AMD Radeon Pro 5600M with HBM2 Benchmarked

[btarunr]

Benchmarks of the new Apple-exclusive AMD Radeon Pro 5600M graphics solution by Max Tech reveals that the new GPU is about 50% faster than the Radeon Pro 5500M, and within striking distance of the Radeon Pro Vega 48 found in Apple's 5K iMacs. The Pro 5600M is an Apple-exclusive solution by AMD, based on the "Navi 12" silicon that features a 7 nm GPU die based on the RDNA graphics architecture, flanked by two 4 GB HBM2 memory stacks over a 2048-bit interface. The GPU die features 2,560 stream processors, but clocked differently from Radeon Pro discrete graphics cards based on the "Navi 10" ASIC that uses conventional GDDR6.

The Radeon Pro 5600M solution was found to be 50.1 percent faster than the Radeon Pro 5500M in Geekbench 5 Metal (another Apple-exclusive SKU found in 16-inch MacBook Pros), and just 12.9 percent behind the Radeon Vega 48. The Vega 56 found in iMac Pro is still ahead. Unigine Heaven sees the Pro 5600M being 48.1% faster than the Pro 5500M, and interestingly, faster than Vega 48 by 11.3%. With 2,560 RDNA stream processors, you'd expect more performance, but this card was designed to meet stringent power limits of 50 W, and has significantly lower clock-speeds than "Navi 10" based Radeon Pro graphics cards (1035 MHz max boost engine clock vs. 1930 MHz and 205 W TDP of the Pro W5700). Find more interesting commentary in the Max Tech video presentation.



View at TechPowerUp Main Site

 

[cucker tarlson]

you'd expect more from a ful 5700xt die with hbm2 than to match vega 56

 

[mtcn77]

you'd expect more from a ful 5700xt die with hbm2 than to match vega 56

50 vs. 150w tdp.

 

[Vayra86]

Its funny how Mac makes a mid range card look great.

 

[mtcn77]

Its funny how Mac makes a mid range card look great.

It is, the perfect version. 50w is the killer deal.

 

[cucker tarlson]

50 vs. 150w tdp.

one would think the editor should mention that

 

[mtcn77]

one would think the editor should mention that

Funny enough chied editor wizzard has entered that in the gpu database - that is how I found out.
PS: I'm something of an editor myself(previously).

 

[btarunr]

one would think the editor should mention that

Done.

Half the clocks and 1/4th the power budget of the Pro W5700. It's a miracle that this card is doing as well as it is.

 

[cucker tarlson]

Funny enough chied editor wizzard has entered that in the gpu database - that is how I found out.
PS: I'm something of an editor myself(previously).

W1zzard is the man

 

[mtcn77]

W1zzard is the man

He is the lord gaben.

 

[Chrispy_]

Sadly all of the other results are hampered by a meagre 50W power budget too, so there's no comparable reference point to a regular 5700XT running on Windows.

All I can see is that a 40CU part is about 50% quicker than the 24CU part , so we know that it's bottlenecked by something other than CUs; Probably power, because it's unlikely to be the HBM2.

With no details on how the test is run or under what conditions, these results are just noise.

 

[InVasMani]

I'm certainly no Apple fanboy, but this type card is well suited for them ironically given their such a closed ecosystem so the quality of support for this type of GPU design is likely to be a bit higher than it might be otherwise.

 

[Valantar]

I am seriously looking forward to seeing this GPU in a broader review including a Windows install and some gaming benchmarks comparing it to laptops with the GTX 1650 (max-p, 50W), 1660 Ti max-Q and 2060 max-Q (both 60W IIRC). That power efficiency seems utterly mind-boggling. It definitely comes at a cost (HBM2 and a bespoke die), but man... this is seriously impressive.

 

[mtcn77]

All I can see is that a 40CU part is about 50% quicker than the 24CU part ,

That is not bad.

so we know that it's bottlenecked by something other than CUs; Probably power, because it's unlikely to be the HBM2.

HBM2 has benefits outside of performance. The signal intervals are discrete from one another, so while buffered writes might keep the optimal ratios in gddr6, it enables far greater control over immediate mode accesses, imo. Like reading from memory with more io traffic.

1. Page operations concurrent with reads and writes.
2. IO pin redundancy for high packaging yield,
3. One differential clock for commands, address, and data. (Unlike GDDR5(X), which has half rate clocks for commands and addresses.)
4. Temperature dependent refresh rates,
5. Double data rate (DDR) for data and address bits.

 

[dj-electric]

That RX 5600M config is what i wish i had in a NUC-sized computer, for real.

 

[Valantar]

Sadly all of the other results are hampered by a meagre 50W power budget too, so there's no comparable reference point to a regular 5700XT running on Windows.

All I can see is that a 40CU part is about 50% quicker than the 24CU part , so we know that it's bottlenecked by something other than CUs; Probably power, because it's unlikely to be the HBM2.

With no details on how the test is run or under what conditions, these results are just noise.

The Radeon Pro 5500M has a rated boost clock of 1300MHz and has a rated TDP of 85W. In other words, this new GPU is delivering ~150% of the performance at ~60% of the power, or up to 2.5x the perf/w. And it's not like the 5500M is being pushed very far up its voltage/frequency curve - we know how high 7nm Navi can go. It is of course possible that the Pro 5500M power/thermal throttles and runs at less than 85W sustained, but nonetheless, this is a massively impressive showing from AMD, particularly when taking into account that these are fundamentally the same architecture.

