The bus has to be the same though right ? They aren’t altering the SOC. Can this be achieved ?
How many raid arrays to make up 2.4GB/s?
That’s a pretty tight fit to serve multiple XSX APUs on a single blade.
How to sell next-gen consoles, Marketing, Positioning, and Pricing [2020]
- Thread starter steveOrino
- Start date
How many raid arrays to make up 2.4GB/s?
That’s a pretty tight fit to serve multiple XSX APUs on a single blade.
You are wrong.
"Azure Managed Disks are the new and recommended disk storage offering for use with Azure virtual machines for persistent storage of data. You can use multiple Managed Disks with each virtual machine. We offer four types of Managed Disks — Ultra Disk, Premium SSD Managed Disks, Standard SSD Managed Disks, Standard HDD Managed Disks"
Lower latency and faster storage is necessary for a number of Azure use cases. They have a significant amount of SSD in their infrastructure.
MS's need for XSX blades means more GDDR6 and SoC purchases than they expect to sell XSX. So they'd have larger orders even if they sell less. If they're going to do a proper global rollout for streaming they're going to need a lot of those parts.
Edit: and yes, they will probably use SSDs for the XSX blades too. Power efficiency, heat, and rack space are serious concerns for something the scale of Azure. In some cases more so than the cost of the hardware itself.
Last edited:
D
Deleted member 11852
Guest
No. Semiconductors are qualified by the technical requirements of the customer. If the customer specifies a chip capable of X Mhz/Ghz and it hits X but not X+1 that the customer's problem.
If Sony commission AMD to design the chip and TSMC are commissioned to fab it, ether by AMD or by Sony. The way these negotiations typically happen is the designer (AMD) and fabber (TSMC) collaborate on layout (which AMD will already be familiar because TSMC's 7nm+ process is mature) and go through a small-scale fab-test-iterate process then TSMC will cite a price based on those results and give a estimate of cost over time for the future. Yield problems incurred through manufacturing problems, which is what the term actually means, are AMD/TSMC's problem.
Semiconductor fabrication that it not cutting edge (e.g. 5nm) is generally predictable and well understood. Even if there is a design issue such as an unexpected thermal issue in one quadrant of the chip, there are two solutions: a redesign of the layout or a different composition of wafer. Wafers come in literally tens of thousands of types and are comprised of different materials optimised for different types of product. If you need really high performant chips you'll use more expensive wafers, which may use more exotic materials (like rare earth metals) and be more expensive. Similarly you can have yield issues by virtual of contamination - poor wafer feed contaminating production lines so you need to clean it all.
Yield issues are rarely permanent problems because there are a variety of solutions to mitigate it. As long as the solution is cheaper than living with the yield issue, it's worth doing. The issue is about who picks up the cost.
D
Deleted member 11852
Guest
Yeah I don't think these are that common configurations in data centres given what they cost. How many high customers demanding such high I/O do you think are using Azure?
If you assume Microsoft are going to replace all their high-performance servers with XSX blades, which I reckon would reduce their global compute capacity by around 90%?
You think they will do with for Xbox, which has a current customer base of 40m users when Azure's whole revenue stream is orders or magnitudes larger?
Sure, I guess if they want to hand the server market to AWS, sure. 
p.s. I liked the opening "You are wrong" opening. It was very bold and dramatic.
how do they make up the bus deficiency on XSX? It's only 320 bit bus. Not enough DDR4 chips to support the 560GB/s. Is there a way around this? Is there such a thing as memory multiplexers? that aggregate low bandwidth into high bandwidth in both directions?Yeah I don't think these are that common configurations in data centres given what they cost. How many high customers demanding such high I/O do you think are using Azure?
If you assume Microsoft are going to replace all their high-performance servers with XSX blades, which I reckon would reduce their global compute capacity by around 90%?
You think they will do with for Xbox, which has a current customer base of 40m users when Azure's whole revenue stream is orders or magnitudes larger?
p.s. I liked the opening "You are wrong" opening. It was very bold and dramatic.
Let me restart and clear this up then. I understood your first point here before, but that is about contracts. The post I was initially replying to was about cost comparison between the PS5 APU and the XBSX APU. Ultimately, your position is that the PS5 APU is smaller, thus producing more chips per wafer and being cheaper. My confusion to this was that having a higher clockspeed meant that more of those chips would be rejected as compared to a lower cock rate. The PS5 being clocked at a much higher rate than the XBSX would mean that PS5 yields would be reduced and costs would be higher than for a smaller chip of the same clockrate as the XBSX. You are basically stating that this may be true early, but that reduced yield due to having a higher needed clockrate is something that can be fixed whereas the larger chip obviously cannot be shrunk.
Sound about correct?
D
Deleted member 11852
Guest
They have their ways. That's not really wide in server-terms. Yes. Yes.
