In Theory: Mapping the Next-Gen Xbox
Digital Foundry discusses the hardware options open to Microsoft for its new console
The Power Mac G5 tucked into the corner of an AV room of a highly respected UK games developer looked like it had seen better days - by the time I saw it back in 2008 it was already obsolete and gathering dust. But this was a machine with history: three to four years prior to my visit it first arrived at the office under heavy NDA - this was one of the first Xbox 360 developer kits.
Recent news from Develop suggests that the hardware cycle has begun again, that the first dev hardware for the next Xbox has been despatched to key Microsoft partners - in this case, Electronic Arts (who quickly released a blanket denial). Develop also cheekily suggests that the new hardware may be revealed at E3, an event where Nintendo is planning to reveal its contender for the next gen crown.
Recent investigations suggest that while the E3 reveal for the "Nextbox" is almost certainly not going to happen ("it's bullshit," a senior Microsoft source told me off the record), the notion of dev kits being out there now is not so far-fetched. The platform holder itself has also recruited top-tier engineers (for example Corrinne Yu at 343 Industries) and it's hard to believe that staff of that calibre are working on existing console architectures when their talents would be better used mapping out the future of gaming technology.
Microsoft has recruited top-tier engineers and it's hard to believe staff of that calibre are working on existing console architectures when their talents would be better used mapping out the future of gaming technology
However, in no way should this be considered any kind of evidence that Microsoft has locked down the design for its next-gen hardware, but what it does suggest is something more promising - that the new wave of consoles will not be rushed to market and that publishers and studios will have a lot of input into the design and the programming tools at a very early stage of its development. That's right about now.
So if Microsoft hasn't finalised the design for its next gen hardware, how can it possibly be in a position to despatch dev kits at all? Well, the answer lies in the make-up of that neglected Power Mac G5 I spied a few years back. The original Xbox 360 development hardware consisted of a dual core G5 IBM POWER4 CPU married up an ATI X800 graphics card, replaced at some point with a card from the X1800/X1900 line that more accurately represented the performance of the final production Xenos GPU. While there are similarities in the core architecture, this set-up had a great many differences compared to the final production hardware.
The Xbox 360 Xenon CPU has three cores at 3.2GHz just for starters, and while the ATI cards Microsoft supplied in lieu of Xenos were of similar ballpark power, they didn't utilise the eDRAM configuration found in the final console. This requires special care and attention from developers: elements like predicated tiling would have been challenging to emulate on the prototype hardware.
In short, Microsoft built up a hardware mock-up of its console that attempted to approximate the make-up of its final design using off-the-shelf hardware, presenting clear compatibilities that would make development easier. However, the differences between the dev hardware and the production console was such that there was little chance that any of the actual launch games would really be using anything like the true potential of the hardware.
The predicated tiling issue certainly had an impact: game-makers worked on the presumption that anti-aliasing would come for free with Xenos - indeed, one of the most important TRCs from Microsoft was that games should run at 720p with a minimum of 2x multi-sampling anti-aliasing. AA is indeed free on Xenos, but with the 10MB eDRAM limitation, certain launch games such as Perfect Dark Zero and Project Gotham Racing had to run at sub-HD resolutions in order get that boost to image quality.
We also need to consider that Microsoft was forced to rush to market with Xbox 360 in order to get its head start over PlayStation 3. At E3 2005, mere months before the console launched, most of the games on display were still running on the Power Mac G5 developer configuration. Indeed, it is rumoured to this day that one of the most impressive games on show actually required some deft overclocking to the dev kit's graphics core to get the game looking fast and smooth enough for presentation at the event.
Things are different this time. Despite the arrival of Project Cafe, Microsoft is under no real pressure to launch its next generation of console, likely to arrive in 2014 (though some rumours suggest that the platform holder may be ready to go one year earlier). There is certainly nothing like the crazy rush to bring a new machine to market in the way that the company was effectively forced to when it became clear that the original Xbox was finished.
While it may seem premature to be sending out dev kits now when the hardware design is most likely still in a state of flux, Microsoft will have clear ideas on the power level it expects from the final machine - the hardware will have been costed out, and the CPU and GPU can only be made at certain fabrication plants whose fabrication processes going forward are a matter of public record.
The Xbox platform holder also enjoys a clear advantage over its competitors: by owning the DirectX 11 standard, Microsoft works in concert with all GPU vendors in defining the roadmap of graphics technology going forward, and also gets the direct input of the leading rendering architects in the business on what they want to see. The challenge going forward is to finalise that hardware design and optimise DX11 to a degree that benefits the fixed architecture - something that most agree was achieved well in refactoring DX9 for Xbox 360.
What is a shame is that the current rumours do not give us any technical insight into the make-up of the dev kit, specifically the graphics processor. In its next-gen Samaritan demo, Epic Games targeted a Core i7 PC with triple GTX580s to give us its vision of how the next-gen would look. The next-gen Xbox would need to be considerably more conservative and efficient in its final design: factoring in power draw and silicon budgets, it would be hugely surprising if Microsoft brought a graphics processor to market that exceeded the capabilities of just one of today's top-end cards.
The challenge here remains in producing a cost-effective graphics core that isn't going to turn the console into a furnace
We can only speculate on the make-up of the final production console of course, but an interesting exercise might be to scale up the cost structure of the existing architecture. Let's put it this way: if a Nextbox launched in 2014 using the same budgets in terms of silicon die space as the Xbox 360, what could we get?
