If it ever ships, which at this point is not exactly guaranteed. When a company announces pre-orders for something, and then drops the date back by an entire year you should run for the refund button ASAP.
Disappointing, but I imagine AMD would've been happy to sell some old Bristol Ridge stock (even if for a knock-down price), and it's not as if you'd need a powerful APU for the task. Just as well really, as these chips hit their power limits quickly and throttle back.
If this is a new chip, doesn't it require new masks? If it does, aren't new mask sets around 80 million?
Supposedly AMD was getting around 80% yields on Ryzen in 2017. This chip is 2x the size at around 400mm^2. If we assume the same yields, then that's around 110 chips per wafer. That comes out at around $42 per chip, but some probably can't hit that clockspeed, so they either have to be sold in other markets (what I expect) or are wasted. If new masks are needed, then there's a big additional cost too.
That said, that's a big die. 209mm^2 is enough for a 4-core ryzen module plus about half as many CUs (2400G). Even if half the die were dedicated to the GPU (die shots show it to be quite a bit less than that), we wouldn't expect more than an extra 100mm^2. I would have personally guessed at around 275mm^2 to simply extend the 2400G. Either they are spacing the transistors way out or there's something extra happening here.
They will not need to create new masks for any of this. They will just use a four core CCX (110mm2)just like any other APU. They got plenty of those laying around since all Ryzen and Epyc uses the same CCX. Based on the CU count on the Vega GPU you are looking at an equivalent sized chip to a RX 570 which comes in at 220mm2.
They produce these on the same die. They create APUs by connecting the CPU and GPU via Infinity Fabric at the test pack facility.
The 393mm2 estimate by Anandtech was just that, an estimate.
However, using GDDR5 may add some area due to it being a different memory type from a standard APU. This is just speculation, on my part.
However, the fact that the CPU and GPU parts needed are produced on every Vega and Ryzen wafer produced makes this incremental business for AMD. No added costs. Especially since semi-custom projects require the majority of R&D to be covered by the customer, not AMD.
Not really. Zen cores and Vega CUs can have their own Layer Mask in a Multi-Layer Mask. There is no need to go for Full Maskset / Single Layer Maskset. This is the preferred method for medium size production volumes.
Also, by going to https://www.euvlitho.com/2017/P33.pdf [EUV MASK TECHNOLOGY AND ECONOMICS: IMPACT OF MASK COSTS ON PATTERNING STRATEGY], on page 11/26 [DESIGN COST FOR SOC] we can see that:
1. The mask set cost is about 2.5% of the total cost 2. Design cost and embedded software cost are more than 90% of the total cost
AMD semicustom design reuses Zen cores and Vega dies in both design and embedded software so the total cost of this APU is an order of magnitude smaller of a full (re)design.
I have no idea how modular you can do masks, but as you say GDDR5 instead of HBM my not be minor and I wonder if they added or replaced it. Longer term I'd like to to see lots of APUs clustered into tight form factors with Ininity Fabrics of small to vast scale: Want more power? Add more cartridges, power and cooling!
Interesting release for sure. When the 8th gen consoles came out, it would be years before a standalone APU could be bought at the performance of the PS4's GPU, and even recent ones lag in some ways. The way things are going now, APUs could be pretty close to what the PS5/XB? are looking like they're going to be, with Ryzen+Vega/Navi, and another year or two and die shrink down to add CUs and cores.
It has nothing to do with what's possible. It has to do with what's profitable. There is absolutely nothing special about the PS4 or Xbox One's SoC. If AMD wanted to launch a nearly-identical APU they could.
They don't for many reasons, but two big ones are that the CPUs on those chips aren't very powerful (and more powerful ones would take up valuable space) and AMD makes more money selling you a separate CPU and GPU than they would from an APU.
As for what the next Xbox and PS consoles will have on board, they're constantly working on new ideas for that, it won't be settled until they have a good business case to launch something. What they have now could be tossed out the window tomorrow.
As this product is mainly a combination of a number of existing products, much of the masks will just be a cut-and-paste job from those products. This should greatly reduce the cost of producing the set of masks. If each of the functional blocks (CPU, GPU etc) has a tiny gap (1000nm would suffice) between it and the next functional block then there should be no optical interference between the masks of the different blocks hence no need to redesign them.
One would assume China first and ROW later if the economics pan out. Unoless Xbox is licensed to this, there is little chance it will go anywhere. Maybe Linux/Steam could use this hardware into a gaming alternative to mid-range rigs ?. Or if big player like TenCent put their customised Android on it with their games and use Kinect-like gesture accessory to make a gamebox. Though China has strong regulations on the types of games permitted ....
Doubtful we'll see this SoC in a future XBox or PS variant. The CPU is much faster than what either of them are using now, but the GPU would be a major downgrade from the top end XBox and more or less a sidegrade on the PS. OTOH I suspect that when both consoles have their next generation high end refreshes they'll probably jump to Zen cores of some flavor to give the CPU a major boost in addition to the larger and more capable GPU.
An interesting product indeed (except the pricing). BTW as one of their content partners (game dev), I'll let them know about the interests from AT community and establish contact with Ian if they have bandwidth. :)
I've wondered about that and I hope someone in the know can share some data points here. On the other hand, do CPUs actually abandon cache line loads, if they notice code only asked for a single word and would now prefer another outside the current GDDR burst or cache line?
For code, I don't think it's much of an issue, that better be well inside the L1-3 caches anyway. Data may be all spread around, but mostly for old style code that's either fast enough or won't use multi-core anyway. With HPC or games much of the data can only benefit from bandwidth and people have taken great care to block things together for speedy access: They now get more benefit but no longer need to worry about copying things forth and back to the GPU.
And HSA workloads!! They become so much more attractive...
Really all we need is the ability to put lots of these nodes into really big NUMA over Infinity Fabric, I guess.
That console/PC combo would be a good value proposition for the US and Europe for about $ 600, and could even be the "steam machine that actually could" match or beat consoles, that is. I would be a possible customer, if the specs hold, the price is right, and a large player like Valve is behind it. Maybe Valve could even have another go at getting Steam OS going (Linux-based, if memory serves), and actually make it run games faster than WIN 10 this time. This setup has the oomph to do it.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
35 Comments
Back to Article
a1exh - Monday, August 6, 2018 - link
Could this be the same AMD SoC powering the Atari VCS?mode_13h - Monday, August 6, 2018 - link
No. Sadly. They already listed the specs, and that thing isn't using Zen cores *or* Vega.But it *will* be a fair bit cheaper than $625.
mode_13h - Monday, August 6, 2018 - link
According to https://www.indiegogo.com/projects/atari-vcs-game-...it will use A10-9630P with 4 (single-threaded) Excavator cores and 6 ancient GCN v1.2 CUs. And only 4 GB of DDR4-2400. Shipping in only 11 months.
The only sort-of bright spot is the base price of $239.
Flunk - Monday, August 6, 2018 - link
If it ever ships, which at this point is not exactly guaranteed. When a company announces pre-orders for something, and then drops the date back by an entire year you should run for the refund button ASAP.silverblue - Tuesday, August 7, 2018 - link
Disappointing, but I imagine AMD would've been happy to sell some old Bristol Ridge stock (even if for a knock-down price), and it's not as if you'd need a powerful APU for the task. Just as well really, as these chips hit their power limits quickly and throttle back.Flunk - Monday, August 6, 2018 - link
"Atari" can't afford semi-custom hardware, they had to pick off the shelf.mode_13h - Wednesday, August 8, 2018 - link
Yeah, but they could at least have gone for a V1000-series embedded Ryzen.https://www.amd.com/en/products/embedded-ryzen-v10...
psychobriggsy - Tuesday, August 7, 2018 - link
A 400mm^2 12nm (I presume, could be 14nm but there's little difference if it's GlobalFoundries) isn't cheap, there's a reason the price is over $600.I wonder if it's an exclusive design for this company, or not.
quadrivial - Monday, August 6, 2018 - link
If this is a new chip, doesn't it require new masks? If it does, aren't new mask sets around 80 million?Supposedly AMD was getting around 80% yields on Ryzen in 2017. This chip is 2x the size at around 400mm^2. If we assume the same yields, then that's around 110 chips per wafer. That comes out at around $42 per chip, but some probably can't hit that clockspeed, so they either have to be sold in other markets (what I expect) or are wasted. If new masks are needed, then there's a big additional cost too.
That said, that's a big die. 209mm^2 is enough for a 4-core ryzen module plus about half as many CUs (2400G). Even if half the die were dedicated to the GPU (die shots show it to be quite a bit less than that), we wouldn't expect more than an extra 100mm^2. I would have personally guessed at around 275mm^2 to simply extend the 2400G. Either they are spacing the transistors way out or there's something extra happening here.
