one to go for when guaranteeing headline frequencies of two,680MHz increase. Shifting over to the third setting, in the meantime, reduces increase pace to 2,500MHz – matching AMD’s card – and removes roughly 50W of board energy, although fairly why you’d need to run Nitro+ Vapor-X slowly begs a special query.

Sapphire makes use of an MBA-matching 20-phase energy provide for the cardboard, with 17 heading to the GPU and three for reminiscence. Rated at as much as 70A every, good luck in getting near 1,400W! Overkill in a great way.

Let’s not beat across the bush. Sapphire’s all-new heatsink design is a delight. The die-cast aluminium body carries three axial followers utilizing a twin ball-bearing design. A part of the Fast Join vary, the followers might be eliminated individually for upkeep and cleansing functions. The trio swap off at temperatures under 50°C after which ramp up in very orderly vogue, helped by what Sapphire says are the tightest fan-speed tolerances thus far.

We noticed not one of the sawtooth fan-speed points afflicting lesser options additionally carrying a zero-RPM mode; the swap between idle and better speeds is sleek and barely detectable, even in a quiet system corresponding to our take a look at rig.

AMD makes fairly the noise about having solely two 8-pin connectors on the MBA card; Sapphire sees issues in a different way through the use of three plugs enabling 525W with ease. Definitely is smart for an overclocked card, although one can argue the single-cable 12VHPWR connector, favoured by Nvidia and able to 600W, facilitates a tidier construct.

Nitro+ Vapor-X is an imposing answer weighing a considerable 1,955g. Moderately than imbue followers with compulsory RGB lighting, Sapphire takes a special tack to earlier designs, as presentation is kind of completely different to, say, the RX 6950 XT Nitro+ Pure.

As a substitute, the 71.6mm top is available in helpful as we’re handled to spectacular ARGB mild bars on each side of the heatsink. Managed by Trixx software program, the colourway might be set to numerous telemetry, together with fan pace or PCB temperature. These desirous to synchronise with the remainder of the system can achieve this with a three-pin ARGB header and cable to the motherboard.

Having a minimalistic shroud and lightbar-accentuated lighting works fantastically properly, lending the cardboard a high-quality look, with additional RGB behind the Sapphire brand on the rear. Sweeping step by step from one facet to the opposite, out-the-box results are easy and with out hotspots sometimes seen on poorer designs. It’s hypnotic to see waves of RGB run throughout the cardboard.

Because the lightbar is current on each side in equal measure, mounting a card in regular orientation leads to an identical illumination, although bouncing off the motherboard and chassis, the impact is perceived in a different way. Prime marks to Sapphire’s implementation.

Our benchmark outcomes are in, however it might be remiss to not first revisit what’s beneath the hood. Right here’s a recap on every part it’s essential find out about how AMD has gone about constructing a chiplet-based GPU structure endowed with RDNA 3 goodness.

RDNA 3 – Sooner And Smarter

Understanding the impetus behind AMD’s latest GPU structure first requires an appreciation of two elements affecting the manufacturing of cutting-edge silicon. As we transfer right down to ever-smaller nodes, the primary is a scarcity of scaling in two essential metrics: reminiscence and analogue I/O, each of that are liberally utilized in GPU design.

a method that has labored wonders within the CPU area with chiplet-based Ryzen processors since 2017

The purpose right here is it’s not price having these two aspects on the most recent, costliest nodes if there is no such thing as a quick profit to doing so, since you’re paying over the percentages for no discernible profit. Alternatively, logic – the constructing blocks for the core blocks – seems to nonetheless scale properly, which means it’s price investing in smaller processors for attendant good points.

The second level is expounded to the primary insofar as the price of transferring between the most recent nodes is turning into considerably dearer. Solely use, say, 5nm manufacturing for those who completely should, as whereas it makes highly effective chips bodily possible, there comes a time after they’re not financially viable. For instance, constructing a monster, monolithic 700mm² die on purely 5nm is an train in inflating end-user value to arguably unpalatable ranges.

– Commercial –

Such considering coalesces into AMD’s elementary fabrication technique for high-end graphics transferring forwards. Understanding GPUs are basically large calculators processing directions in parallel, why not break down the constituent components of the design into small blocks, which naturally yield higher, and lasso them along with high-speed smarts? It’s a method that has labored wonders within the CPU area with chiplet-based Ryzen processors since 2017. But when one thing was simple, it might have been carried out by now.

GPU Chiplets Present Distinctive Challenges

Although good on paper, there are myriad explanation why a chiplet-based GPU is a troublesome proposition, exemplified by the graphic under.