Servers often solve problems in very different ways to consumer devices. Not always, many server problems are just the same as local PC problem but orders of magnitude larger.
Let me take a step back first, the higher clocks required by PS5 may mean it is using more expensive wafers than those used to provide APUs for XSX. But when people are talking about "yields" it is invariably in relation to the number of viable chips from a the wafer is different from what was predicted. This is a subtle but critical point. If Sony order 10m APUs from TSMC/AMD at $140 a piece, yield issues are not Sony's problem. They've paid a price per viable chip. Yield issues are usually where production deviated from testing and there is usually a reason for that and it's usually something the can be solved. The reason the cost of chips always fall over time is because it becomes cheaper to manufacturer then, on average, compared to earlier on. The reasons for this may be because an optimised layout was used, or different wafer, or different fabs, or a combination of all three.
This is the expertise for which Sony and Microsoft pay AMD and TSMC. But, Sony's APUs are smaller. So unless there is something particular abut the higher clocks at which they are qualified for acceptance, issues with yields for PS5 should be better. This is based on decades on consumer chip production demonstrating that larger quantities of smaller chips will be cheaper than smaller quantities of larger larger chips simply because the larger your die, i.e the greater surface area of the silicon (or volume in the case of 3D chips), the greater the increase of an imperfection taking it out - even with redundancy.
This is why there is a finite size of memory chips, and solid state cell chips. Going larger has not porven to be economical from a silicon production angle to make more individual smaller chips. While sometimes you can't go smaller (e.g. complex CPU, GPU designs), when you can, you do. When you do need to "go large", it's where people will pay regardless so yield issues increasing cost don't hamper commercial appeal.
I think my confusion was related to improved yields. I know they get better with time. Both time the node has been in use and the time a particular chip is being produced on said node. I just always thought the improvements in yield were limited to fewer defects making chips unsuitable. Not that the yield could increase the number of chips that hit the acceptable clock. In short - I thought the number of chips per wafer would not get better with time with regards to those being rejected due to not hitting the desired clocks. Only defects.
Learn something more every time I turn around here.
Silent_Buddha
Legend
I can't recall where I read it and I can't find it again at the moment, but I'm sure I remember reading that Microsoft had support for ECC memory built into the Anaconda SOC.
Regards,
SB
Yep they do, as GDDR6 didn't support ECC so they rolled their own implementation.
You made an assertion that MS's SSD capacity was definitely 1/1000th of their storage capacity or less. They don't need to be that common for you to be wrong about that.Yeah I don't think these are that common configurations in data centres given what they cost. How many high customers demanding such high I/O do you think are using Azure?
If you assume Microsoft are going to replace all their high-performance servers with XSX blades, which I reckon would reduce their global compute capacity by around 90%?
You think they will do with for Xbox, which has a current customer base of 40m users when Azure's whole revenue stream is orders or magnitudes larger?
p.s. I liked the opening "You are wrong" opening. It was very bold and dramatic.
And I think perhaps you misunderstood part of what I was saying too; I said nothing about MS replacing their existing non Xbox infrastructure with XSX blades. They are constantly adding capacity and creating new data centers, nothing is zero sum about Azure. They're replacing One S based blades and then scaling up the amount of XSX blades as they expand xCloud. It won't touch their other capacity, but it does represent spend on XSX components.
MS needs more XSX SoCs and GDDR6 modules than just what they need for retail XSXes; this should be obvious and straightforward. Sony can sell more PS5s than MS sells XSX but it won't necessarily mean MS hasn't bought more GDDR6 and XSX SoCs (and probably SSDs for use with them). That'll depend on how well xCloud does.
Because they need the XSX blades as close as possible to their users they have to invest fairly massively to make the service work; they can't offload to other hardware, so they'll need enough capacity to match, really exceed so there's some overhead, peak demand in each geographic region. If it takes off they'll need many millions of them. That won't happen overnight, but it is something they need to be preparing for. They'll need enough baseline capacity for the initial roll out, coming out of beta, and initial demand will be highest there.
I will grant that if they can work around differences with the ram speed without the solution requiring any developer work to account for, they might not need GDDR6. Given how close to the metal console games can get I'm not sure how feasible it is. But they still need the extra SoCs in all cases.
Got you, good memory. LOL . If you didn't jog my memory I wouldn't have remembered where to look.
Eurogamer
There are customisations to the CPU core - specifically for security, power and performance, and with 76MB of SRAM across the entire SoC, it's reasonable to assume that the gigantic L3 cache found in desktop Zen 2 chips has been somewhat reduced. The exact same Series X processor is used in the Project Scarlett cloud servers that'll replace the Xbox One S-based xCloud models currenly being used. For this purpose, AMD built in EEC error correction for GDDR6 with no performance penalty (there is actually no such thing as EEC-compatible G6, so AMD and Microsoft are rolling their own solution), while virtualisation features are also included. And this leads us on to our first mic-drop moment: the Series X processor is actually capable of running four Xbox One S game sessions simultaneously on the same chip, and contains an new internal video encoder that is six times as fast as the more latent, external encoder used on current xCloud servers.
