Sony Reportedly Has RDNA 5 to Itself on Handhelds. The Leaked Chip Still Is Not a PS6

2026-10-26 · News and analysis

Affiliate disclosure: This guide contains affiliate links. As an Amazon Associate, we earn from qualifying purchases at no extra cost to you. Learn more.

Digital Foundry's Richard Leadbetter said this week that Sony has "an exclusive of sorts" on AMD's next graphics architecture for handhelds. Within hours it was everywhere, reported as Sony locking every rival out of RDNA 5.

Here is the thing. That phrase came from one sentence on a Patreon only podcast episode. Nobody outside that paywall has heard the full context. Every article you have read today is quoting the same hedged half sentence secondhand.

The claim is probably true anyway. It just does not mean what the headlines say, and the leaked specs for Sony's handheld tell a much more interesting story than "Sony wins." Run the arithmetic and you find a device with the same number of graphics cores as a handheld already on shelves, clocked at under half the speed, carrying sixty percent more memory bandwidth.

That is not a PS6 you can put in a bag. It is something else.

The Short Version

• What was said: Sony has "an exclusive of sorts" on RDNA 5 for handheld, per Richard Leadbetter on a Patreon exclusive Digital Foundry Direct. One sentence, no public transcript.
• The mechanism everyone repeats is the weakest part. AMD supposedly demanded ten million units before building custom handheld silicon. That claim comes from a single leaker, and a Windows Central reporter says it "isn't even slightly true."
• The mechanism nobody mentions is the one on the record. Under Project Amethyst, Sony's Mark Cerny has been co-designing the architecture itself. Sony is not a customer who won a bidding war. It helped write the thing.
• The leaked PS6 handheld is called Canis. 16 RDNA 5 compute units, four Zen 6C cores plus two low power cores, 192 bit memory.
• The leaked PS6 console is called Orion. 52 compute units at up to 3.6 GHz. That is roughly a ten times gap between the console and the handheld, inside one generation.
• Canis has the same compute unit count as a Ryzen Z2 Extreme you can buy today, at under half the clock speed. Its edge is memory bandwidth, not raw throughput.
• "Exclusive" means custom silicon only. Off the shelf RDNA 5 is coming, just not in a part that fits a handheld power budget.
• Nothing here ships before late 2027, and more likely 2028.

What Was Actually Said

The quote, as it reached the press, is that "there has been some behind-the-scenes discussions that Sony has an exclusive of sorts on RDNA 5 for handheld with its mobile device that's coming."

Read that again slowly, because almost every word is hedged. "Behind the scenes discussions," not a confirmed agreement. "An exclusive of sorts," not an exclusive. "Some," not sources plural and named.

That is a careful person flagging something he has heard without claiming he can prove it. Which is fine. It is how good reporters talk about things they half know. The problem is what happened next, where that hedging got sanded off on the way to a hundred headlines.

It also matters that the episode is Patreon only. There is no public transcript. Nobody writing about this today has quoted more than the one fragment above, which means nobody outside the paywall knows what qualifiers came before or after it.

So treat this as a credible hint from a credible person. Not a scoop with receipts.

The Mechanism Everyone Is Repeating, and Why It Is the Weakest Link

Here is the explanation attached to nearly every version of this story.

AMD has supposedly become choosy about building semi custom chips. Semi custom is the division that made the PS5 and Xbox Series silicon, and the Steam Deck's Aerith. The claim is that AMD now wants a guarantee of about ten million units before it will spend engineering time on a bespoke handheld part, and that Microsoft walked away from a first party Xbox handheld rather than sign up for that.

The volume arithmetic does make it plausible. Steam Deck is estimated around five million lifetime. Asus ROG and Lenovo Legion handhelds are estimated at one to two million each. Nobody in that group can honestly promise ten million. Sony, with the PlayStation brand behind it, is the only company that could.

So the "exclusivity" would not be a deal at all. It would be arithmetic. Sony is not the winner of an auction. It is the only bidder who clears the minimum.

That is a tidy story. Two problems with it.

The ten million figure comes from a single leaker, KeplerL2. And Windows Central's Jez Corden, who covers Xbox more closely than almost anyone, is reported as saying the claim "isn't even slightly true."