Leaping from a monolithic CPU design to a chiplet-based strategy requires invention of high-speed, low-latency interconnects between what are termed Core Advanced Dies (CCDs) and a central hub. That hyperlink is AMD-developed Infinity Cloth, which connects stated chiplets with 100s of wires between dies.

A unbelievable feat of engineering, AMD can’t take the identical strategy when a chiplet-based GPU, and the explanation has to do with the large quantities of information traversing a GPU at any second. Having upwards of 5,000 cores means the bandwidth required is an order of magnitude larger, and a normal Infinity Cloth strategy merely can’t carry that quantity of data in a design that has to promote for a set charge to generate income.

Placing this drawback into intelligible context, AMD estimates a multi-chip GPU wants over 10x the bandwidth supplied by the most recent Epyc server processors. What to do and learn how to do it, eh? That is the place sensible engineering involves the fore.

AMD’s strategy is to flip the issue on its head, and by that we imply construct a multi-core chip the place the bandwidth-hungry reminiscence partitions are manifested as smaller chiplets whereas the massive computational engines are stored in a bigger central die. For those who recall, Ryzen processors do the reverse by transferring cores to chiplets and reminiscence to the IOD.

RDNA 3 is subsequently constructed by having a number of Reminiscence Cache Dies (MCDs), fabricated on a 6nm course of, circling a bigger, 5nm-based Graphics Core Die (GCD). Every MCD carries L3 cache and GDDR6 reminiscence hyperlinks to the GCD, and accumulating the contents of every MCD gives a very good inkling of a card’s memory-side potential.

The GCD in the meantime, and identical to Highlander there is just one, is dwelling to acquainted Compute Items (CUs) resident with shader cores, ray tracing cores, texture models, and the like.

Linking MCDs to the GCD with the requisite bandwidth is the true triumph for RDNA 3. AMD calls this particular sauce Infinity Cloth Fanout Know-how (IFFT), illustrated within the slide above. Going by AMD’s displays, every wire is way, far smaller than its CPU counterpart. If we crunch the numbers, GPU-optimised IFFT is over 50x as dense and gives 10x larger bandwidth than Infinity Cloth On Bundle (IFOP) used for CPUs. Take a second for that to sink in.

We wanted a brand new bundle expertise and a brand new hyperlink expertise. And that’s what we have now carried out.

Sam Naffziger, AMD Senior Vice President and Product Know-how Architect

With out IFFT, make no mistake, there is no such thing as a chiplet-based RDNA 3, as scaling incumbent CPU IFOP expertise would end in large quantities of energy diverted solely for the interconnects, not to mention any die-space ramifications. However, constructing hyperlinks between chiplets isn’t a free power lunch. IFFT takes essential area and energy to carry out its high-speed information shuttles at 9.2Gb/s. We estimate including related hyperlinks, versus a purely monolithic design, will increase die space by shut to fifteen per cent whereas including treasured watts to general consumption.

As it’s, every TSMC-fabricated 6nm MCD weighs in at round 37mm² and homes some 2bn transistors. That’s a mixed 222mm² and 12bn transistors proper there for Navi 31 silicon. CU-housing GCD is a major 300mm², which means the mixed dimension of the Navi multi-chip GPU is 522mm².

An attention-grabbing query arises. Is it higher to have a 5nm monolith chip at 450mm² – the scale we estimate Navi 31 can be on a single piece of silicon; there can be no space- and power-taking IFFT hyperlinks – or a multi-chip 5nm/6nm at 522mm²? Not a straightforward one to reply, and one we’re positive AMD’s bean counters wrestled with.

One other drawback going through an IFFT answer is, on paper, considered one of latency. AMD acknowledges working Infinity Hyperlinks will all the time create extra latency than conducting transfers inside one monolithic die. This problem is proven by the Navi 31 bars marginally larger than Navi 21. Nonetheless, the mix of a 43 per cent larger IF clock and better recreation clock ameliorates the difficulty. In truth, AMD contends Navi 31 has decrease latency… however how a lot decrease might it have been on a monolithic design?

We come away from the chiplet dialogue with the considering that AMD has spent numerous time and useful resource in mitigating apparent points emanating from happening a multi-chip route on a GPU. The ache is well worth the long-term advantages, in line with AMD, nevertheless it looks as if numerous effort.

The chiplet strategy must play dividends additional down the road when we have now, say, 12 MCDs encompass one or two GCDs. Breaking down the components into manageable chunks, with presumably higher yields than a monolithic monster, is the place AMD is aiming. Whether or not it’s chiplets or tiles, we are able to foresee Nvidia undertake an identical strategy in years to return.

