Showing results for 
Search instead for 
Did you mean: 

General Discussions

AMD Raven Ridge APUs benchmarked: HALF as fast as Vega GL inside Intel Gen8 processors.

Maybe this is why AMD's stock is down another 4% today...

AMD Raven Ridge Ryzen APUs With Vega Graphics Performance Tested | WCCFTech


1 Reply

These are showcasing performance pretty much where they should be., actually what this showcases more strongly is that the APU Graphics (Desktop Versions) are actually very competitive with Discreet Solutions.

Keep in mind that the RX Vega 11(CU) is operating at 1250MHz (1.76TF) compared to the RX Vega M GL (20CU) operating at 1010MHz (2.59TF)

Now as an experiment (given no one bloody uses 3D Mark 11 for Benchmarking given it was deprecated in favour of 3D Mark Firestrike / Timespy) I did some runs with my RX 480 and R5 1600 (Stock 3.2 / 3.5GHz)

At Stock (1300MHz / 5.99TF) 18,836 / Performance (720p) Graphics Score • 15,126 / 3D Mark Score • 9,962 / Physics Score

At MGL (565MHz / 2.60TF) 8,138 / Performance (720p) Graphics Score • 8,167 / 3D Mark Score • 9,798 / Physics Score

At RXV11 (385MHz / 1.77TF) 5,840 / Performance (720p) Graphics Score • 5,955 / 3D Mark Score • 9,720 / Physics Score

Now why the Under Clocked RX 480 was 16% Faster than the APU RX Graphics., which as a note IS NOT caused by the Memory Frequency / Bandwidth... as I did run tests with the Memory set to 1250MHz (5000MHz Effective)... while this isn't 1:1 with the 2400MHz (4800MHz Effective)... it's close enough that we can eliminate this as a potential performance bottleneck. As the results were within Margin of Error., compared to 1750MHz (7000MHz Effective) Stock.

Instead we can explain the performance differential here (5000 Vs. 5800 / +16%) being caused by the Shared Instruction Pipeline., specifically speaking while this is greatly improved on Zen over Carrizo., chiefly we're talking about a 6 Instruction Pipeline with 2 Operation Per Clock... Vs. 4 Instruction Pipeline with 2 Operation Per Clock... this means instead of 50% we're looking at 33%., this is as a note further Halved due to SMT (16.5%).

As such the RX Graphics in the APU are *EXACTLY* where we'd expect them to be in terms of performance compared to Discreet Solutions.. although we're not actually seeing any real benefit over Polaris beyond Lower Power Draw (i.e. better Efficiency). Mind given the Test we're running is using DirectX 11.0, this actually isn't surprising as for all intended purposes both GCN 4.0 and 5.0 should showcased *identical* performance to GCN 3.0 (R7 Graphics) in the previous APU Clock-for-Clock. When we account for the CPU Bottleneck,. then yeah this is exactly what we actually see.

Where we run into peculiarity is the RX Vega M., as it simply SHOULD NOT be seeing the Performance it is.

Now this said there are 2 Versions... a 20CU (GL) and 24CU (GH) Version., which again the Article simply doesn't reference which they're testing... this is important as well because the 24CU Version Boosts to 1100MHz., which means instead of 2.59TF we're talking 3.38TF; this isn't a "Minor" Performance Increase.

Running the Test again at 733MHz / 3.38TF we see the results 10,478 / Performance (720p) Graphics • 10,198 / 3D Mark • 9,845 / Physics... which is entirely inline with what we're seeing in terms of the RX Vega M GL.