Then there is the obvious hole. Microsoft never announced a first party handheld. You cannot report the cancellation of a product that was never a product. What shipped instead was the ROG Xbox Ally line, which is an Asus device with Xbox branding and an off the shelf AMD chip. That is a different thing entirely, and it is doing fine.

The Mechanism Nobody Is Citing

There is a much better explanation sitting in plain sight, and it does not need a single leak.

Sony and AMD have a public, announced, jointly promoted collaboration called Project Amethyst. Both companies have talked about it on camera. Out of it came three technologies they have now described publicly: Radiance Cores for ray tracing, Neural Arrays that link the machine learning hardware across compute units so a large chunk of the screen can be processed at once, and Universal Compression that compresses data types wherever it can to cut bandwidth demand.

And Mark Cerny, the architect behind the PS4 and PS5, has said that big chunks of RDNA 5, or whatever AMD ends up calling it, are coming out of engineering he is doing on the project.

Sit with that for a second. Sony did not buy early access to this architecture. Sony helped design it.

That explains "an exclusive of sorts" better than any minimum order quantity ever could. When your own lead architect has been inside the design for years, you do not need a contract to be first. You are first because the silicon has your fingerprints on it. And a company in that position naturally gets the first implementation while everyone else waits for the general release.

The confirmed part of this story explains the rumoured part better than the rumoured mechanism does.

One naming note while we are here. "RDNA 5" is not an AMD product name. AMD folded its gaming RDNA line and its compute CDNA line into a single architecture called UDNA. Everyone keeps saying RDNA 5 because it is the obvious next number, and Cerny himself hedged with "or whatever AMD ends up calling it." When you see RDNA 5, read it as shorthand for the next architecture after RDNA 4.

The Spec Math

Now the part the headlines skip. Here is the leaked Sony handheld chip, codenamed Canis, against hardware you can actually buy, and against the PS5.

Bandwidth is calculated as bus width in bits divided by eight, multiplied by memory speed. Raw FP32 is compute units multiplied by 128, multiplied by clock speed. We ran all of these ourselves so you can check them.

Look at the Canis row next to the Ryzen Z2 Extreme row.

Same compute unit count. Under half the clock speed. Sixty percent more memory bandwidth.

In raw FP32 throughput, the leaked 2027 PlayStation handheld is slower than a Windows handheld you can buy this afternoon. That sounds damning until you notice what it does have, which is 205 GB/s of memory bandwidth against 128 GB/s.

Our calculated 205 GB/s also lands right on the roughly 200 GB/s the original leak claimed, which is a small but real sign the memory figures in this leak are internally consistent. Good leaks usually are. Bad ones usually are not.

Memory bandwidth is the thing that actually strangles handheld graphics chips. All of these parts share one pool of memory between the CPU and the GPU, and 128 bit buses have been the ceiling for years. A 192 bit bus is the first genuine widening anyone has put in a handheld class part. Add an architecture claimed to be fifty to one hundred percent faster per compute unit than RDNA 3, plus dedicated ray tracing and machine learning hardware from Project Amethyst, and low clocks stop looking like a weakness.

Low clocks are how you get battery life. A 16 compute unit chip at 1.20 GHz on a 3nm process sips power compared to a 16 compute unit chip at 2.7 GHz. Sony appears to be spending its silicon budget on width and bandwidth instead of frequency, which is exactly what you would do for a device that has to last a plane ride.

The PS6 Handheld Is Not a PS6

This is the part that answers the question people are actually asking.

The leaked PS6 console, codenamed Orion, is reported at 52 RDNA 5 compute units running up to 3.6 GHz, on a 280mm² die with up to a 160W power budget and GDDR7 memory. Canis is 16 compute units at about 1.20 GHz on a 135mm² die.

Both are the same architecture, so you can compare them fairly by multiplying compute units by clock speed. Orion lands at 187. Canis lands at 19.

That is a gap of roughly ten times, between two devices in the same console generation, from the same company.

A ten times gap is not a docked and handheld mode difference. It is a different class of machine. Whatever Canis is, it cannot natively run PS6 games built for Orion. The leaks saying it plays PS5 and PS4 titles are describing the only thing a chip like that could plausibly do.

