AMD Bristol/Stoney Ridge Thread

[amd6502]

The multithread advantage of BR does disappear when comparing to newer gen big cores; this is vs a 2200u which might be 15w (though it could be anything from 12W to 25W) against top binned 15w A12:

https://browser.geekbench.com/geekbench3/compare/8720276?baseline=8459093

If they can get the wattage under 10W for dual core, do you guys think that a low wattage RR-L that's 70% to 57% of the die area of RR (for 6CU and 3CU version, respectively) would be worth the cost of the project?

I never understood the Polaris 12 (8 CU at 101mm2) project vs Polaris 11 (16CU 123mm2) with the lesser chip being just 83% of the die area while half as capable.

 

[NTMBK]

Thanks, that's a good place to start. Though XV has 400 MHz clockspeed advantage there according to the benchmark. An i3-2130 would have fared better:

https://browser.geekbench.com/v4/cpu/11934784

Still sucks in MP though. For a broader perspective (note that Bench does not have the i3-2130 in its database):

https://www.anandtech.com/bench/product/1684?vs=1901

The Anandtech Bench comparison is where I was looking. Pretty much every single threaded benchmark was a loss for XV. And bear in mind that Carrizo was tested with DDR3-2400- I'm not able to track down the full test set up for the 2120, but I doubt it had such fast memory.

 

[DrMrLordX]

If they can get the wattage under 10W for dual core, do you guys think that a low wattage RR-L that's 70% to 57% of the die area of RR (for 6CU and 3CU version, respectively) would be worth the cost of the project?

If they want to go that direction, they'd be better off trying to build it on 7nm using Renoir as a base. 14nm/12nm may have lower design costs, but the baseline power usage for IF etc. may make it hard for them to fit under a 10w power envelope. They're already making Renoir on 7nm, and I think getting into the correct power envelope will make such a product workable. Plus die area would be significantly reduced.

The Anandtech Bench comparison is where I was looking. Pretty much every single threaded benchmark was a loss for XV. And bear in mind that Carrizo was tested with DDR3-2400- I'm not able to track down the full test set up for the 2120, but I doubt it had such fast memory.

It was probably DDR3-1600 CAS9 or something. If I had to guess. I could be wrong though.

I am not really sure an i3 is the correct thing to bench an XV against, though. At least not from the Sandy generation. Those are 2011 Intel CPUs against 2015/2016 AMD CPUs. One of the big advantages behind Sandy is that it clocks better than XV ever will (not that you can enjoy such an advantage with an i3). There isn't an XV made that will square off well against a 2500k in terms of pure performance.

 

[ET]

AMD said that Zen IPC is 52% higher than Excavator. Difference to Piledriver is more than that.

Okay, checked it out. From AMD's site (footnotes here):

SPECint_base2006 estimates: “Zen” vs. “Piledriver” (31.5 vs. 20.7 | +52%), “Zen” vs. “Excavator” (31.5 vs. 19.2 | +64%). Cinebench R15 1t scores: “Zen” vs. “Piledriver” (139 vs. 79 both at 3.4G | +76%), “Zen” vs. “Excavator” (160 vs. 97.5 both at 4.0G| +64%).

So, first of all, it's interesting that AMD used estimates for SPECint. Though the Cinebench results aren't estimates. My understanding is (from Anandtech) that "Piledriver" refers to FX CPUs, which have L3, which would account for the advantage in SPECint over Excavator.

Still, bottom line is Zen is 64% faster than Excavator in single threaded Cinebench R15. Would have been interesting to see how well Excavator would work with L3 cache.

So yeah, pretty bad. Pity we don't have Excavator with L3 to compare to, but this doesn't look promising.

 

[NTMBK]

It was probably DDR3-1600 CAS9 or something. If I had to guess. I could be wrong though.

I am not really sure an i3 is the correct thing to bench an XV against, though. At least not from the Sandy generation. Those are 2011 Intel CPUs against 2015/2016 AMD CPUs. One of the big advantages behind Sandy is that it clocks better than XV ever will (not that you can enjoy such an advantage with an i3). There isn't an XV made that will square off well against a 2500k in terms of pure performance.