RDNA 3 – Effectivity At The Core

Feast your eyes on that, will you? Having marked off the primary a part of our dialogue, because it pertains to a chiplet design and the whys and wherefores of happening that street, the second is RDNA 3 structure.

In comparison with Navi 21, AMD claims a 54 per cent performance-per-watt enhance

It’s onerous to overlook AMD’s claims of RDNA 3 ‘architected to exceed 3GHz.’ This actually isn’t the case within the first two playing cards based mostly on this design, particularly Radeon RX 7900 XTX and XT, suggesting AMD hasn’t achieved lofty frequency aspirations within the first retail salvo. A misjudgement of the structure or comparatively poor frequency yields from foundry companion TSMC? In all probability a little bit of each.

However, there may be constructive information. In comparison with Navi 21, which you’ll know as high-end Radeon RX 6000 Collection, AMD claims a 54 per cent performance-per-watt enhance. That’s nothing to be sniffed at, in fact, and the remainder of our dialogue focusses on how AMD has achieved this quantity.

Efficiency naturally goes up for Navi 31 because it matches in additional Compute Items than its quick predecessor, and the 80CU-96-CU notation offers a 20 per cent increase proper off the bat. Shoving in additional CUs is solely anticipated when transferring between architectures and processes. Low-hanging fruit and all that.

But the satan could be very a lot within the particulars. Take one other have a look at the 2 CU blocks, or pairs. AMD refers to it as a single CU, of which, for those who recall, there are 96 within the Massive Boy design. Nevertheless it’s not one CU, is it, because the second is a mirror of the primary.

This has essential ramifications for efficiency, and why evaluating generation-on-generation is quick bordering on pointless. AMD has successfully doubled every CU’s ALU functionality – initialisms, ahoy – which means there’s double the throughput for floating-point duties on a shader-to-shader foundation.

Breaking it down, every RDNA 3 64-ALU block, or CU, has twice as many FP32 pathways of RDNA 2. Right here’s the place terminology turns into troublesome when referring to ALUs, FP32 models and INT32 models, however suffice to say, FP operations, that are a mainstay of gaming, are a lot improved.

On the integer facet, AMD has all-new AI Matrix Accelerators which embody bfloat16 and WMMA64 Dot4 assist, primarily to assist with convolution, which is without doubt one of the elementary operations of deep neural networks with demanding matrix computation. In different phrases, it’s AMD’s reply to Nvidia’s Tensor Cores. Little extra is understood of the expertise apart from fundamental specs and a purported 2.7x matrix pace enhance in comparison with RDNA 2.

Cache Me In

Having a extra sturdy, sooner core and reminiscence subsystem helps, however most iterative GPU architectures are inclined to dedicate rising quantities of area for bigger caches. Simply see what Nvidia did with RTX 40 Collection.

Going by way of the data, it’s simple to see the place AMD has doubled sizes in comparison with RDNA 2. Definitely not the large will increase instigated by Nvidia, but it stands to purpose that caches have to be upgraded to cope with extra innate processing energy.

The one outlier from this strategy is AMD Infinity Cache. These acquainted with how AMD builds current GPUs will know Radeons have a slab of goodly cache residing between on-chip L2 and exterior GDDR6. Name it an L3 if you’ll. RDNA 2’s Infinity Cache topped out at 128MB on Radeon RX 6950 XT. AMD drops it to 96MB for RDNA 3. If bigger caches are higher and there’s been loads of enhance on L0, L1, and L2, what offers? Good query.

It’s not the scale, it’s how you utilize it, appears to be AMD’s response. Moderately than scale this obvious L3 to 128MB+, which takes up helpful real-estate area, AMD has elevated the conduit between L2 and Infinity Cache by 2.25x. That’s no small potatoes. Aforementioned RX 6950 XT runs off a 1,024 bytes per clock and has an SoC working at 1.93GHz attaining complete bandwidth of 2TB/s. Navi 31’s greatest SoC operates at 2.3GHz. Do the mathematics and we attain 5.3TB/s at this juncture alone, with out taking the a lot wider, sooner GDDR6 into consideration.

In that regard, Navi 31 (RX 7900 XTX) has a most 384-bit reminiscence bus feeding GDDR6 reminiscence working at 20Gbps, whereas Navi 21 (RX 6950 XT) is just 256 bits broad and 18Gbps. 576GB/s can’t compete with 960GB/s on right this moment’s champ.