So the honest read is that Sony is not building a portable PS6. Sony is building a portable PS5 with a back catalogue and a streaming client, sold inside the PS6 generation. Which, if you remember how the PlayStation Portal landed, is a considerably more ambitious version of a thing Sony already tried.

While we are checking arithmetic, one circulating figure does not survive it. Orion is widely reported at 34 to 40 TFLOPS. But 52 compute units at 3.6 GHz gives about 24 TFLOPS using standard RDNA math, or about 48 TFLOPS if RDNA 5 doubles up FP32 per clock the way RDNA 3 did. To land at 34 to 40 you need roughly 182 to 214 FLOPs per clock per compute unit, which matches neither. One of those numbers is wrong, or RDNA 5 changes the maths in a way nobody has explained. We are not publishing a derived TFLOPS figure for Orion, and you should be suspicious of anyone who does.

You should also be suspicious of "faster than a PS5" headlines about the handheld. 2.5 TFLOPS against the PS5's 10.3 is not faster. Architecture gains and machine learning upscaling could plausibly close a lot of that at a lower internal resolution, and that is a real thing modern consoles do. But there is no leaked benchmark here. Only a spec sheet and some optimism.

Does This Actually Lock Everyone Else Out?

No, and this is where the precise wording matters most.

The claim is about custom silicon. It is not a claim that no handheld will ever have RDNA 5. Those are very different statements, and almost every write up has merged them.

Off the shelf RDNA 5 is reportedly coming. AMD's rumoured Medusa Halo, the Ryzen AI MAX 500 series, is said to pair Zen 6 with RDNA 5 in 2027 or 2028, possibly with LPDDR6 on a 256 bit bus for around 460 GB/s. Any manufacturer could buy that without promising anyone ten million units.

Except Medusa Halo is a 55W to 120W class part. That is mini PC and thin laptop territory, not a handheld you hold for three hours.

And the off the shelf part that does fit a handheld, Medusa Point, which is the lineage the Z2 Extreme comes from, is reportedly staying on RDNA 3.5 or a variant some reporting calls RDNA 4m. The same reporting has RDNA 3.5 living on in mainstream parts into 2029.

Put those together and you get a two step conclusion that nobody seems to have printed. In the 15W to 30W band where handhelds actually live, the part with RDNA 5 is too hot, and the part that fits does not have RDNA 5. So Sony's advantage holds for most of this decade. Just for a reason that has nothing to do with an exclusivity agreement.

Who is actually affected, then:

• A first party Xbox handheld. Would need bespoke silicon. Genuinely blocked, if the volume story is true at all.
• Valve. Would need bespoke silicon too, which is the whole question around the Gainsborough leak. More on that in a moment.
• Asus, Lenovo, MSI, GPD, OneXPlayer. Buy off the shelf AMD parts and always have. Largely unaffected on this horizon, because they were never getting custom silicon anyway.
• Anbernic, Retroid, AYN, Powkiddy, TrimUI. Build on Snapdragon, Allwinner and Rockchip. Completely unaffected. None of this touches the budget retro handheld market at all.
• Intel. Quietly interesting. The Arc G-Series handheld wave is the one competing platform that does not depend on AMD's roadmap decisions.

About Our Own Steam Deck 2 Article

We have to flag this one, because it is our page and it is now in tension.

On 1 October we wrote up the Gainsborough APU leak, an AMD semi custom chip widely assumed to be the Steam Deck 2 part, with claimed specs of 4 to 8 Zen 6C cores and 12 to 24 RDNA 5 compute units on TSMC N3P. If Sony has handheld RDNA 5 to itself, a Valve chip with RDNA 5 is hard to square with it.

Here is what makes it stranger. Both leaks come from the same person, Moore's Law Is Dead. And he himself noted that Gainsborough might be close to Sony's Canis, with both on the same 3nm class node.

Our October article listed four possible explanations for the oddly wide Gainsborough spec range. Two of them now look a lot more likely. Possibility three was that Gainsborough is not one product and the range covers more than one customer. Possibility four was that Gainsborough is not the Steam Deck 2 at all.

We cannot tell you which. What we can tell you is that the Steam Deck 2 having an RDNA 5 chip and Sony having handheld RDNA 5 exclusivity cannot both be straightforwardly true, and both claims trace back to one leaker. We have added a note to that article too.