As for testing conditions: these are laptops. Beyond controlling ambient temperature and background applications, running them out of the box is the only thing necessary.

 

[InVasMani]

Ironically 6 of these GPU chips on a single PCB in a dual heatsink/blower design would only pull 150w TDP. On the other hand 3 would be 75w TDP and only could just be PCIE bus powered. These seem rather efficient and compact in design as well.

 

[cucker tarlson]

gtx 1660 performance at 50w.
nice.

 

[Valantar]

Ironically 6 of these GPU chips on a single PCB in a dual heatsink/blower design would only pull 150w TDP. On the other hand 3 would be 75w TDP and only could just be PCIE bus powered. These seem rather efficient and compact in design as well.

What? There is a single GPU in question here, at a 50W TDP. 3x50 =150, not 75. The other chips on the package are HBM.

 

[InVasMani]

I thought this was a dual chip GPU design with HBM at 50w except it's made for laptops.

 

[Chrispy_]

The Radeon Pro 5500M has a rated boost clock of 1300MHz and has a rated TDP of 85W. In other words, this new GPU is delivering ~150% of the performance at ~60% of the power, or up to 2.5x the perf/w. And it's not like the 5500M is being pushed very far up its voltage/frequency curve - we know how high 7nm Navi can go. It is of course possible that the Pro 5500M power/thermal throttles and runs at less than 85W sustained, but nonetheless, this is a massively impressive showing from AMD, particularly when taking into account that these are fundamentally the same architecture.

As for testing conditions: these are laptops. Beyond controlling ambient temperature and background applications, running them out of the box is the only thing necessary.

Isn't this set of graphs specific to the 16" MBP though? That's the same exact strict 50W limit on the 5300M and 5500M.

Sure, Wikipedia and TPU will give you 85W as the nominal reference TDP but that's a generic spec which in the case of mobile parts is rarely adhered to by the laptop manufacturer.

In terms of apples-to-apples comparison, we are literally comparing Apple MBP 50W GPUs with the same cooling solution and TDP.

 

[Darmok N Jalad]

Now the next question would be, will they drop this into a refreshed iMac Pro and let the clocks fly? I imagine it would just earn itself the 5700 Pro label.

 

[R0H1T]

I thought this was a dual chip GPU design with HBM at 50w except it's made for laptops.

I like what you're thinking, too bad Apple isn't going to splurge millions on such a solution that'll likely fetch them less than a few billion

 

[mechtech]

Can I get one on a small card for my HTPC please?

 

[Valantar]

I thought this was a dual chip GPU design with HBM at 50w except it's made for laptops.

Dual chip? No. It is (obviously, as no such thing has ever been done) a single GPU chip on the package. There have of course been dual GPU graphics cards, but those have two GPU dice on separate packages. This has one GPU die + two HBM 2 dice + what looks like two tiny ancillary dice (something ECC-related, perhaps?) on one package). As such it isn't radically different from a Vega 10 or 20 die or a Fiji die, all of which had a single GPU die + multiple HBM(2) dice on a single interposer and in a single package. If this was the first MCM, multi-GPU die GPU, it would be a very, very big deal. There have been persistent rumors of AMD working on this, but so far nothing real, and RDNA(1) is definitely not designed to accommodate this. I would expect something like that for RDNA 3 at the earliest.

Isn't this set of graphs specific to the 16" MBP though? That's the same exact strict 50W limit on the 5300M and 5500M.

Sure, Wikipedia and TPU will give you 85W as the nominal reference TDP but that's a generic spec which in the case of mobile parts is rarely adhered to by the laptop manufacturer.

In terms of apples-to-apples comparison, we are literally comparing Apple MBP 50W GPUs with the same cooling solution and TDP.

Nvm, I misread that. Seems like the Apple-only Pro 5500M is indeed also at 50W, and the RX 5500M is 85 - not a very skimmable text on Notebookcheck's part, but I should have read more carefully nonetheless. Still, even 1.5x perf/W is quite impressive on the same architecture. It certainly makes AMD's promises of up to 50% increased perf/W for RDNA 2 seem more plausible.

Can I get one on a small card for my HTPC please?

Beat you to it (though in the launch thread ):

If one has the money for that, that's a good approach, particularly with AMD cards that for the past near decade have been pushed too far up their DVFS curves.

I'm just imagining a ... let's call it RX 5700 Nano - why not? It would sure be a worthy successor to the near-legendary R9 Nano - at 75W, two slot HHHL form factor, probably ~1300MHz or possibly a bit more (given that this does 1035 at 50W). That would absolutely destroy the current highest performance HHHL GPU, the GTX 1650. They could even make it a harvested die with a couple of CUs cut off, letting Apple run off with the best chips. It could still deliver ~1660 Ti performance unless the frequency scaling falls off a cliff at low clocks (considering how low the MBP version is clocked, there should be plenty of headroom without getting notably inefficient). It might not be the highest volume product ever - not by any means given the premium pricing something like this would demand - but it would be the darling of the SFF crowd for years to come.