I mean... the i3 is exactly the kind of part that Carrizo is intended to compete with. Ever since Llano, AMD has been positioning these quad cores with poor per-thread performance against Intel's dual-core hyperthreaded chips. (I mean if you want to compare it against an i3-6320 instead then go ahead, but the results aren't pretty for Excavator: https://www.anandtech.com/bench/product/1684?vs=1783 ) If we're not comparing it with an i3, then what should we compare it against? The Atom derived chips? Not exactly a resounding endorsement of the architecture.

 

[NTMBK]

So, first of all, it's interesting that AMD used estimates for SPECint. Though the Cinebench results aren't estimates.

I think that's due to some legal shenanigans, where unless the result is published on the SPEC website you are required to label it as an estimate: https://www.spec.org/fairuse.html#DefineEstimate

 

[DrMrLordX]

If we're not comparing it with an i3, then what should we compare it against?

Depends on the generation, but for anything as old as Sandy, I'd say the i5s.

 

[ET]

I never understood the Polaris 12 (8 CU at 101mm2) project vs Polaris 11 (16CU 123mm2) with the lesser chip being just 83% of the die area while half as capable.

It's my go-to example whenever people talk about what AMD should or shouldn't or is expected to do, talk about costs of creating a new chip, etc. It's one reason I'm willing to entertain the thought of AMD creating a CPU outside the mainstream Ryzen.

BTW, far as I remember, it's actually 10 CUs, but only 8 active.

I can think of a few reasons, none of them really great, for Polaris 12. I mainly think of it as a replacement for Oland. I think it should have ended a little smaller (2.2B vs. 3B transistors ~= 73% but 101mm2 vs 123mm2 ~= 82%). A good low end chip to shove at laptop OEMs once Oland outlives its welcome or AMD gets out of 28nm.

 

[Insert_Nickname]

BTW, far as I remember, it's actually 10 CUs, but only 8 active.

It is 10CUs (640:40:16). The rare 550X has all enabled. I don't think it's available for regular consumers though.

I can think of a few reasons, none of them really great, for Polaris 12. I mainly think of it as a replacement for Oland. I think it should have ended a little smaller (2.2B vs. 3B transistors ~= 73% but 101mm2 vs 123mm2 ~= 82%). A good low end chip to shove at laptop OEMs once Oland outlives its welcome or AMD gets out of 28nm.

Two reasons. One as an Oland replacement for OEMs. Two, more importantly, as a basic IGP-like chip for IGP-less Ryzen CPUs that don't need graphics horsepower.

 

[Shivansps]

They really need a replacement to hd6450 and 5450s.

 

[NTMBK]

They really need a replacement to hd6450 and 5450s.

It's called the IGP.

 

[Shivansps]

It's called the IGP.

unfortunately they are still selling them, in fact they got rebranded to R5 230 like 2 years ago. There is a market for it, just like the GT710.

 

[ET]

unfortunately they are still selling them, in fact they got rebranded to R5 230 like 2 years ago.

That was about 5 years ago. They are no longer being sold.

 

[Shivansps]

That was about 5 years ago. They are no longer being sold.

https://www.newegg.com/Product/Prod...Description=R5+230&ignorear=0&N=-1&isNodeId=1

I do agree it is more rare now, as the GT710 is just better option as a non-gaming GPU. It really depends on country at this point, here Gigabyte is dropping their R5 230 chaper than Gt710s.

 

[ET]

https://www.newegg.com/Product/Prod...Description=R5+230&ignorear=0&N=-1&isNodeId=1

I phrased my previous post badly. I meant 'they are no longer being produced'.

The sheer number of options might indicate that they are still being made, but I still doubt that. I can find GeForce 9800 being sold at Newegg, GeForce 8400, multiple versions of GeForce 210, Radeon 4350. They're marked as new, probably old stock. Same as CPUs. I bought an FX 8350 recently, new in box. Date on CPU is 2011, date on box is 2015.

 

[amd6502]

Well, when old 28nm GPU stock runs out they can consider using an APU with CPU disabled to make low end graphics cards. However, I think it must be a declining market, outside of the computer repair business. I'd like to see all Ryzen get a small iGPU in the 3000 and 4000 series.

So a PD regressed XV quadcore with small cache (1MB L2 only) run at under 10W would fall into atom territory. Same thing if they were to attach two jaguar cores run at half frequency to Stoney. The oversized front-end could be balanced that way if a jaguar core shares a decoder with a dozer core.

 

[NostaSeronx]

Since, there is GPU discussion. The 22FDX APU/GPU will most likely be skipping GFX9. We also should be getting small compute units versus standard compute units.

Standard Compute Unit;
64 Wavefront, 64KB LDS, 4x64KB VGPR, etc
^-- Remains premium, Stoney -> RV2

Small Compute Unit;
32 Wavefront, 32KB LDS, 4x16KB VGPR, etc
^-- Goes to this for higher power efficiency, Stoney -> 22FDX

Not only will the CPU be refined for clocks, so will the GPU.

 

[amd6502]

What's the advantage of smaller CU?

Could a CU be used for FPU calculations? They had a whole HSA concept 5+ years ago. I think they should get back to it now that Zen is so well along and can take the backburner a bit.

What would a 4 thread CCX look like in a Nosta APU?

 

[DrMrLordX]

Could a CU be used for FPU calculations? They had a whole HSA concept 5+ years ago. I think they should get back to it now that Zen is so well along and can take the backburner a bit.

They never stopped. Different software stack, but the kernel fusion driver lives on (Linux only I'm afraid).

What would a 4 thread CCX look like in a Nosta APU?

We'll never know, since nothing Nosta discusses actually exists, or will exist.

 

[hojnikb]

Well, when old 28nm GPU stock runs out they can consider using an APU with CPU disabled to make low end graphics cards. However, I think it must be a declining market, outside of the computer repair business. I'd like to see all Ryzen get a small iGPU in the 3000 and 4000 series.

So a PD regressed XV quadcore with small cache (1MB L2 only) run at under 10W would fall into atom territory. Same thing if they were to attach two jaguar cores run at half frequency to Stoney. The oversized front-end could be balanced that way if a jaguar core shares a decoder with a dozer core.

Thats highly unlikely. That kind of design would be too costly for a budget GPU due to VRM requirments and you'd need a chipset as well as far as i'm aware. Neat idea though.

 

[hojnikb]

They really need a replacement to hd6450 and 5450s.

Yes, that would be awsome. A cheap, 2-3CU GPU with as low power as possible and all the latest video decoding support (H265 10 BIT, VP9, perhaps AV1?).

It would work a treat with old PCs, that dont have codec support and only thing used is media consumption. Or perhaps if someone made a OpenCl based H265/VP9 decoder for older gpus (like 5450).

 

[Insert_Nickname]

Yes, that would be awsome. A cheap, 2-3CU GPU with as low power as possible and all the latest video decoding support (H265 10 BIT, VP9, perhaps AV1?).

It would work a treat with old PCs, that dont have codec support and only thing used is media consumption. Or perhaps if someone made a OpenCl based H265/VP9 decoder for older gpus (like 5450).

Best you can do right now is the GT1030/DDR4 version. 20W and has all modern codecs covered. It's not cheap though. The RX550 is the only other option, but lacks HW VP9 and is more expensive.

Then there is the whole 4K Netflix debacle. Where you need a 1050(3GB) or RX550. Don't even think about 4K UHD blurays. That's a minefield beyond belief.

 

[Shivansps]

Best you can do right now is the GT1030/DDR4 version. 20W and has all modern codecs covered. It's not cheap though. The RX550 is the only other option, but lacks HW VP9 and is more expensive.

Then there is the whole 4K Netflix debacle. Where you need a 1050(3GB) or RX550. Don't even think about 4K UHD blurays. That's a minefield beyond belief.

Imagine a 4gb ddr4 64 bit Vega 3 GPU... What would be that? 10w?

 

[Insert_Nickname]

Imagine a 4gb ddr4 64 bit Vega 3 GPU... What would be that? 10w?

Depends on frequency. In the Athlon, it consumes 10-15W. So you could most likely keep it below 10W with some tweaking. Polaris12 can already be contained in a 25W TDP. (in WX2100 guise)

It'd be a very small chip though. The GT1030 is already packing 1.8 billion transistors on a measly 70mm2 using 14nm process. I doubt it's that practical to go much smaller.

 

[ET]

From my point of view the right solution to this is to a mobile CPU hooked to USB. Think a media stick with USB on one end and HDMI on the other. I really don't understand the need for a weak, low power GPU card (I mean, beyond a certain point). I can understand the desire for a cheap card, but ...