What This Would Mean for Emulation

Our usual job here is to extrapolate to emulation, so here is the unglamorous version.

Directly, almost nothing. Canis would be a locked PlayStation console. Locked consoles do not run RetroArch, RPCS3 or Dolphin. Whatever that chip can do, you will not be pointing it at a GameCube ISO. Compare it to the PS5 homebrew scene, where the interesting work only became possible years after launch through a security hole, not through a feature.

Indirectly, the bandwidth is the exciting bit. If a 192 bit bus becomes normal for handheld class silicon, that helps everything. GameCube and Wii emulation at 4x and 5x internal resolution is bandwidth hungry. So is PS2 upscaling. So are the FSR and XeSS paths that cost GPU time a Steam Deck does not have spare.

And here is the counterintuitive one. Canis is reported with four Zen 6C cores plus two low power cores for the operating system. A Ryzen Z2 Extreme has eight Zen 5 cores. For the workloads that actually define the ceiling on a PC handheld, which are PS3 emulation in RPCS3 and Xbox 360 emulation in Xenia, the bottleneck is CPU behaviour and thread count, not graphics throughput. Four big cores on a newer architecture would not automatically beat eight older ones there.

Newer silicon does not mean better at everything. It means better at the things it was designed for, and Canis was designed to run PlayStation games at a good battery life, not to brute force an emulator.

The old lesson still holds. Emulator maturity is the gate, not hardware. PS4 emulation is limited by how complete shadPS4 is. Switch 2 emulation has no commercial game running at all. Faster chips do not boot games that do not boot.

What To Do Right Now

Nothing. Genuinely nothing.

The chip enters production in mid 2027 at the earliest, with a launch window of late 2027 or 2028. Sony has announced no console, no handheld, no price and no date. Every number in this article is a leak, and specs on silicon this far out change constantly. Our Gainsborough piece makes the same point about Valve, and it is still true: there will always be a device two years out.

It is also worth remembering that memory prices have been climbing all year, hard enough that the Steam Deck OLED took a price increase. A handheld that needs 24GB minimum and prefers 32GB, which is what developers are reportedly asking for on Canis, gets designed against 2028 memory costs. Nothing about the current trend suggests that will be cheap.

If you want the best SteamOS handheld today, the Steam Deck OLED is still the one we recommend most, and our Steam Deck OLED review explains why it keeps holding up. SteamOS getting out of the way is most of the reason it is the better emulation machine of the premium options.

If you want maximum raw performance now and Windows does not bother you, the ROG Ally X is the Z2 Extreme class device to look at. It is also the realistic floor for anyone attempting PS4 emulation.

Our best PC gaming handhelds guide covers the whole field, and the Steam Deck versus ROG Ally X versus Legion Go S comparison is the head to head if you are choosing between the premium options today.

Frequently Asked Questions

Does Sony really have exclusive rights to RDNA 5 for handhelds?

Probably something like it, but not in the way the headlines suggest. The claim is one hedged sentence from Richard Leadbetter on a Patreon exclusive Digital Foundry Direct, describing "behind the scenes discussions" and "an exclusive of sorts." There is no confirmed agreement, no public transcript and no second source. It most likely refers to custom silicon only, not to RDNA 5 appearing in handhelds at all.

Why would Sony get RDNA 5 first?

The most likely reason is public and has nothing to do with exclusivity deals. Under Project Amethyst, Sony's Mark Cerny has been co-designing the architecture with AMD, and he has said big chunks of it come out of engineering he is doing. Sony is a co-author of the design rather than a customer buying early access, which naturally puts it first in line.

What are the leaked PS6 handheld specs?

The chip is codenamed Canis. Leaks describe a roughly 135mm² die on TSMC's 3nm process, with four Zen 6C cores plus two low power Zen 6 cores, 16 RDNA 5 compute units at roughly 1.20 GHz in handheld mode and 1.65 GHz docked, and a 192 bit LPDDR5X-8533 memory bus. Developers are reportedly asking for 24GB of memory as a minimum. None of this is confirmed by Sony, and the clock speeds are the least corroborated part.

Would the PS6 handheld be faster than a Steam Deck or a ROG Ally X?