Reference Benchmarks:

AMD Radeon RX 480 video card benchmark result - AMD Ryzen 5 1600,ASUSTeK COMPUTER INC. PRIME B350-PL... (Stock)

AMD Radeon RX 480 video card benchmark result - AMD Ryzen 5 1600,ASUSTeK COMPUTER INC. PRIME B350-PL... (RX Vega M GL / 20CU)

AMD Radeon RX 480 video card benchmark result - AMD Ryzen 5 1600,ASUSTeK COMPUTER INC. PRIME B350-PL... (RX Vega 11 / 11 CU)

AMD Radeon RX 480 video card benchmark result - AMD Ryzen 5 1600,ASUSTeK COMPUTER INC. PRIME B350-PL... (RX Vega M GH / 24CU)

While you might think this isn't an ideal approach., the reality is that comparing the results between an Under Clocked RX 480 with R7 360 and R9 280 (simply accounting for the Theoretical TF) actually does provide reasonably identical results on Benchmarks like 3D Mark. If you have older Radeon Hardware I strongly recommend you try it for yourself to compare and contrast.

This actually allows for a good High-Level Estimation in Performance Requirements, even for APU Graphics when we account for the CPU Bottlenecks.

As noted above the Memory Frequency is less important than you might think... unless you're starving the GPU Data., which frankly even at 1300MHz you're only going to need 1600MHz DDR3 before Memory becomes the Bottleneck on an 8CU GPU.

The bigger performance bottleneck will always be the Shared Instruction Cache, as the two components aren't operating independently... although this is resolved with HBCC., something I do hope Zen 2 incorporates assuming memory prices calm down.

Something else to also keep in mind is the cost differential here.

Sure the R5 2400G is Gaming Wise ~50% the Performance of the i7 8709GH but ... that said we know the price of the R5 2400G is $170., but we don't know how much these CPU + Discreet will be.

Keep in mind they pack HBM2, plus will require a Custom Motherboard Socket (and Chipset).

Now if I were Intel I'd be looking at given it's essentially an i3 8450K ($180) + GTX 1050 Ti ($150 / MSRP) then it /should/ be about $350 ... but I'd wager it's more likely to be $410 - $500... even still at the lower price point that would still just break even with the R5 2400G in terms of Price/Performance.

This said something I'd suggest might be interesting would be Threadripper 2-Series CPU w/RX Vega Mobile.

Given it has 2 Dummy CCX., it's entirely possible that these could be replaced with RX Vega Mobile GPU providing up to 48CU with 8/12/16C and 16/24/32T.

That would make an excellent Content Creation / Developer Platform, so say $1,000 ... as you're effectively getting a High Performance Workstation CPU with an RX Vega 48., performance being somewhere between RX 580 and RX Vega 56.

Yet because there would be 2 GPU, would make it perfect for VR Content Creation (like the Pro Duo is)... plus if it has small Local HBCC (48-96MB) this would keep the memory costs down utilising the Shared System Memory which can provide substantially more Memory than would really be financially reasonable on a Dedicated GPU.

Bare in mind that RX Vega Mobile isn't Intel Specific., rather it's just an Embedded SoC that AMD Produce.

With this said Stock prices always rise and fall. It's best not to pay much attention to the Week-by-Week swings., as these are driven by idiots who panic easily with news like this that they barely understand.

Keep in mind as well that Spectre and Meltdown are still somewhat dominating., so news about that is more likely responsible to current swings you see.

Still again, let idiots be idiots and panic sell. Just means you can invest more without the premium stock price., then when it recovers (which it will because panic selling occurs due to Rumours and Speculation not Facts) then everyone will be in a stronger position. You as the Investor and AMD for their Investment Reserves., which are currently at a healthy $1B ... and not something they're dipping into as they're actually operating with a profit this year and expecting to see doubled growth over last year (which they saw double their predicted growth... so if anything the recent news should be driving a MUCH stronger Stock Price, especially with the news that AMD are guaranteeing to patch Spectre and Spectre-Like vulnerabilities in Zen 2... due out next year) we don't have a similar guarantee from Intel yet ... which could drive Sector growth towards AMD until Intel do announce such because Business' will see to shift their Ecosystem from Intel to AMD.

The price at present already makes it enticing, but the guarantee of a maintained compatible ecosystem and removed vulnerabilities; which are already lower, well that's not just enticing but essentially going to mean business will be currently pricing up and planning their upgrade cycle around AMD (almost certainly for Q3/Q4 2018).