Ray Tracing Overhaul

AMD launched devoted {hardware} ray tracing models in RDNA 2. But whereas rasterisation efficiency was roughly analogous to price-comparable Nvidia RTX 30 Collection, RT lagged behind significantly. The hole has change into a chasm with the introduction of RTX 40 Collection, so in order for you best-in-class lighting, Group Inexperienced is the best way to go.

Seeking to arrest this gulf, AMD has invigorated the RT models inside RDNA 3. As RT {hardware} is tied to CUs on a one-to-one foundation, there may be extra ray tracing potential by way of sheer numbers – 96 Ray Accelerators vs. 80.

we’d put the best-case RT state of affairs as one matching premium RTX 30 Collection playing cards

AMD’s give attention to RDNA 3 isn’t a wholly new RT unit – that might be too pricey to implement at this stage – however considered one of enhancing what it has to work with. A part of this rests with effectivity, insofar as not doing work that’s of no use. RDNA 3 introduces an early subtree culling to take away pointless calculations by skipping components of the acceleration construction throughout traversal.

There are additional enhancements, too, from having extra rays in flight to pure uplifts brought on by bigger caches, CUs, and so forth, however the backside line is that mixed efforts are not any silver bullet; we’re not going to see Nvidia RTX-like numbers anytime quickly. In truth, although AMD contends an 80 per cent enchancment for RDNA 3 over RDNA 2, we’d put the best-case RT state of affairs as one matching premium RTX 30 Collection playing cards. RTX 40 Collection will stay in a special league.

Bits And Bobs

Head again up and look throughout to the developments within the pixel pipe. Complementing the elevated top-end muscle, AMD ups ROP rely by 50 per cent, from 128 to 192.

We touched on bandwidth hikes earlier, but it’s prudent to recap. AMD’s Infinity Cache is 165 per cent sooner and exterior reminiscence affords 67 per cent over the earlier era. You gotta feed the CU beast.

On the show facet of issues, AMD retains to HDMI 2.1a of the earlier era and provides in DisplayPort 2.1, which is a function lacking on rival Nvidia RTX 40 Collection playing cards.

A Twin Media Engine brings hardware-accelerated assist for AV1 encode and decode as much as 8K60, amongst different niceties, and outputs are such {that a} single card can drive 4 4K144 shows.

Enter Radeon RX 7900 XTX and XT

Radeon	RX 7900 XTX	RX 7900 XT	RX 6950 XT	RX 6800 XT
Launch date	Dec 2022	Dec 2022	Could 2022	Nov 2020
Codename	Navi 31	Navi 31	Navi 21	Navi 21
Structure	RDNA 3	RDNA 3	RDNA 2	RDNA 2
Course of (nm)	5/6	5/6	7	7
Transistors (bn)	57.7	57.7	26.8	26.8
Die dimension (mm2)	522	522	520	520
Compute Items	96 of 96	84 of 96	80 of 80	72 of 80
ALUs	6,144	5,376	5,120	4,608
Increase clock (MHz)	2,500	2,400	2,310	2,250
Peak FP32 TFLOPS	61.44	51.61	23.65	20.74
RT cores	96	84	80	72
AI cores	192	168	–	–
ROPs	192	192	128	128
Infinity Cache (MB)	96	80	128	128
Reminiscence dimension (GB)	24	20	16	16
Reminiscence kind	GDDR6	GDDR6	GDDR6	GDDR6X
Reminiscence bus (bits)	384	320	256	256
Reminiscence clock (Gbps)	20	20	18	16
Bandwidth (GB/s)	960	800	576	512
Energy (watts)	355	315	335	300
Launch MSRP ($)	999	899	1,099	649

All of our dialogue has rightfully centred on the overarching RDNA 3 structure and consequent Navi 31 GPU. Each phrases describe the accessible {hardware} instruments from which AMD constructs retail playing cards. These are Radeon RX 7900 XTX and Radeon RX 7900 XT, each hewn from Navi 31, albeit in a different way.

Beginning with the very best Radeon thus far, RX 7900 XTX makes use of the complete 96-CU complement accessible on the die – there is no such thing as a holding again, not like how Nvidia historically retains cores unused on shopper graphics playing cards. It’s an enormous beast, alright, with 57.7bn transistors unfold over the six MCDs and GCD. Basic math informs us 12bn of these go in the direction of the previous and 45bn to the cores.

Relying on how one defines ALUs RX 7900 XTX has both 6,144 (how we want to know it) or 12,288, as every pair is able to floating-point operations. It doesn’t matter, really, as peak TFLOPS is identical regardless of which facet of the ALU fence you sit on.

61.44TFLOPS is precisely 10 billion instances bigger than the standard ALU rely. We arrive at this determine by multiplying the height increase clock of two,500MHz by the operations per clock, which on this case is 4. Seen on paper alone, RX 7900 XTX batters RX 6950 XT – the very best of the final era – into submission by way of large firepower.