In memory bandwidth, clearly yes. Its 192 bit bus works out to about 205 GB/s, against 88 GB/s on a Steam Deck and 128 GB/s on a Ryzen Z2 Extreme. In raw graphics throughput, no. It has the same 16 compute units as a Z2 Extreme at under half the clock, which is about 2.5 TFLOPS against 5.5. Architecture gains and machine learning upscaling would close some of that, but it is not a brute force win.

Is the PS6 handheld actually a PS6?

No, and the leaked numbers make that fairly clear. The PS6 console chip is reported at 52 compute units up to 3.6 GHz while the handheld is 16 compute units at about 1.20 GHz. Both are the same architecture, so comparing compute units times clock gives a gap of roughly ten times. A handheld that far behind cannot natively run games built for the console, which is why the leaks describe it playing PS5 and PS4 titles instead.

Does this stop the Steam Deck 2 from using RDNA 5?

Possibly, and that is the open question. Valve would need custom silicon, which is exactly the thing the claim covers. It also creates a contradiction, because the Gainsborough chip widely assumed to be the Steam Deck 2 part was reported with RDNA 5, and both that leak and the Sony one come from the same leaker. Those two claims are hard to reconcile, and nobody has resolved them yet.

Will any other handheld get RDNA 5?

Eventually, but probably not soon in a handheld power budget. AMD's rumoured Medusa Halo pairs Zen 6 with RDNA 5 but is a 55W to 120W class part, which suits mini PCs rather than handhelds. The part that does fit a handheld, Medusa Point, is reportedly staying on RDNA 3.5 or a variant of RDNA 4, with RDNA 3.5 reported to continue in mainstream chips into 2029.

Does any of this affect budget retro handhelds?

Not at all. Anbernic, Retroid, AYN, Powkiddy and TrimUI build on Snapdragon, Allwinner and Rockchip chips, so AMD's roadmap decisions do not reach them. If you are shopping in that part of the market, this story has no practical effect on what to buy.

Would the PS6 handheld be good for emulation?

Directly, no. It would be a locked PlayStation console, and locked consoles do not run emulators. The PS5 only got a homebrew scene years after launch through a security flaw. Indirectly it is more interesting, because a 192 bit memory bus becoming normal in handheld class silicon would help bandwidth heavy emulation everywhere. Worth noting its four main CPU cores are fewer than a Z2 Extreme's eight, which matters for CPU bound emulators like RPCS3 and Xenia.

When would the PS6 and its handheld launch?

Leaks point to the Canis chip entering production in mid 2027, with a handheld launch in late 2027 or early 2028, and a similar window for the console. Sony has confirmed nothing. Treat the whole timeline as an estimate built on unverified information.

Related Reading

• The Steam Deck 2 APU Leak: Zen 6, RDNA 5, and the Low End Problem
• Best PC Gaming Handhelds in 2026
• Steam Deck OLED Review
• Steam Deck OLED vs ROG Ally X vs Legion Go S
• FSR and XeSS on Handhelds: What You Actually Get
• Intel Arc G-Series Handhelds Explained
• PS5 Homebrew: Switch and PS3 Emulation on a PS5
• PS4 Emulation on Handhelds in 2026
• Two New PS5 Emulators Are Real
• The RAM Shortage and Handheld Prices

Written 26 October 2026. The "exclusive of sorts" claim originates with a single hedged sentence from Richard Leadbetter on a Patreon exclusive Digital Foundry Direct episode and reached us through secondary coverage dated 26 October 2026. We have not heard the episode. Project Amethyst, Radiance Cores, Neural Arrays and Universal Compression are publicly announced by Sony and AMD and are the only confirmed elements of this story. All Canis, Orion, Gainsborough, Medusa Point and Medusa Halo specifications are unverified leaks attributed to Moore's Law Is Dead and KeplerL2, reaching us through third party reporting. The AMD ten million unit claim is disputed by a Windows Central reporter, and Microsoft never announced the handheld it supposedly affected. Every bandwidth and FP32 figure in the comparison table was calculated by us from published specifications so that readers can check the arithmetic. Sony has announced no PS6 console, no PlayStation handheld, no specification, no price and no date. Nothing here is a product announcement.