AMD Raven Ridge 'Zen APU' Thread

[mohit9206]

If you just want a cheap gaming PC, sure, but if you want a laptop that is used for gaming among other general tasks, a Zen APU is going to be the best bet. The top model will probably offer performance comperable to a 460 which is going to be enough for people playing e-sport titles on their laptops.

Yes but i was talking about desktops.
Laptops, yes Zen apu is a great option.
Secondly no its barely going to be half as fast as rx460 on laptop.

 

[IntelUser2000]

I am also curious if the CPU will downclock if the iGPU is being used, like how this is done for the current APU's.

That's pretty much a given. Shared setups are slower than discrete parts even if everything else is equal, like TDP.

 

[PeterScott]

This will likely be true for the next few years.

Nevermind the fact that 7700HQ paired up with RX 550 beats R7 graphics in Carrizo by 4-5x. If we normalize for GPU we're still looking at 2-3x. Compare it to to RX 460 and we're back up to 3-4x.

1/2 the ROP
45% of memory bandwidth if it uses DDR4-3200, and it has to share it with the platform
TDP shared with CPU, meaning CPU won't be as performant

Totally agree. I was just looking at the one parameter he was focused on, but in totality, it will be many times slower than an RX 460/560.

But, hey it will likely beat my 8800GT I am still gaming on .

 

[ao_ika_red]

No, anyone who wants to play aaa games, will always buy a dedicated card unless they have no clue about pc market. $160 zen apu much worse than Pentium and 1050.

Tell that to my friend who still use 5800k apu to play witcher 3 with ancient 19" HD LG display. He loves to buy new games, but he won't spend a dime even for an SSD.

 

[whm1974]

Tell that to my friend who still use 5800k apu to play witcher 3 with ancient 19" HD LG display. He loves to buy new games, but he won't spend a dime even for an SSD.

A bit silly there considering that an SSD will make everything much faster and is well worth getting.

 

[ao_ika_red]

A bit silly there considering that an SSD will make everything much faster and is well worth getting.

There's a lot of people outside this forum who's not as picky as us here.

 

[whm1974]

There's a lot of people outside this forum who's not as picky as us here.

Yes I'm well aware of this.

 

[Mopetar]

Vega clocks higher when chowing down on 300 watts. This is a laptop part, and so far the only leaks show the GPU clock at 800 MHz (72.7% of RX 460). 78.5% * 72.7% = ~57% of RX 460.

That's also 64 CUs running at 1600 MHz. Running 11 CUs at 1600 MHz would be about 50 W from napkin math, so it's rather likely that the 800 MHz we've seen so far isn't the top speed or is an engineering sample. I expect 1300 MHz (if not more) for the desktop parts, and 1100 MHz doesn't seem unreasonable for the the best notebook part.

1/2 the ROP
45% of memory bandwidth if it uses DDR4-3200, and it has to share it with the platform
TDP shared with CPU, meaning CPU won't be as performant

An 11 CU APU should have 11 ROPs, so it's not quite as bad as you make it out to be. Memory bandwidth shouldn't be as bad either depending on how Vega improvements targeting memory bandwidth play out.

Obviously I don't expect the mobile APU with a 35W TDP to trounce a discrete graphics card, but it's not hard to imagine that the desktop part with a higher 65W TDP being competitive. Just doing napkin math with the numbers we have for Vega, an 11 CU part with a 1100 MHz clock, should have a ~33% FLOP advantage over a 550, so unless it hits a bottleneck due to memory or ROPs, it's going to have more throughput.

Another way to look at it is that if the APU were always going to hit a memory bottleneck, why not just use an 8 CU design instead since it would be cheaper and would only result in a disproportionately small performance hit? I think AMD went with a wider GPU (Initially I expected them to top out at an 8 CU design precisely because I didn't think they would use HBM and would have memory bandwidth limits) because they developed or improved technologies to address their memory bandwidth limitations and work around them as much as possible.

 

[Glo.]

Do we have any clue of how latency and power consumption compares between HBM2 and DDR4?

Power consumption: not more than 5W each HBM2 stack, and depends on voltage and frequency. 160 GB/s bandwidth can be achieved with 350 MHz(I think...), and will allow for sufficient enough bandwidth to feed the GPU.

 

[PeterScott]

That's also 64 CUs running at 1600 MHz. Running 11 CUs at 1600 MHz would be about 50 W from napkin math, so it's rather likely that the 800 MHz we've seen so far isn't the top speed or is an engineering sample. I expect 1300 MHz (if not more) for the desktop parts, and 1100 MHz doesn't seem unreasonable for the the best notebook part.

From what I can tell, even the RX 550 has more than triple the Memory bandwidth of Raven Ridge. So I expect RR to perform under the RX 550, and be clocked significantly lower, so they are likely bandwidth starved even running a lower speed.

I expect 800 MHz for the Notebook part and maybe 1000 MHz for the desktop part. Having more CUs running at a slower speed usually gives better perf/watt for GPU operations than running fewer CUs faster. Perf/Watt is something AMD desperately needs to compete with Intel for Notebooks, so that explains more CUs running at slower speeds, as we have seen in leaks so far.

 

[AtenRa]

Actually that should fit the bill on an APU. An A10-7870K is between a R7 240/250. Which is between 30-65 TDP.

I think some people here are underestimating low GPU's/APU's. Sure you can't play most of the latest games on 1080p/60fps, but 30fps with moderate settings is sufficient for most games and the people playing them. Which games are supposed to be unplayable on a A12-980/A10-7870?

At 1080p ?? many

As for the RR iGPU performance, with 40% higher than current APUs it should be ok to play most games at 1080p Low/Medium settings (desktop).
If they will keep the same prices (sub $150 for top SKU) and at 65W TDP or bellow, then they should be the perfect entry level Slim SFF APUs in 2018.

 

[PeterScott]

At 1080p ?? many

As for the RR iGPU performance, with 40% higher than current APUs it should be ok to play most games at 1080p Low/Medium settings (desktop).
If they will keep the same prices (sub $150 for top SKU) and at 65W TDP or bellow, then they should be the perfect entry level Slim SFF APUs in 2018.

Ryzen 5 4C/8T is already higher than that without an integrated GPU.
Ryzen 5 1500X 4/8 65 W $189

So pricing will obviously be MUCH higher, as this will now be a much more credible i5 competitor on both the GPU and CPU side.

There is room for cut down 2c/4t APU version that may occupy old price points, and the new 4C/8T high end at higher price points.

 

[insertcarehere]

Ryzen 5 4C/8T is already higher than that without an integrated GPU.
Ryzen 5 1500X 4/8 65 W $189

So pricing will obviously be MUCH higher, as this will now be a much more credible i5 competitor on both the GPU and CPU side.

There is room for cut down 2c/4t APU version that may occupy old price points, and the new 4C/8T high end at higher price points.

I can't see a $200+ APU gaining much traction, given stuff like Ryzen 3 + GTX1050/RX 460 (both of which are available in SFF) exist at that point and is a more powerful + flexible platform.

 

[AtenRa]

yea at 200$ its very close to a 6-Core R5.
I could see the top RR at $189 on launch date but not more.

 

[PeterScott]

yea at 200$ its very close to a 6-Core R5.
I could see the top RR at $189 on launch date but not more.

$189 is already what their top 4 core R5 without GPU costs.

You expect them to charge ZERO extra for adding a GPU?

That doesn't seem very realistic.

 

[AtenRa]

$189 is already what their top 4 core R5 without GPU costs.

You expect them to charge ZERO extra for adding a GPU?

That doesn't seem very realistic.

RR will go against mainstream Core i3/i5.
In 2018 Intel will have a 6-Core + iGPU mainstream Core i7 at $300-350. That means Core i5 could be a 6-Core (no HT) + iGPU at $230-240 with entry level 6C 6T Core i5 at $180-190

So how exactly AMD will release a 4C 8T + iGPU at more than $189 if they want to be competitive ??

 

[PeterScott]

I expect you will be disappointed by Intel pricing as well.

 

[Glo.]

$189 is already what their top 4 core R5 without GPU costs.

You expect them to charge ZERO extra for adding a GPU?

That doesn't seem very realistic.

189 mm2 Die size can cost to manufacture AMD as low as 45$ with packaging, testing and validation.

APU is monolithic design, with single CCX, and GPU added. The die size is what matters judging the manufacturing costs. Even if Raven Ridge is bigger(up to 200 mm2) we are looking at manufacturing costs increased by a margin of 1$. Cheapest Ryzen CPUs using Zeppelin die will cost 109$.

It depends what AMD is going to target.

IMO HBM2 design can lock in 55-60$ price range, considering die size, and complexity of manufacturing.

AMD is working on those designs, altho not announced yet, which has been said by Don Woligroski in TomsHardware AMD, the question is whether we ever will see them out in mainstream offerings. Most likely yes, the possibilities are too huge to avoid them.

Its cheaper for AMD to make single solution for 85% of market targeting the cheapest hardware, that can offer 350-400$ hardware level of performance(GPU+CPU).

APU without HBM2 on package has little benefit to the client, and gives no business possibilities. HBM2 design can sell for 299$, and give AMD a lot of margin. How come. Because if the performance is right - they suck up Intel+Nvidia, AMD+Nvidia potential customers in this market.

 

[AtenRa]

I expect you will be disappointed by Intel pricing as well.

We already know there will be a 6+2 CoffeLake to replace 4+2 Core i7 7700K at $300-350 price point.

 

[PeterScott]

We already know there will be a 6+2 CoffeLake to replace 4+2 Core i7 7700K at $300-350 price point.

No we don't. We know there will be a 6 Core part. Pricing is unknown.

I expect Intel to charge a premium for 6 core and still have the bulk of their sales in 2 core and 4 core parts, as such they won't want to lower the pricing of 2 core and 4 core parts.

 

[Mopetar]

From what I can tell, even the RX 550 has more than triple the Memory bandwidth of Raven Ridge. So I expect RR to perform under the RX 550, and be clocked significantly lower, so they are likely bandwidth starved even running a lower speed.

I expect 800 MHz for the Notebook part and maybe 1000 MHz for the desktop part. Having more CUs running at a slower speed usually gives better perf/watt for GPU operations than running fewer CUs faster. Perf/Watt is something AMD desperately needs to compete with Intel for Notebooks, so that explains more CUs running at slower speeds, as we have seen in leaks so far.

AMD is currently selling Bristol Ridge APUs that clock at 900/1100 MHz respectively in their mobile and desktop configurations. Sure they have parts with lower clock speeds to hit lower TDP targets, but moving to 14 nm and an architecture designed to achieve higher clock speeds is going to allow far more than 800 and 1000 MHz. Historically, AMD has offered desktop APUs with 65W TDP, and in the past they even had some at 95W TDP.

The desktop parts will allow for higher clocks, and if you use faster memory, the total bandwidth falls closer to half of what's available for a 550. If Vega allows for more effective use of that bandwidth, then the performance difference is further mitigated. If you're running a game that isn't bound by available memory bandwidth, then it won't matter at all. Are the type of games (e-sport titles) that are typically targeted by something like the 460/550 all that memory heavy or bandwidth sensitive?

 

[Glo.]

Let me give you some example of potential in Raven Ridge.

My company has made an interesting experiment lately. We have bought 15 inch Touch Bar Macbook Pro, with Radeon Pro 555. Then we have built a lop power, high performance computer targeted for 1080p.

It has Core i7 7700T, 16 GB 2400 MHz RAM, and GTX 1050 Ti, to make comparable performance of both Computers. It may not look like they were comparable at first glance, but when you will go to end of this post, you will be surprised.

Both GPUs, Polaris and Pascal have exactly the same core count, and have the same 128 Bit memory bus, with GDDR5 memory.

What we wanted to look at is how both architectures are comparing clock, for clock in games.

So we decided to buy Palit KalmX GTX 1050 Ti, and declock it to the lowest possible value on core, and lowest possible value on Memory.

The end result was that the GPU was running at 875 MHz, and 6000 MHz on memory. Standard clocks: 1392/7000 Mhz.

Radeon Pro 555 has 768 GCN cores, 855 MHz core clock, and 5000 MHz memory clocks.

Test was fairly simple. We have run in Windows 10 with latest drivers available Overwatch in 1080p Epic settings, and we tested it by playing 5 runs in Quick Match.

Radeon Pro 555 averaged 32 FPS in those settings, over those runs.
GTX 1050 Ti, declocked to the mentioned values: 34 FPS.

Thoughts? Overwatch is heavily Nvidia optimized game, and the GPU has higher memory bandwidth which really affects Nvidia GPUs in Blizzard games.

Per clock both GPUs are comparable. Both have 768 Cores.

What this means for Raven Ridge.

If Raven Ridge with HBM2 has 12 CU's the only thing that will define its performance compared to GTX 1050 ti is its core clock. 160 GB/s from single stack of HBM2 will be enough to feed those cores, and will be very easy to cool and maintain stable. I can very easily see that 65W TDP APU with 4C/8T and 2 GB/s of HBM2, made on 14 nm FF+ has 1200 MHz core clock on GPU.

Per clock Polaris is on the same level as Pascal. There is a lot actually to root for AMD to deliver with Vega Raven Ridge APUs. So 768 GCN core APU, with sufficient enough bandwidth will actually be clock for clock on the same level as GTX 1050 Ti, if Vega does not bring any changes to IPC of the GPU.

7th Gen APU: 9800 65W TDP, 1108 MHz, 512 GCN cores.
Raven Ridge APU mentioned: 35W TDP, 800 MHz, 704 GCN cores.

Raven Ridge actually has lower TFLOPs number, but is still 40% faster. This slide has not been made by AMD GPU marketing team, but by CPU marketing team. On CPU they were able to deliver. Will they be able to deliver also GPU?


P.S. Imagine that AMD develops two SKUs for HBM2 RR APUs. 65W and 95W. And 65W with HBM2 costs 199$, and offers GTX 1050 Ti level of performance, with 4C/8T CPU.

Possibilities for SFF, HTPC market are endless.

 

[PeterScott]

The desktop parts will allow for higher clocks, and if you use faster memory, the total bandwidth falls closer to half of what's available for a 550. If Vega allows for more effective use of that bandwidth, then the performance difference is further mitigated. If you're running a game that isn't bound by available memory bandwidth, then it won't matter at all. Are the type of games (e-sport titles) that are typically targeted by something like the 460/550 all that memory heavy or bandwidth sensitive?

Every game that requires a GPU is going to have some memory BW related limitations. Even running the fastest available DDR4 you still won't have half the BW of RX 550. It really won't challenge an RX 550, let alone the 460/560. AMD themselves are only calling for an up to 40% GPU performance increase.

It's a better APU, not a replacement for current discrete GPUs.

 

[insertcarehere]

Per clock Polaris is on the same level as Pascal. There is a lot actually to root for AMD to deliver with Vega Raven Ridge APUs. So 768 GCN core APU, with sufficient enough bandwidth will actually be clock for clock on the same level as GTX 1050 Ti, if Vega does not bring any changes to IPC of the GPU.

I dont like to be a sourpuss here, but if you've visited the gpu forums a bit over the last few weeks, whatever information we're getting out of the vega architecture, matching or exceeding Polaris IPC doesn't seem to be it, disregarding that Pascal seems to be able to run higher clocks than Polaris/Vega at the same power consumption to begin with..

 

[Thunder 57]

189 mm2 Die size can cost to manufacture AMD as low as 45$ with packaging, testing and validation.

APU is monolithic design, with single CCX, and GPU added. The die size is what matters judging the manufacturing costs. Even if Raven Ridge is bigger(up to 200 mm2) we are looking at manufacturing costs increased by a margin of 1$. Cheapest Ryzen CPUs using Zeppelin die will cost 109$.

It depends what AMD is going to target.

IMO HBM2 design can lock in 55-60$ price range, considering die size, and complexity of manufacturing.

AMD is working on those designs, altho not announced yet, which has been said by Don Woligroski in TomsHardware AMD, the question is whether we ever will see them out in mainstream offerings. Most likely yes, the possibilities are too huge to avoid them.

Its cheaper for AMD to make single solution for 85% of market targeting the cheapest hardware, that can offer 350-400$ hardware level of performance(GPU+CPU).

APU without HBM2 on package has little benefit to the client, and gives no business possibilities. HBM2 design can sell for 299$, and give AMD a lot of margin. How come. Because if the performance is right - they suck up Intel+Nvidia, AMD+Nvidia potential customers in this market.

I don't think many expect to see HBM with Raven Ridge. We may see some niche solutions but I wouldn't count on that either. I wouldn't expect to see HBM come into play until whatever comes after RR. What would seem more realistic is pulling an Intel and equipping RR with some embedded DRAM for the premium models.

APU without HBM2 on package has little benefit to the client, and gives no business possibilities.

I strongly disagree with this statement. The larger problem for most people right now is that APU's are still based on the Construction Cores. We have seen that Zen is very power efficient and still delivers great performance. Having a Zen based APU that eliminates the need for discrete graphics in most use cases should be very popular.