In terms of the graphics core, it's pretty hard to determine an equivalent scaled up form due to the ultra-fast eDRAM attached to Xenos, which also holds the eight ROPs. Indeed, Microsoft may prefer to bin eDRAM completely for the next console and use the silicon budget elsewhere. However, if eDRAM remains, for the same silicon budget as the launch Xbox 360, 40nm should give us around 30-40MB - perhaps enough for a full 1080p framebuffer with multi-sampling anti-aliasing. At 28nm the memory benefit scales up still further to 70-75MB. In all cases, 16 ROPs should be sufficient for supporting 1080p and 720p stereo 3D at 60Hz.
Scaling up the silicon budget for the launch 360 to factor in present day graphics cards at the current 40nm presents a difficult match. Right now the only fit from AMD's line would be the Juniper core, as found in the Radeon HD 5770. However, we'd still have around 20 per cent die-space spare and who knows what kind of customisations and extra power AMD could factor into that.
So what does this give us in comparison to the present day Xenos? An HD 5770, scaled up to the Xenos die-size, could give us 960 stream processors, 12 texture mapping units or 48 texels per clock. Xenos is approximately 240 SPs, so that's more than four times shading power per cycle when you consider efficiencies. Juniper has an 850MHz core clock, so that makes it even faster: 1.7 times Xenos's clock speed with almost seven times the shading power. Xenos has 16 TMUs, but only 16 texels per clock, yielding five times the texturing capabilities of the current chip - so, yay for filtering or higher precision textures.
In the here and now, the speculation is fun but we really won't have much more of an idea what's going on until we see how the 28nm Cayman cards from AMD pan out. The potential of what Microsoft could achieve with a 28nm graphics core with the same surface area as the launch Xenos is quite phenomenal - you can essentially exchange our notional Juniper core for the latest AMD Cyprus tech. Bearing in mind that at full load the current production Cyprus at 40nm can draw up to 190 watts of power, a 28nm die-shrink would be essential to incorporate this tech into a console.
AMD remains the more likely partner for Microsoft not just because of the issues it had with NVIDIA on the original Xbox GPU but also because of the challenges ahead in producing the necessary die-shrinks needed to make consoles cheaper and greener over the course of their lifespans. Intel leads the way with its "3D Tri-Gate" tech, but AMD isn't so far behind, but the position with NVIDIA is not so clear. Overall, the challenge here remains in producing a cost-effective graphics core that isn't going to turn the console into a furnace, and which can be scaled down in future to make for cooler, greener, cheaper consoles.
While AMD looks like the more preferable partner for the GPU element, it's unclear where Microsoft can go with its CPU for the Nextbox. An evolution in the form of IBM POWER7 architecture might seem logical, and Microsoft really wanted out-of-order execution for the 360, and never got it. POWER7 provides it, but in its current form it's big, expensive and power-hungry - basically all the things you don't want in a console CPU. (Updated: POWER7 does indeed have out-of-order execution.)
Intel and AMD may seem to be the obvious choices here but the question is really whether either company has the resources available to generate a bespoke CPU design for a new console. As crazy as it may sound, ARM's reduced power CPUs may be a good fit.
An alternative approach for Microsoft may well be to consider where PC gaming will be in a few years time and design around this notion. Right now we're still at the stage where modern architecture with two cores and four threads still manages to run most games really well - anything beyond four cores/eight threads seems like overkill when the silicon budget could be spent on GPU compute shading, opening the door to offloading elements such as AI and physics onto the graphics processor.
The real challenge with the next-gen of consoles comes down to the fact that there are key issues that prevent cutting edge tech being scaled down and fit into a small box
The platform holders also face a real issue in terms of sourcing appropriate RAM tech too. The usual convention is that memory increases by eight times between generations, giving us a target of 4GB for the Nextbox. The issue here concerns the format in which cutting-edge RAM is currently delivered. The current best option is GDDR5, which in itself presents three challenges: IO per chip (bits), latency, and cost.
GDDR5 is currently only 32-bits per chip, so you'd need four chips for a 128-bit bus. It can go up to eight banks of memory, but that means a more expensive 256-bit bus – something not typically associated with consoles. Assuming 4GB of RAM is off the table, this also means manufacturers must produce 512MB chips so that the console would have 2GB system RAM on the 128-bit bus. Latency also becomes an issue for the CPU: this might not be so much of an issue if there were enough threads, but it's still not ideal to have so much latency between the CPU and memory. The other option for RAM might be DDR4, but there's little news on when this will become a mainstream proposition.
RAM - and the lack of it - is probably the single biggest gripe developers have about the current generation architecture, but a related issue concerns transfer speeds: how quickly that space is filled with data. The next-gen consoles will be using optical drives, probably faster iterations of the Blu-ray tech currently seen in the PlayStation 3. We'd like to think that Microsoft will not pursue the mandatory install route, so a further challenge would be in streaming data to get players into the game as quickly as possible while still making the most of the extra memory.
This in turn leads us onto another controversial aspect of next-gen technology: onboard storage. Recent rumours suggest that Project Cafe relies on USB and SD card storage, with only 8GB of onboard flash for cache and downloads. Xbox 360 to this day retains two SKUs: one with a small amount of flash, the other with a fairly generous hard drive. Whether this becomes a core component of the machine as it did with PS3 remains to be seen, but developers will certainly be in favour and Microsoft clearly has much more to gain by including a decent hard drive than Nintendo in terms of online resources and services it can make money from.
At the end of the day, the real challenge with the next-gen of consoles comes down to the fact that there are key issues that prevent cutting edge tech being scaled down and fit into a small box. A true generational leap in power presents a range of technical challenges that the boffins have yet to come up with viable solutions for. It also explains quite nicely why Nintendo would concentrate on the controller as the focus for Project Cafe, and why the rumoured hardware specs are seemingly so close to current 360 and PS3 performance. After all, if the true technological advances aren't open to you, a faster GPU alone isn't going to present the kind of mass-market appeal you can build a console around...