D
Deleted member 11852
Guest
The term 'yield' can mean different things to different people, but when when people in the industry are talking about yields it is generally referring to the number of qualified ICs relative to what was expected or contracted to be produced. I don't know that PS5s APU costs less than XSX, it may not. But over time, it has greater scope to be cost reduced if yields (the number of fault/viable chips per wafer) is a problem - the scenario you presented. The same is true for XSX. I think folks tend to look at PS5's APU, particularly the GPU and see the high clocks and infer a production challenge. And this may be a challenge. XSX's CPU is clocker high and what also be a challenge.
But fundamentally, the more chips can fit on a wafer, the cheaper it is over a long haul and consoles are produced for quite a few years.
Good catch, I'd forgotten this. If Microsoft want to use server ECC RAM in Series X Azure blades they would need to engineer a very weird RAM design if they want to replace GDDR6 with whatever variant of ECC they can get cheap. Using ECC in Series X server blades would erode the case for the economics of scale argument of getting GDDR6 cheap, i.e. it wold ne less GDDR6 because they're only putting it in retails consoles.
Fair enough, conceded!
Fair point. We really know very little about the Azure angle of supporting the gaming business and we may get more on this when Microsoft release their 2020 Annual Report. Their 2019 report was published on 1 August 2019 but I assume COVID-19 has impacted the 2020 report. In last year's report they flagged Azure as an arm of the business receiving significant capital investment. It's surely won't all be Xbox-related but you would except this year's report to at least talk a little about how they spent that cash and because Microsoft are releasing new consoles this year, you'd export more detail on any spend to support this part of the business.
Before PlayStation 5 launches this holiday, we wanted to give you one more look at some of the great games coming to PS5 at launch (and beyond!).
Our next digital showcase will weigh in at around 40 minutes, and feature updates on the latest titles from Worldwide Studios and our world-class development partners.
Tune in live on Twitch or YouTube this Wednesday, September 16 at 1pm PDT / 9pm BST / 10pm CEST to see what’s next for PS5.
Our next digital showcase will weigh in at around 40 minutes, and feature updates on the latest titles from Worldwide Studios and our world-class development partners.
Tune in live on Twitch or YouTube this Wednesday, September 16 at 1pm PDT / 9pm BST / 10pm CEST to see what’s next for PS5.
D
Deleted member 7537
Guest
From a Spanish Retailer. Looks like PS5 will be sold at 399/499.
D
Deleted member 11852
Guest
I am so buying a Sony Áurea for 499.
D
Deleted member 7537
Guest
I am so buying a Sony Áurea for 499.
Áurea means gold, so there is that. Maybe you are buying a PS5 in gold. In any case could be that the guy just entered fake names and prices.
399$/€ really seems to be the sweetspot for PS5. If they achieve it, PS5 DE is going to sell like pancakes.
Hotcakes
My only gripe is that digital games are higher priced then disc based ones. I'm all for digital otherwise, on pc discs dont exist anymore.Dudes, Sony offered the Folding@Home with the PS3. Sony didn’t need special server class hardware for that function.
Investment in GP server hardware is a sunk cost. If MS simply used XSX blades as storage for recipes of ramen noodles lovers around the world, MS would only need enough revenue to cover software development and electricity costs.
I didn’t buy my car to move furniture and there are a ton of more capable vehicles for that function. But if someone paid me to move a few lamps, I’m not going to rent a UHaul.
It’s a 12 Tflops GPU with an 8 core Ryzen. With all the things moving to the cloud, I’m sure there are a ton of light to mid-weight processing that MS could migrate to XSX.
Not to say MS will convert everything to XSX blades or that MS will totally mitigate the cost of XSX server hardware with ancillary business. But if you have exa-flops lying around being underutilized, you can find someway to use that hardware outside of streaming games.
Investment in GP server hardware is a sunk cost. If MS simply used XSX blades as storage for recipes of ramen noodles lovers around the world, MS would only need enough revenue to cover software development and electricity costs.
I didn’t buy my car to move furniture and there are a ton of more capable vehicles for that function. But if someone paid me to move a few lamps, I’m not going to rent a UHaul.
It’s a 12 Tflops GPU with an 8 core Ryzen. With all the things moving to the cloud, I’m sure there are a ton of light to mid-weight processing that MS could migrate to XSX.
Not to say MS will convert everything to XSX blades or that MS will totally mitigate the cost of XSX server hardware with ancillary business. But if you have exa-flops lying around being underutilized, you can find someway to use that hardware outside of streaming games.
Last edited:
