Next Generation Hardware Speculation with a Technical Spin [2018]

[McHuj]

They probably didn't actually read the articles they referenced. This all comes from near the PS4 Pro launch timeframe where, Cerny said you'd need around 8 Teraflops for native 4K rendering of current of current gen games.

The rough google translate quote: "Personal calculations would be necessary if rendering natively at 4K, but at least 8 TFLOPS would be necessary," says Sirney." from a 2 year old article.

https://av.watch.impress.co.jp/docs/series/rt/1026717.html

 

[MrFox]

''Mark Cerny, PS4’s lead architect, told Digital Foundry last year that the realistic limits for a next-gen console GPU would most likely top out at eight teraflops''

From that link.

Gamesradar can't read good

 

[chris1515]

He said 4k for all PS4 games need at least 8 Tflops. (He did not said it but it needs in theory 4 times the bandwith), after it depends where the bottleneck of a game is (ALU, Bandiwith,fillrate...)

 

[MrFox]

The context was about requiring new techniques to reach 4k efficiently because without anything different a brute force 4k of existing games would need 4 times more of everything in hardware. It was hypothetical, it was never about a real hardware. There were similar statements about 120fps on psvr needing more effort with temporal reconstruction instead of just using existing engines.

 

[vipa899]

They indeed cant read good. On the other hand it probally depends on what kind of flops, if its 8TF of turing it would be 2070 performance which isnt unrealistic.
Even if it aint 1080 performance its quit a jump. I wonder how far amd have come on the gpu front. Navi desktop should be there 2019 so we probally get an indication then.
As for zen 2 we know that ryzen is a huge step up from fx series, with zen 2 being slightly faster it wont be a problem even at lower clocks.
Intel/nvidia are much ahead but not in the price/performance ratio.

 

[iroboto]

hand it probally depends on what kind of flops, if its 8TF of turing it would be 2070 performance which isnt unrealistic.
Even if it aint 1080 performance its quit a jump. I wonder how far amd have come on the gpu front. Navi desktop should be there 2019 so we probally get an indication then.
As for zen 2 we know that ryzen is a huge step up from fx series, with zen

Architecture matters slightly less on console, since they can voptimize for architecture.

 

[vipa899]

Architecture matters slightly less on console, since they can voptimize for architecture.

Yeah slightly less, but i assume that architectures play a big role for consoles too, Navi will be better per TF i assume. And doesnt that go both ways, if they can optimize for turing or navi, then they would get more power out of the same amount of flops from a 2012 GPU with as many TF?

 

[iroboto]

Yeah slightly less, but i assume that architectures play a big role for consoles too, Navi will be better per TF i assume. And doesnt that go both ways, if they can optimize for turing or navi, then they would get more power out of the same amount of flops from a 2012 GPU with as many TF?

FLOPs are a unit of measure of maximum theoretical throughput on the compute side of things. There are all sorts of other bottlenecks that could impede the output of a GPU, but not necessarily be related to FLOPs for instance.

And thus, if you can optimize your title around that bottleneck, you can use the FLOPs to the fullest.
So, the answer should be, it depends on what you're trying to do.
Nvidia FLOPS aren't any faster than AMD FLOPs, a FLOP is a FLOP. But if the bottleneck is Geometry or Triangle throughput then this would be problematic. It's easier for people to simplify by saying Nvidia FLOPS are faster than AMD FLOPS, but that really just boils down to bottlenecks on the architecture or card that is not being well addressed.

 

[3dilettante]

@3dilettante

How much faster is a FX8350 over the jaguar core found in the base ps4 anyway, how much were talking about?

This is tricky to say given how long ago Jaguar and Vishera were considered relevant, and the lack of head to head testing for them. Jaguar on its own de-emphasized performance, so AMD typically didn't try to benchmark it competitively. The consoles make it more difficult since they change much of the platform and neither platform has that kind of public benchmarking available.

Some examples of SpecInt estimates for the Jaguar-based Opteron X1000 put a 2GHz quad core at around 10 for single-threaded integer and 28.9 multithreaded (https://hexus.net/tech/items/cpu/55...performance-small-core-x86-server-processors/).
The SpecCPU2006 results for the closest FX chip I could search was the FX-8150, which is 10% or so slower generally than the FX-8350. Single-core integer is between 20 and 22, multithreaded is 106-115.

There are uncertainties on how representative these would be for the specifics of modern games on a console, but integer performance seems to be in the ballpark of 2.5-3x better for single-threaded integer work, and similarly better on multithreaded.
Doubling the throughput score for the 8-core console would bring the console chip to about half of the FX, but this is not correcting for clock speed or core reservation effects.

FP is harder to find tests for that far back. Geekbench has more modern results, although the specifics of its test patterns and how they'd compare to the evaluations done years ago are unclear.

The Jaguar-based A4-5100 https://browser.geekbench.com/v4/cpu/10981167
vs.
FX-8350 https://browser.geekbench.com/v4/cpu/10986666

Integer 964 vs 3003 single, 2259 vs 15693 multi.
FP 709 vs 2031, 2393 vs 10708.

Multi-threaded, if similarly naively doubled for the PS4 gives a worse deficit for integer than SpecInt. FP shows a 2x advantage for the FX over a 2x Jaguar.

That said, this is rating a lower-end netbook SOC at 15W versus a higher end desktop 125W chip with significantly more silicon.
Bringing the core count and clocks down to equalize would make Jaguar look better--aside from single-threaded. Steamroller would have done more to improve per-clock performance and power efficiency while keeping a very large single-threaded advantage, which is likely why it was considered for a time.

 

[vipa899]

Ok wow well thanks both of you, i understand now

 

[Garrett Weaving]

Wouldn't it make more sense to compare with one of those AM1 Athlons instead? Needless to say, since there isn't an 8core Jaguar part for desktops, there's not much you can do to approximate the performance.

 

[3dilettante]

Wouldn't it make more sense to compare with one of those AM1 Athlons instead? Needless to say, since there isn't an 8core Jaguar part for desktops, there's not much you can do to approximate the performance.

I haven't kept track of all the variants for Jaguar or their inconsistent benchmarking.
There are Geekbench scores for Jaguar-based Athlons like the 5150.
https://browser.geekbench.com/v4/cpu/9645913
The differences in FP seem to be in line with it being incrementally faster than the previously mentioned Jaguar score. Integer single-core seems to be generally in line. The integer multicore has a notable increase to 3224, which is notably higher than expected from a clock bump or general variability. Whether this is somehow tied to the better memory scores or higher TDP granted to the 5150 (25W), or some other factors related to the less formal and more varied submission base of Geekbench is unclear.
This would nudge Jaguar up in the integer multithreaded score closer to its deficit in SpecInt (to a somewhat lower IPC architecture than Piledriver, which is slower than Steamroller), ~2.25 if naively doubled. The scaling would not be that perfect, but it serves to give a general estimate of the high-end. The higher power rating may impact whether a console would see the same benefit, depending on what the console has budgeted for CPU. The dominant consumer would be the GPU, and given the 120-140W range that had been measured for the launch consoles a 2x15W versus 2x25W basis would be a measurable difference.

 

[see colon]

Here's geekbench for the FX 8310. 3.4GHz and 8 cores, so just over twice the clock and twice the cores. Also, it's a 95w part.
https://browser.geekbench.com/v4/cpu/10490694

 

[McHuj]

Samsung just announced their new QVO SSDs. 1TB for $149, 2TB for $299, and 4TB for $599. It wouldn’t surprise me if by 2020, you could get a TB for half that price at retail.

I think the chances are increasing that next gen will come with a SSD as the costs are getting reasonable and will scale down better than hard dives.

That’s one of my next gen wishes: a default 1TB SSD drive.

 

[goonergaz]

Samsung just announced their new QVO SSDs. 1TB for $149, 2TB for $299, and 4TB for $599. It wouldn’t surprise me if by 2020, you could get a TB for half that price at retail.

I think the chances are increasing that next gen will come with a SSD as the costs are getting reasonable and will scale down better than hard dives.

That’s one of my next gen wishes: a default 1TB SSD drive.

I appreciate the optimism, but what is that based on? I mean, historically has a similar product shown such a drop? Also, how much did they spend at this launch on the HDD and how would that compare and impact the launch price?

 

[Shifty Geezer]

Flash prices will reduce better than HDD because there are no mechnical limits of the same sort capping minimum price. HDDs can't drop below some $20+ costs because of all the parts whereas flash is as cheap as the silicon needed to make it. This is why you can get flash drives for a few bucks while HDDs have never been that cheap and never will be. eg. I can get a 32 GB USB drive for £8. Once upon a time, a 32 GB HDD cost hundreds. Then a 32 GB HDD cost £30. But a 32 GB HDD has never been £8 because it can't get lower than the mininum cost, and instead the minimum storage capacity for that bottom $30 price has increased.

 

[goonergaz]

Flash prices will reduce better than HDD because there are no mechnical limits of the same sort capping minimum price. HDDs can't drop below some $20+ costs because of all the parts whereas flash is as cheap as the silicon needed to make it. This is why you can get flash drives for a few bucks while HDDs have never been that cheap and never will be. eg. I can get a 32 GB USB drive for £8. Once upon a time, a 32 GB HDD cost hundreds. Then a 32 GB HDD cost £30. But a 32 GB HDD has never been £8 because it can't get lower than the mininum cost, and instead the minimum storage capacity for that bottom $30 price has increased.

Thanks for the explanation, it's just from my own experience I purchased a 1TB SSD from CEX and 2 years later the same drive is selling for 30% less...so a 2 year older product devalued 30% over 2 years vs the expectation of above that a new product will be 50% cheaper 2 years later. And even then there's the consideration of what that impact on increased BoM will have on retail...likely a ~£50+ premium?

I'm more than happy to pay it, and personally I would want SSD...just a thought on how realistic the wish was...

 

[Kaotik]

I appreciate the optimism, but what is that based on? I mean, historically has a similar product shown such a drop? Also, how much did they spend at this launch on the HDD and how would that compare and impact the launch price?

Samsung 860 EVO 1TB launched early this year, it was priced initially (in Finland) 329-339€ depending on shop, it's now it's down to ~200€ in normal pricing (BF prices were around 140€)
Going down over 1/3rd of the price in a year should allow -50% in 2020

 

[Shifty Geezer]

Thanks for the explanation, it's just from my own experience I purchased a 1TB SSD from CEX and 2 years later the same drive is selling for 30% less...so a 2 year older product devalued 30% over 2 years vs the expectation of above that a new product will be 50% cheaper 2 years later.?

As I understand it, there's an expectation that prices are going to drop faster than they have because price fixing has been addressed. A launch console with a $50 SSD versus a $30 HDD is also viable as a loss-leader as that same flash will drop to $25 and then $15 within a couple of years, possibly. It's also smaller and lighter than a drive, allowing for a smaller console with additional savings through the retail chain per unit.

 

[Scott_Arm]

As I understand it, there's an expectation that prices are going to drop faster than they have because price fixing has been addressed. A launch console with a $50 SSD versus a $30 HDD is also viable as a loss-leader as that same flash will drop to $25 and then $15 within a couple of years, possibly. It's also smaller and lighter than a drive, allowing for a smaller console with additional savings through the retail chain per unit.

SSDs should also be a lot more durable, so they'd probably have to do somewhat less repairs.