Making issues extra complicated, AMD decouples the front-end clock from the shader clock. The explanation to take action rests with saving energy on the cores by working them extra slowly, although with a lift clock of two,500MHz, each shader and front-end clocks change into synced once more.

Proper-o, let’s get to what you’ve been ready for.

Efficiency

Our logs present the Sapphire Nitro+ averages 2,794MHz throughout benchmarks, albeit with reminiscence at inventory speeds, naturally elevating efficiency above the MBA card by a discernible diploma.

Huge heatsink meets 420W GPU. The winner, girls and gents, is Sapphire’s triple-slot cooling.

Raytracing numbers are improved from the final era, in fact, but Nvdia’s peerless RTX 4090 is on a special stage.

There’s been loads of dialogue whether or not the mesh shader operation on Navi 31 isn’t working appropriately. Futuremark’s numbers assume so – even with the most recent updates – and we have now reached out to AMD for remark.

Murderer’s Creed Valhalla

4K100 with ease? No drawback for Sapphire’s best-ever card.

Cyberpunk 2077

The spectre of substandard raytracing efficiency continues to hang-out AMD. Not capable of match the very best of Nvidia’s last-generation playing cards, RTX poster youngster, Cyberpunk 2077, wants extra frame-boosting expertise to hit 60fps at a QHD decision.

Far Cry 6

A high-quality RX 7900 XTX is ready to squeak previous a far dearer RTX 4090 on this title. Meals for thought.

Remaining Fantasy XIV: Endwalker

One must count on glorious framerates in older video games. 4K100 is a breeze in Remaining Fantasy XIV: Endwalker.

Forza Horizon 5

One other recreation, one other 4K100-busting benchmark rating. Radeon RX 7900 XTX actually is a premium GPU; Sapphire provides the icing on high.

Marvel’s Guardians Of The Galaxy

The raytracing Achilles heel is in proof when run in essentially the most taxing mode.

Tom Clancy’s Rainbow Six Extraction

Radeon RX 7900 XTX is a superb card for video games counting on rasterisation to color pixels. Including raytracing to the combo pushes suggestions in the direction of GeForce.

Vitals

Geared up with 420W board energy, required to run the upper clocks, causes system-wide energy consumption to rise by a major quantity over the MBA card. However, there’s no have to go looking for a 1,500W PSU.

We double-checked figures to make sure the 57°C was correct. A unbelievable return when cooling so many watts, and even the hotspot temperature of 78°C is nothing to fret about.

Higher nonetheless, with the followers working at round 1,600RPM when gaming the cardboard is barely any louder than idle. Extraordinarily cool, extraordinarily quiet. Extraordinarily good.

Overclocking

Right here is the place it turns into actually attention-grabbing. Overclocking the most recent Radeons isn’t any simple process as there’s no good automated device. Moderately, one wants to extend the utmost energy price range by 15 per cent after which, counterintuitively, undervolt the cardboard. This mixture permits the GPU to scale frequency because it is aware of there may be energy headroom to take action. Growing clocks farther by pushing sliders to the suitable results in a trial-and-error strategy.

After an inexpensive time fiddling round making certain frequencies match as much as elevated efficiency, we settled on a 3,165MHz increase frequency and reminiscence working at an efficient 22.4Gbps. Right here’s the way it performs out in choose video games examined at 4K.

No GeForce RTX 4090-killing efficiency, however each little helps at 4K.

Conclusion

Sapphire understands AMD has raised the bar with in-house Made By AMD reference boards for its newest RDNA 3-powered GPUs. Taking every part good about these playing cards and including significant upgrades on high leads us to the Nitro+ Vapor-X.

Constructed like a tank and that includes essentially the most elegant RGB implementation we’ve seen thus far, the cardboard is an train in much less bling meets extra model. Beneath that hulking exterior lies top-notch cooling which tames the cardboard’s 420W energy price range with ridiculous ease.

Providing as much as 5 per cent additional fps than common Radeon RX 7900 XTX playing cards, Sapphire deserves credit score for reinventing its design strategy within the face of spiralling energy budgets and garish competitor playing cards. The corporate has left no stone unturned in a quest to construct the best RDNA 3 card attainable.

The selection between Radeon RX 7900 XTX and Nvidia GeForce RTX 4080 is a tricky one, but when your coronary heart is ready happening the AMD route, the massive, daring and delightful Sapphire RX 7900 XTX Nitro+ Vapor-X is the one to get.

– Commercial –