IntelUser2000 - Tuesday, August 7, 2018 - link
Raven Ridge Ryzen has 11 CUs at its largest configuration. Based on early leaks this has 28, with 4 disabled. That's 2.5 times as much.RX 580 with 36 CUs have a 232mm2 die size. Intel Kabylake-G's graphics die with 24 CUs(but with more-than-normal ROPs) is over 200mm2 as well.
400mm2 is easily explained.
Intel999 - Tuesday, August 7, 2018 - link
I@quadrivialThey will not need to create new masks for any of this. They will just use a four core CCX (110mm2)just like any other APU. They got plenty of those laying around since all Ryzen and Epyc uses the same CCX. Based on the CU count on the Vega GPU you are looking at an equivalent sized chip to a RX 570 which comes in at 220mm2.
They produce these on the same die. They create APUs by connecting the CPU and GPU via Infinity Fabric at the test pack facility.
The 393mm2 estimate by Anandtech was just that, an estimate.
However, using GDDR5 may add some area due to it being a different memory type from a standard APU. This is just speculation, on my part.
However, the fact that the CPU and GPU parts needed are produced on every Vega and Ryzen wafer produced makes this incremental business for AMD. No added costs. Especially since semi-custom projects require the majority of R&D to be covered by the customer, not AMD.
silverblue - Tuesday, August 7, 2018 - link
AMD's Enterprise, Embedded, and Semi-Custom revenue was up in Q2. I wonder if this has anything to do with it, or is it solely EPYC?T1beriu - Tuesday, August 7, 2018 - link
>They will not need to create new masks for any of this.Each die has a new separate mask. You can't recycle masks because the die layout is different from Zeppelin to this.
>They create APUs by connecting the CPU and GPU via Infinity Fabric at the test pack facility.
Where have you read this junk? :D What what have you been smoking? :D
sgeocla - Tuesday, August 7, 2018 - link
> Each die has a separate mask.Not really. Zen cores and Vega CUs can have their own Layer Mask in a Multi-Layer Mask. There is no need to go for Full Maskset / Single Layer Maskset.
This is the preferred method for medium size production volumes.
http://anysilicon.com/understanding-maskset-type-m...
sgeocla - Tuesday, August 7, 2018 - link
Also, by going to https://www.euvlitho.com/2017/P33.pdf [EUV MASK TECHNOLOGY ANDECONOMICS: IMPACT OF MASK COSTS ON PATTERNING STRATEGY], on page 11/26 [DESIGN COST FOR SOC] we can see that:
1. The mask set cost is about 2.5% of the total cost
2. Design cost and embedded software cost are more than 90% of the total cost
AMD semicustom design reuses Zen cores and Vega dies in both design and embedded software so the total cost of this APU is an order of magnitude smaller of a full (re)design.
abufrejoval - Wednesday, August 8, 2018 - link
I have no idea how modular you can do masks, but as you say GDDR5 instead of HBM my not be minor and I wonder if they added or replaced it. Longer term I'd like to to see lots of APUs clustered into tight form factors with Ininity Fabrics of small to vast scale: Want more power? Add more cartridges, power and cooling!tipoo - Monday, August 6, 2018 - link
Interesting release for sure. When the 8th gen consoles came out, it would be years before a standalone APU could be bought at the performance of the PS4's GPU, and even recent ones lag in some ways. The way things are going now, APUs could be pretty close to what the PS5/XB? are looking like they're going to be, with Ryzen+Vega/Navi, and another year or two and die shrink down to add CUs and cores.Flunk - Monday, August 6, 2018 - link
It has nothing to do with what's possible. It has to do with what's profitable. There is absolutely nothing special about the PS4 or Xbox One's SoC. If AMD wanted to launch a nearly-identical APU they could.They don't for many reasons, but two big ones are that the CPUs on those chips aren't very powerful (and more powerful ones would take up valuable space) and AMD makes more money selling you a separate CPU and GPU than they would from an APU.
As for what the next Xbox and PS consoles will have on board, they're constantly working on new ideas for that, it won't be settled until they have a good business case to launch something. What they have now could be tossed out the window tomorrow.
mode_13h - Monday, August 6, 2018 - link
> 'introduce AAA-level monopoly and phenomenal level games at home and abroad'...gave me a good chuckle. Thanks for that.
> What exactly they mean by using government relations at this time is unknown.
Bonus!
Duncan Macdonald - Monday, August 6, 2018 - link
As this product is mainly a combination of a number of existing products, much of the masks will just be a cut-and-paste job from those products. This should greatly reduce the cost of producing the set of masks. If each of the functional blocks (CPU, GPU etc) has a tiny gap (1000nm would suffice) between it and the next functional block then there should be no optical interference between the masks of the different blocks hence no need to redesign them.jjj - Tuesday, August 7, 2018 - link
Have you asked AMD if Subor has exclusivity for the hardware and if that's in any way limited to China or consumer?fteoath64 - Tuesday, August 7, 2018 - link
One would assume China first and ROW later if the economics pan out. Unoless Xbox is licensed to this, there is little chance it will go anywhere. Maybe Linux/Steam could use this hardware into a gaming alternative to mid-range rigs ?.Or if big player like TenCent put their customised Android on it with their games and use Kinect-like gesture accessory to make a gamebox. Though China has strong regulations on the types of games permitted ....
DanNeely - Tuesday, August 7, 2018 - link
Doubtful we'll see this SoC in a future XBox or PS variant. The CPU is much faster than what either of them are using now, but the GPU would be a major downgrade from the top end XBox and more or less a sidegrade on the PS. OTOH I suspect that when both consoles have their next generation high end refreshes they'll probably jump to Zen cores of some flavor to give the CPU a major boost in addition to the larger and more capable GPU.eva02langley - Tuesday, August 7, 2018 - link
It is actually as strong as a PS4 Pro with 4 TFLOPS.piroroadkill - Tuesday, August 7, 2018 - link
Very interesting - is this the first generally available Windows machine with GDDR5 as system RAM instead of a flavour of regular DDR?msroadkill612 - Tuesday, August 7, 2018 - link
Indeed. This is seismic.It would be an extreme memory overclock afaik, for an am4 cpu to have 40GB/s+ memory bandwidth, and TR4 to have ~2x that.
An RX580 dgpu with similar specs to this console, has ~256GB/s bandwidth, more than 6x current AM4 system memory speeds.
A cpu using gddr memory, or even better, pooling it, is a major paradigm shift.
overseer - Tuesday, August 7, 2018 - link
An interesting product indeed (except the pricing). BTW as one of their content partners (game dev), I'll let them know about the interests from AT community and establish contact with Ian if they have bandwidth. :)edzieba - Tuesday, August 7, 2018 - link
GDDR for the CPU? That's going to have some interesting performance weirdness due to GDDR's increased access latency.eastcoast_pete - Tuesday, August 7, 2018 - link
Yes, might be, but how much more latency does GDDR5 have vs. DDR4? Also, I'd love to see what over four times the memory speed would do for the CPU.abufrejoval - Friday, August 10, 2018 - link
I've wondered about that and I hope someone in the know can share some data points here.On the other hand, do CPUs actually abandon cache line loads, if they notice code only asked for a single word and would now prefer another outside the current GDDR burst or cache line?
For code, I don't think it's much of an issue, that better be well inside the L1-3 caches anyway. Data may be all spread around, but mostly for old style code that's either fast enough or won't use multi-core anyway. With HPC or games much of the data can only benefit from bandwidth and people have taken great care to block things together for speedy access: They now get more benefit but no longer need to worry about copying things forth and back to the GPU.
And HSA workloads!! They become so much more attractive...
Really all we need is the ability to put lots of these nodes into really big NUMA over Infinity Fabric, I guess.
eastcoast_pete - Tuesday, August 7, 2018 - link
That console/PC combo would be a good value proposition for the US and Europe for about $ 600, and could even be the "steam machine that actually could" match or beat consoles, that is. I would be a possible customer, if the specs hold, the price is right, and a large player like Valve is behind it. Maybe Valve could even have another go at getting Steam OS going (Linux-based, if memory serves), and actually make it run games faster than WIN 10 this time. This setup has the oomph to do it.Alistair - Tuesday, August 7, 2018 - link
The obvious thing for Microsoft to do is add Windows 10 S mode to the next Xbox. Wipe the floor with this box. I wonder if they will or not.tipoo - Tuesday, August 21, 2018 - link
> AMD sites its commitment to technologies such as Rapid Packed MathCite maybe?
>The website for the system is now avaialble
;)
Balaji - Friday, March 29, 2019 - link
can anyone share me a purchase link for subor z+ ?ASAPfrenkie - Thursday, August 1, 2019 - link
Wow This is amazing. If I have this, I need not play pubg game on pc with nox: https://bignox.org